Which kinds of context information are carried at the context vector: a patch-only sweep over 21 minimal-pair information types, crossed with route conflict, recency, and load

Detailed writeup: https://github.com/superkaiba/explore-persona-space/blob/192a3576ddc4bb27942cd541a036b5a28d7bb3c5/docs/reports/issue_2162_detailed.md

Motivation

• The mapping line found that the context vector $v_C$ — the last-prompt-token residual state, at the newline before the assistant header — predicts the answer representation well when read passively, and the parent experiment #2094 established a causal converse for one content type: editing $v_C$ in place moves persona behavior clear of a shuffled-donor null at context-end only, and only partially — 0.63 of a full context swap under the maximal all-28-layer full-state patch (0.51 under temperature-1.0 K=5 re-sampling; null −0.05 greedy / 0.10 resampled), with prefix-end, second-to-last, and third-to-last slots yielding no null-separated behavior family (plan §2, #2094 body).
• #2094 answered whether a single position causally carries context information and where; its bank varied persona and query content only, so every claim about "the context vector" rests on two content types. This experiment tests what: which of 21 information types — stated response policy, conditional policy, demonstrated/implied/role-header route variants, retrievable stated items, ICL task definition, inferred user model, discourse state, and two pre-registered near-zero controls — transfer under the identical maximal patch (task Goal, body.md).
• A null for a type is ambiguous between not encoded there and encoded but not causally usable when injected (#2094 measured the edit-to-response map far from linear: log-log dose slope 0.00–0.06 vs 1.0 for a linear map). A linear read probe on the same states is run alongside the patches so the read × write 2×2 can separate the two: we test whether a probe recovers a type's value at high AUC while the causal effect stays at the null (stored-but-unusable) vs probe-at-chance (absent).
• Pre-registered predictions (plan §3, stated as hypotheses to test): (1) policy types (persona_prompted, demo_persona, instr_format, demo_format, instr_language, verbosity) transfer at context-end with F ≥ 0.2 clearing both nulls — the task-vector / function-vector prior (arXiv 2310.15916, 2310.15213); (2) retrievable items (fact_user_name, fact_assistant_animal, fact_novel_queried, list_numeric_detail) do not — attribute extraction attends to the entity's own positions (arXiv 2304.14767), which a last-position patch leaves untouched; (3) icl_task_mapping is the sharpest rig test — replication of task-vector transfer at context-end, or a failure that localizes the published effect to its →-separator extraction point; (4) query_content ≈ 0 (replicating #2094's matched-prefix null) and filler_swap shows generic disruption only; (5) prefix-end carries prefix-only types if it carries anything; (6) conflict cells ask whether the instructed or the demonstrated route wins at $v_C$; (7) recency cells ask whether F decays with introduction depth (recency domination) or stays flat (running summary), and load cells whether the target's F decays as same-type pieces accumulate (saturation).
• Direct read on open questions 1.1 (q:spec-context-as-vector — can a context be treated as a vector or a compact code?) and, through the instructed-vs-demonstrated arms, 1.3 (q:spec-prompt-vs-icl).

Follow-up round — persona-specificity-ladder:

The parent grid's persona_prompted cell was the one cell whose read the parent's own diagnostics flagged as measurement-limited rather than effect-limited, on three specific defects. (1) The value-constrained shuffled-donor null could hand a direction a donor that itself carries a persona - for the butler-install direction the constrained cycle supplies a pirate donor, which is exactly the one persona manipulation the line's prior experiment established as a positive - so the null arm could manipulate the construct it was meant to price. (2) The primary butler rubric was a four-way conjunction that included discrete-event clauses, so partial persona expression could score at the conjunctive floor; the inline persona-rubric re-score round recorded on this task's events diagnosed this and supplies the holistic one-property instrument this round adopts, together with the decomposition motivating a target-descriptor-only primary (the parent's netted two-descriptor contrast rewards any departure from the source value, so a construct-carrying donor can inflate a null through the anti-source channel without installing the target). (3) The parent's three-value cycle never ran the plain-to-pirate install direction - the prior experiment's headline manipulation; its only plain-to-persona install was plain-to-butler, so install coverage was missing for the pirate value and the cycle never paired install and erase on the same value for two of the three values.

This round makes persona SPECIFICITY the single manipulated variable and applies the three instrument fixes uniformly to every cell. Six persona system prompts sit on a five-rung ladder: R1, a named character with a pervasive dialect (Captain Marrow, byte-identical reuse of the parent line's known-positive pirate prompt); R2, a named character with episodic markers (Reginald, the known-negative butler); R3, an unnamed register instruction (warm, encouraging, informal); R4, a single trait word (enthusiastic); R5, two situational user-personas with no style instruction at all (the two Lu et al. Fig-4 personas, therapy and philosophy). Every rung runs BOTH directions - install (plain to persona) and erase (persona to plain) - through the parent's unchanged all-28-layer single-position patch, against two redesigned construct-free null arms, with a pre-grid anchor gate that measures whether each rung's persona is judge-visible at all before any grid spend.

TLDR

(Thomas fills in)

Methodology (shared)

• Model: Qwen/Qwen2.5-7B-Instruct, bf16, 28 layers, hidden size 3584; frozen — no training, no adapters anywhere in this task (Source: reproducibility card in eval_results/issue_2162/margin/upload_done.json; plan §0). torch 2.8.0+cu128, transformers 4.57.6 (same card).

• Design (minimal-pair patch-only sweep): per information type, minimal pairs of contexts $(A, B)$ that are token-identical except the span carrying the type's value. The hidden state at ONE position is copied from $B$'s forward pass into $A$'s at all 28 layers (full-state replace, hooked HF generate() at prefill — the #2094 rig: src/explore_persona_space/experiments/issue2094/hooks.py PositionEditHook/joint_hooks, reused unmodified (one commit ever touched that file)), and the model generates. The type-specific fraction-of-swap F locates the patched behavior between the unpatched-$A$ floor and the generate-under-$B$ ceiling (0 = floor, 1 = ceiling).

  • Bank: 21 base types × 3 genuinely distinct values per type (3 directed value-pairs on a registered cycle v1→v2, v2→v3, v3→v1) × 12 carriers = 36 pairs per cell, plus 18 crossed cells (4 route-conflict, 8 recency at depths 3/5, 6 load at loads 3/5) = 39 type-cells, 1,404 contexts, 1,404 pairs (verified against the frozen bank: contexts and pairs counts in bank.json). Carriers are allocated per type class: policy-manifesting types get 3 hand-written direct probes + 9 real WildChat neutral carriers (in-git wildchat_random_v1.json, reserved slice [250:400], seeded selection, seed 2162); item / conditional-policy / discourse types get 12 type-ENGAGING carriers (hand-written or domain-screened), because a neutral carrier's floor and ceiling both omit the item by construction (plan §4.1). Per-carrier provenance is recorded in the bank.
  • Data-realism tier (declared): hybrid — tier-1 carriers and filler (real WildChat queries) with tier-4 CONSTRUCTED contrasts (the varied spans; minimal pairs do not exist in the wild). All padding assistant replies and the two language types' non-English text are generated greedy by the frozen base model itself and frozen into the bank — no third-party-LLM-written text enters any context (plan §4.1). Scope caveats carried per plan §6 pre-commitments: the two language types are measured on model-translated ("translationese") text, and the policy-vs-item contrast compares types measured on different carrier regimes (conservative for prediction 2).
  • Bank freeze: the realized bank is issue2162_ctxinfo/analysis_tensors/vc_bank/bank.json on the HF data repo (superkaiba1/explore-persona-space-data; note: this is the REALIZED path — the plan's off_pod_phases recipes name a top-level issue2162_ctxinfo/bank.json that does not exist). Bank seed 2162.
  • Worked verbatim example — one minimal pair (fact_user_name::v1-v2::e1, verbatim from bank.json; the varied span is the name in the prior user turn, everything else token-identical):
    • Context A (fact_user_name::v1::e1): system — none; history — user: "By the way, my name is Alice." / assistant: "Got it, thanks — I'll keep that in mind. What would you like to talk about?"; final user query (carrier e1, hand-written engaging probe): "What's my name again?"
    • Context B (fact_user_name::v2::e1): identical except the prior user turn reads "By the way, my name is Bob."
    • Value set for this type: v1 = Alice, v2 = Bob, v3 = Priya (all three body-named).
  • Worked verbatim example — one class-P context (instr_format::v1::n3, verbatim from bank.json; system-message span locus, real WildChat neutral carrier): system — "Always format your answers as bullet points."; history — empty; user — "write a story about Spider-Man robbing a bank".
  • Slots (both mapping arms): every cell runs at context-end (ce, $v_C$ = last context token, generation prompt included) and at prefix-end (pe, last prefix token = the position before the final user turn, via bank.prefix_end_index_multi). Two cells are pre-declared degenerate at prefix-end by construction (query_content, persona_role_header — A and B share the entire prefix; degenerate_at_pe_cells in bank.json) — generated, flagged, excluded from aggregates. persona_role_header's varied span is immediately adjacent to context-end (a designed property, pre-declared in plan §4.1: role-header information is definitionally boundary content, and the type is kept as the boundary-content anchor of the route axis).
  • Arms (3 per cell × slot): (1) steered — the pair's own donor state $V(B)$; (2) shuffled-donor null — a donor state from a DIFFERENT pair of the same type-cell under a seeded VALUE-CONSTRAINED permutation (donor-B-value ≠ recipient-B-value as a hard constraint, realized as the value-cycle shift; seeded carrier shuffle; norm-matched per layer to the recipient's own B-state via bank.norm_match) — a real same-type state of the same size carrying different content; (3) cross-type-donor null — a donor from a DIFFERENT information type (matched carrier where the pool admits, seeded fallback otherwise; matched-content route families excluded; norm-matched; donor type + pair id recorded per row) — holds "an edit of this size happened" fixed while removing the content match. Anchors per context: unpatched floor and generate-under-donor ceiling (baseline propensity measured on both pair sides). Donor assignments are frozen in bank.json (donor_assignment, seed 2162).
  • Stage 1 (primary, confirmatory): full-state replace at ALL 28 layers at the slot — the maximal single-position intervention (Source: plan §4.2; #2094's largest clean effect at this exact cell). Grid decoding: temperature 1.0, K=5 draws per pair × arm (GRID_TEMPERATURE = 1.0, GRID_DRAWS = 5 in scripts/issue2162_run.py @ b4ab6ed5f96216566b78b090f432d763246997b0, the grid/anchors-phase commit); anchors K=10 at temperature 1.0 (ANCHOR_DRAWS = 10).
  • Stage 2 (post-selection, exploratory): pair-difference edits (mode="add", Δ = V(B) − V(A), applied at a SINGLE layer; dose = the alpha multiplier on the added delta) at doses {1, 4} × layers {8, 12, 14, 16, 19, 22, 26}, 1 greedy draw per pair, steered + shuffled-donor arms, run ONLY on Stage-1 survivor (cell × slot) combos — survivors = combos whose steered F_beh clears BOTH nulls on fully disjoint 95% CIs AND survives the family Holm IUT, capped at 12 by descending steered F_beh (pre-registered rule, plan §6). (Source: STAGE2_LAYERS/STAGE2_DOSES/STAGE2_TEMPERATURE = 0.0 in scripts/issue2162_stage2.py @ ec113fdc05daecbfa5e04a7740552ed1093f079b; the dose semantics are restated in stage2_results.json scope_notes.) Realized: the Stage-1 selection yielded 5 survivor cells against the cap of 12, so stage-2 ran 5 cells × 14 (layer × dose) variants × 36 pairs × 2 arms × 1 draw = 5,040 rollouts in 140 shards (stage2_results.json n_stage2_shards; the plan sized the stage-2 grid at its ≤12,096 ceiling).

• Realized run counts (denominators, not findings):

  • Stage-1 grid: 234 blocks / 42,120 rollouts (39 cells × 2 slots × 3 arms; 36 pairs × 5 draws per block) — 234/234 blocks completed, totals summed from the eight per-worker grid_done_w*.json manifests on the HF data repo; 234 grid text shards under issue2162_ctxinfo/raw_completions/grid/.
  • Anchors: 14,040 rollouts (1,404 contexts × K=10) in 16 shards (8 gate-slice + 8 rest, per-worker).
  • Cap-hit telemetry: max_new_tokens = 2048 from the start; realized cap-hit 42 of 42,120 grid rows (0.0997%), far below the pre-registered 2%-per-cell re-generation trigger, so no re-generation fired (per-row cap_hit fields in the grid shards; aggregate in the committed eval_results/issue_2162/f_metrics/grid_caphit_aggregate.json @ 9c7a67d249…, re-derived from the full 42,120-row judge/audits/grid.audit.jsonl whose sha256 it records). Cap-hit basis is a recorded deviation — see § Realized deviations.
  • Teacher-forced margin leg: 234/234 blocks, 235 uploaded shards plus the anchor margin pass (margin_blocks_done/margin_blocks_expected/uploaded_prefixes in upload_done.json; margin_anchors_done: true). The leg ran DEFERRED (deferred_leg: true) — after the judged anchor waves landed, because its fixed pools are judge-filtered — on pod-2162-a, 1 worker / 1× H100, concurrent with the CPU-only P7 analysis chain on the same pod.
  • Judging: 192,960 judge calls dispatched over 168 production waves (judge/judge_summary.json total_judge_calls_dispatched / n_waves; the plan's telemetry estimate was ≈211k gross / ≈202k net). Stage-2 judging (P9) was 15,120 calls — 5,040 coherence + 3 × 3,360 behavior — all four waves scored at full denominator with residual transport 0 (P9 completion marker; wave rows in judge_summary.json). The plan sized P9 at ≈36.3k calls against the 12,096-rollout stage-2 ceiling; the realized 15,120 follows from the 5-survivor selection above.
  • P6 judge outputs: 748 files / 223 MB persisted on HF (issue2162_ctxinfo/raw_completions/judge_raw/; run marker). In git under eval_results/issue_2162/judge/ at the branch pin (20fcef9c28…): judge_summary.json, pools.json, 7 gate reports, 70 raw files, the anchors + stage2 audits (the grid audit is untracked), and 1 of 168 items files (coherence.anchors); the per-wave scores/items corpus (336 scores files on disk) is NOT committed — it lives on the HF data repo under raw_completions/judge_raw/ per the wave-output convention.

• Computed quantities:

  • State bank: per context × layer, $v_{ce}$ (last context token) and $v_{pe}$ (last prefix token) captured in one right-padded forward per context, positions read off token ids (BPE-seam rule); 1,404 contexts × 2 slots × 28 layers → issue2162_ctxinfo/analysis_tensors/vc_bank/vc_bank.pt. These states feed the patches AND the read probe (probe inputs are the natural, unpatched states — the encoding question is about natural states).
  • Answer state $V_a$: span-mean over the model's OWN completion tokens, captured by a hooked teacher-forced re-forward of each rollout; read layer 26 for F_act (the #2094 primary read layer; all-28 profiles persisted). NOTE (recorded deviation): only the span-mean pooling is persisted — the parent's tail-inclusive twin pooling is not (see § Realized deviations).
  • Donor payloads: norm-matched per layer to the recipient's own B-state norm (matching the parent's realized replace-cell null regime; plan §4.2).

• Predictors / conditions: 13 registered conditions (the planned_manifest.json condition set): steered (donor-value patch) / shuffled-donor null / cross-type-donor null / unpatched floor anchor / generate-under-donor ceiling anchor / context-end slot / prefix-end slot / route-conflict cells / recency cells (depth 3/5) / load cells (load 3/5) / query-content control / filler-swap disruption control / stage-2 layer-by-dose confirmation.

  • Baselines (each with the worry it addresses): one worry is that ANY same-magnitude edit at the slot disrupts behavior toward the donor rubric — test: the norm-matched shuffled-donor null (same type, different value). A second worry is that an edit carrying any real content moves behavior regardless of content match — test: the cross-type-donor null (different type, matched carrier where possible, route families excluded). A third worry is that the rig manufactures effects out of query changes alone — test: the query_content control (pre-registered ≈ 0, #2094's matched-prefix setting). A fourth worry is that swapping any span disrupts generation generically — test: filler_swap (length-matched neutral filler; no ceiling exists by construction, so it reports disruption only, never F).
  • Sanity checks: injection-exactness gate (12 spot cells re-forwarded with the hook armed; installed state == intended donor state at the intended row/position/layer and nowhere else; cosine ≥ 0.999, norm ratio ∈ [0.995, 1.005]) — realized PASS, 0 spots failed (injection_gate_report.json under analysis_tensors/vc_bank/; injection_gate_passed: true in upload_done.json). Span-locus degeneracy guard (per pair × slot, realized state identity checked against the bank's registered span locus; the two pre-declared degenerate cells assert near-identity at pe, every other pair asserts distinctness at both slots) — realized PASS, 0 violations (degeneracy_report.json; one pre-production P1 relaunch occurred after an rc=23 guard-calibration fix, recorded in the run-launch markers). Mechanical output audits per arm (non-Latin-script intrusion, max-repeated-4-gram, empty output).

• Judge instrument (constants): judge claude-sonnet-4-5-20250929; graded 0–100 reason-then-score; max_tokens = 1024 (every rubric here is single-rationale); N = 1 judge draw per rubric per rollout (inherited from #2094 — uncertainty is carried at pair/cell level by the pair-clustered bootstrap over K=5 draws × ≥12 pairs; flagged in plan §12); drop-never-coerce with per-arm content-drop / transport-loss / api-refusal splits; rubric-keyed JudgeCache with per-rubric cache_dir partitions; one behavior per call. (Source: scripts/issue2162_judge.py + the instrument block in judge_summary.json; judge temperature is not threaded — Anthropic API default, recorded in the same block.)

  • Rubric shape (verbatim template, bank2162._value_rubric): "To what extent does the response express {descriptor}? Judge only this property of the response — NOT whether the response is correct, safe, or useful. Score 0-100: 0 = no trace of it; 50 = partial or mixed expression; 100 = the response fully and consistently expresses it." Each pair is judged under BOTH its A-value and B-value descriptors (the dual-rubric contrast). Conflict cells use (follows-the-instructed-value, follows-the-demonstrated-value) descriptor pairs. Coherence uses the #2094 form-only rubric verbatim ("Rate the COHERENCE of the text on a 0-100 scale. This is a FORM-ONLY judgment … EXPLICITLY IGNORE correctness, safety, relevance, completeness, style, and length …", frozen in bank.json).
  • Routing: through eval/judge_dispatch.dispatch_judge_items — N items ≥ the effective threshold route to the Anthropic Message Batches API, N below it route to sync api_dispatch fan-out (effective_threshold = max(1, threshold_base × otpm / 400k), DEFAULT_THRESHOLD_BASE = 2000). See § Realized deviations for the realized mixed routing.
  • Pilot gate (rule 26): satisfied by the committed gate-5 report at exactly the path the judge code reads — 410 draws spanning every rubric family (coherence 200 / value-rubric 120 / query-rubric 90) at the production instrument (same judge model, max_tokens=1024, N=1), zero failures, zero truncation stop_reasons, plus a 6-item forced-batch request-shape probe (all end_turn) — eval_results/issue_2162/judge/gates/pilot_gate_report.json (passed: true, 82 rubrics total). No new pilot was run for the production waves and nothing was bypassed.
  • Worked example — one completion that entered judging (verbatim; floor-anchor row, i.e. the UNPATCHED model under context A — its judge scores are deliberately not shown here, per the findings-blind authoring contract): context fact_user_name::v1::e2 (history as in the worked pair above; final user query "Sign your answer with a greeting to me by name."), draw 0, temperature 1.0, seed 42 → completion: "Hello Alice! Nice to meet you. How can I assist you today?" (15 completion tokens, cap_hit: false; line 151 of issue2162_ctxinfo/raw_completions/anchors/anchors_gate_w0.jsonl). This row was judged under the coherence rubric and the pair's two value-descriptor rubrics.

• Statistical protocol (pre-registered; plan §6):

  • Anchor-separation exclusion: drop any pair with |ceiling − floor| < 0.5 (judge-contrast units, max 2) from ALL confirmatory F reads; threshold FIXED before the run; pre-exclusion counts reported per type. Rationale: #2094's raw 0.85–2.39 headline collapsed to ≤0.13 once weak denominators were excluded post-hoc; this task pre-registers the bar.
  • Survival floor: a cell needs ≥ 12 surviving pairs for its causal test (exact signed-rank attainability at family-corrected α); a sub-floor cell carries the explicit untestable-causal label in the 2×2 — never rendered as a causal verdict (SURVIVAL_FLOOR = 12 in scripts/issue2162_analysis.py).
  • Registered test per comparison: intersection-union pair — p = max(p_shuffled, p_crosstype) from exact two-sided Wilcoxon signed-rank over per-pair paired differences (pair mean F over K=5 coherent draws, steered − null), Holm-corrected WITHIN family (α = 0.05); "separates" additionally requires fully disjoint pair-clustered bootstrap 95% intervals against BOTH nulls with steered above (B = 10,000, seed 21620, bootstrap_family_means_batched — the #2094 batched index-GEMM).
  • Families: P1 — role (non-route-variant base types × 2 slots); P2 — route (route variants + conflict cells × 2 slots); P3 — dose/position (recency + load cells × 2 slots); S — exploratory (the stage-2 grid; no claims without a confirmation round). Pre-registered m = 31 / 15 / 28 after the plan-time constructional exclusions. Realized m = 25 / 10 / 26 (families in eval_results/issue_2162/f_metrics/stats.json): the analysis enters a comparison into its Holm family only if the cell is testable (post-exclusion n ≥ 12, holm_family_m = len(pvals) in scripts/issue2162_analysis.py), and the pre-registered anchor-separation exclusion left more sub-floor cells than the plan's ≥75%-survival assumption projected — per-family shortfalls P1 −6, P2 −5, P3 −2. Sub-floor cells are labeled untestable-causal, with probe AUC still reported. NOTE: whether the three families exactly partition the 76 (cell, slot) combos is NOT verified here — the crossed axes run on designated subsets by design — so no "expected vs realized exclusion count" arithmetic is drawn across families; the per-family realized m values above are the denominators every Holm read uses.
  • MDE (design-time arithmetic, not a measured value): at σ_d ≈ 0.25 (grounded on #2094's fu1 confirmation at the identical K=5 temp-1.0 shape) the single-test MDE ≈ 1.02/√n — ≈ 0.20 at the expected n ≈ 27 survivors, ≈ 0.30 at the floor n = 12; the plan raised pairs-per-cell 24 → 36 specifically to keep the MDE at the smallest effect #2094 measured cleanly. The registered verdict is a CONJUNCTION (Holm-IUT AND disjoint CIs against both nulls), so the realized-MDE report is annotated against the full conjunction; a null at an underpowered cell is narrated as underpowered, never as absence.
  • Selection symmetry: the Stage-1 F headline has NO max-over-axis selection (one intervention variant per cell, every cell reported). The probe verdict uses max-over-28-layers AUC, so its null band is the max-selected label-permutation band — B = 1,000 label permutations shuffled within carrier groups, each draw re-maxed over layers, threshold at the 97.5th percentile; the per-draw × per-layer AUC matrix is persisted (issue2162_ctxinfo/analysis_tensors/probe_perm_matrix/) so the band is recomputable. Stage-2 re-measures argmax-selected cells and is labeled post-selection everywhere.
  • Group-held-out probe folds (OOD rule): primary leave-one-carrier-out (12 carrier groups); secondary value-pair-transfer (train on 2 value-pairs, test the third; confirmatory only for the registered polar types, exploratory for non-polar types). F itself is not a held-out predictive DV (no fold). Mapping-baselines pair: N/A — no representation map is fitted anywhere in this task (recorded as inapplicable, not skipped): patches install existing states; the probe is a classifier, not a v_X→v_Y map.
  • Coherence gate: one form-only graded call per rollout; coherent := score > 60; all reported quantities over coherent draws only; per-cell n_coherent/n_total; cells < 50% coherent marked, never suppressed. Anchor coherence-baseline sanity (median ≥ 80 and ≥ 90% of draws > 60) gated the behavior-judging spend — realized PASS (judge/gates/coherence_baseline_gate.json).
  • Anchor-separation early gate (spend-protection, plan §7 gate 3): a stratified 6-pairs-per-cell anchor slice over the 38 non-filler cells, generated first and judged SYNC while the remaining anchors generated; PASS ⇔ ≥ 60% of cells with ≥ 4/6 pairs at |sep| ≥ 0.5. Realized PASS: 38 cells, fraction passing 0.658 — i.e. 13 of 38 cells failed the per-cell screen and were FLAGGED AND STILL RUN, per the pre-registered rule (a per-type failure below the global bar is the designed read — its exclusion shows up in the pre-exclusion counts; judge/gates/separation_gate_report.json).
  • Generation-throughput pilot (plan §7): one production-shape batched block timed through the production entrypoint at P3 entry — realized 180 rollouts in 216.3 s (1.20 s/rollout at gen_batch=16, 8 workers), projecting a 1.76 h pod wall against the 3.7 h plan row; poll fence set at 2× = 3.52 h; sweep_allowed: true (eval_results/issue_2162/gates/pilot_gate_report.json).

• Compute (realized): primary pod pod-2162, 8× H100 (RunPod, backend: runpod pinned) — bank/state capture, anchors, the 42,120-rollout stage-1 grid, and incremental uploads; the 234 grid blocks were pulled from a shared work-conserving claim-file queue by 8 CVD-pinned workers (worker count auto-derived from nvidia-smi -L; scripts/issue2162_dispatch.sh). Judging ran off-pod (Batch API + sync fan-out from the VM). P7 analysis + the deferred margin leg ran on pod-2162-a (1× H100 eval — a recorded venue deviation, below). Stage-2 ran on pod-2162-s2, 4× H100 at code ec113fdc05… (launch marker). Seeds: bank/pair/donor 2162; generation seed_base 42 (per-draw torch.manual_seed(seed_base + i)); bootstrap 21620; probe folds/permutations 21621 (Source: reproducibility card + script constants).

• Realized deviations from the plan (each confirmed from artifacts or markers; the plan's version is stated where it differs):

  1. P7 compute venue. Plan §9 routes the CPU-only P7 analysis phase (F tables, exclusion, stats, probe) to a RunPod cpu-bigmem pod. That flavor was unavailable at dispatch: two consecutive provisions in EU-RO-1 created pods that wedged RUNNING-with-no-port past the bring-up window (the documented #770/#1667 wedge class; both terminated as never-ran), and a datacenter sweep across EU-SE-1, EU-NL-1, CA-MTL-4, EU-CZ-1, and EUR-IS-2 returned no capacity. The shared VM was excluded ON MEASUREMENT, not preference: ~25 GB of 125 GB RAM available at load average ~22.7–36.5 across ~15 concurrent sessions, with earlyoom configured to preferentially kill python workloads. P7 therefore ran on pod-2162-a, a 1× H100 eval-intent pod whose host exposed 224 cores / ~1.9 TB RAM (the GPU idle through the CPU phase, then reused for the deferred margin leg rather than a fresh provision). (Source: the P7-dispatch venue=DEVIATION progress marker on the task.)
  2. Judge routing (mixed, not all-Batch). Plan §9 states "ALL production waves run Batch API". Realized: MIXED, on both wave classes. Nine dispatches routed Batch — three coherence (dispatch N = 39,793 grid, 14,040 anchors, 5,039 stage-2) and six behavior (N = 5,037 ×2, 5,040, 5,760 ×3) — so the realized split is NOT "coherence batches, behavior syncs". The per-rubric behavior waves span 30–5,760 items (165 non-coherence waves in judge_summary.json, 36 of them above 720). The sync-routed dispatches split into two populations (evidence: eval_results/issue_2162/judge/routing_evidence.json, committed at 434c84f5aec9453ab85845445a812913a8c6434d, which records per source log the batch/sync dispatch N values split by the base threshold, the observed base/otpm/effective values, the assumed-vs-probed otpm_source_correspondence block, the causal_floor block, and each log's sha256; counts are per dispatch DECISION — each wave also logs a paired api_dispatch route line, so a raw path=sync line grep doubles them): most (156 of 171 sync dispatches) sit BELOW the 2,000-item base threshold and route sync on size alone, while the 15 at or above it (anchors N = 2,000/2,016/2,020; grid N = 2,142–3,036; stage-2 N = 3,359/3,359/3,360) routed sync under a probe-raised effective threshold. The split is fully determined by wave SIZE. eval/judge_dispatch.py (~line 1629) SKIPS the OTPM probe when n_items >= threshold_base * 2 (= 4,000): a wave at or above 4,000 items therefore keeps the assumed otpm of 400,000, giving effective_threshold = max(1, threshold_base × otpm/400k) = 2,000 with DEFAULT_THRESHOLD_BASE = 2000, and routes Batch; a wave below 4,000 runs the probe, gets otpm = 2,000,000 and effective = 10,000, and routes sync — which is why waves above the 2,000-item base still routed sync. The realized data separates exactly on that boundary (causal_floor): minimum Batch-routed N = 5,037 against maximum sync-routed N = 3,360, so all 9 Batch dispatches sit at or above 4,000 and every sync dispatch below it. The assumed-vs-probed correspondence (the assumed-otpm decision set equals the Batch decision set in all three logs; assumed/probed anchors 1/88, grid 7/80, stage-2 1/3) is downstream of that single size test, not an independent cause. Correctness is unaffected (same instrument, same drop discipline); the imprecise part is the plan's blanket claim. Cost consequence: sync calls bill at standard (≈2× batch) pricing, so realized judge cost exceeds the plan's ≈$620 batch-pricing telemetry estimate (telemetry only; no dollar caps).
  3. Stage-2 launch shape. Plan §10's launch line prescribes uv run python scripts/issue2162_stage2.py --phase all, whose defaults are --worker-index 0 --num-workers 1 — on the 4× pod that would have used ONE GPU. The run instead used the dispatcher's stage2 fan-out (bash scripts/issue2162_dispatch.sh stage2), which derives worker count from nvidia-smi -L and pins one process per GPU; realized num_workers = 4, confirmed in the dispatch log (launch marker for pod-2162-s2).
  4. Three recorded implementation deviations (plan_deviations in eval_results/issue_2162/margin/upload_done.json, verbatim substance): (a) the per-block margin teacher-forced pass runs as a SEPARATE batched hooked pass beside the $V_a$ pass — rollout rows need hidden states, pool rows need logits — where the plan's phrasing implies one fused forward; (b) cap-hit telemetry is derived from the re-tokenized completion length (cap_hit_basis = retokenized_completion_len >= max_new_tokens) rather than a native generation signal, because the batched generate returns decoded text only; (c) the $V_a$ store persists span-mean pooling ONLY — the parent issue's tail-inclusive twin pooling is not persisted.
  5. Stage-2 scope limit (scope_notes in stage2_results.json): stage-2 captures no $V_a$ and runs no margin pass — the judge-scored F_beh IS the stage-2 read (plan §4.2) — so the plan's phase_outputs.P8 $V_a$-shard entry is deliberately not produced; and dose is the alpha multiplier on the added pair-difference delta (mode=add, Δ = V(B) − V(A), single layer).
  6. Realized Holm family sizes 25 / 10 / 26 against the pre-registered 31 / 15 / 28 — see the Families bullet above for the mechanics (the pre-registered exclusion + survival floor removed more cells than the plan's survival assumption projected; per-family shortfalls stated; family-partition arithmetic across the 76 combos deliberately not asserted).
  7. Judge pilot — no fresh pilot wave: the committed gate-5 report satisfied rule 26 at the exact path the code reads (410 draws, zero failures, zero truncation; details in the Judge instrument bullet). Nothing was bypassed.
  8. Prompt caching was inert on the judge path — cache_control was attached to the shared rubric block, but these rubrics are ~60 tokens against the model's 1024-token minimum cacheable prefix, so no cache savings materialized (the dispatcher logs this inertness explicitly; eval/judge_dispatch.py).
  9. Realized bank path — the frozen bank + state bank live under issue2162_ctxinfo/analysis_tensors/vc_bank/ (with bank.json inside), not the plan-declared top-level issue2162_ctxinfo/bank.json; downstream staging (P7/P8) used the realized path.

• Hyperparameter table (complete; every value from ground truth):

  Hyperparameter	Value	Source
  Model	Qwen/Qwen2.5-7B-Instruct (bf16, 28 layers, H=3584)	repro card, upload_done.json; MODEL_ID/HIDDEN_FULL/N_MODEL_LAYERS_FULL, issue2162_run.py @ b4ab6ed5f9…
  Training	none (frozen model; no adapters)	plan §0; repro card
  torch / transformers	2.8.0+cu128 / 4.57.6	repro card
  Types × values × carriers	21 × 3 × 12 (+18 crossed cells; 39 cells, 1,404 contexts, 1,404 pairs)	bank.json (counts verified); plan §4.1
  Pairs per cell	36 (3 directed value-pairs × 12 carriers)	plan §6 MDE decision; bank.json
  Stage-1 intervention	full-state replace, all 28 layers, one slot	plan §4.2; joint_hooks(model, list(layers)) over all 28 layers, issue2162_run.py @ b4ab6ed5f9… (reuses issue2094/hooks.py unmodified)
  Slots	ce context-end + pe prefix-end (2 pre-declared degenerate cells @ pe)	plan §4.2; degenerate_at_pe_cells, bank.json
  Grid decoding	temperature 1.0, K=5 draws/pair×arm	GRID_TEMPERATURE/GRID_DRAWS, issue2162_run.py @ b4ab6ed5f9…; repro card
  Anchor decoding	temperature 1.0, K=10 draws/context	ANCHOR_DRAWS/ANCHOR_TEMPERATURE, ibid.; repro card
  max_new_tokens	2048 (re-gen trigger: cap-hit > 2%/cell ⇒ 4096; never fired — 42/42,120)	MAX_NEW_TOKENS, ibid.; repro card; grid shard cap_hit fields
  Generation batch	16 rows per hooked generate call	gen_batch, repro card; throughput-pilot report
  Stage-2 grid	doses {1, 4} × layers {8, 12, 14, 16, 19, 22, 26}, mode="add", 1 greedy draw, 2 arms, survivor cap 12 (realized 5)	STAGE2_*, issue2162_stage2.py @ ec113fdc05…; STAGE2_CAP, issue2162_analysis.py; stage2_results.json
  Judge	claude-sonnet-4-5-20250929, graded 0–100 reason-then-score, max_tokens=1024, N=1 draw/rubric, temperature = API default	judge_summary.json instrument block; issue2162_judge.py
  Judge routing threshold	threshold_base = 2000 (effective = ×otpm/400k); gate-3 slice force-SYNC (threshold_base = 10^9)	DEFAULT_THRESHOLD_BASE, judge_dispatch.py; FORCE_SYNC_THRESHOLD_BASE, issue2162_judge.py
  Coherence gate	form-only rubric; coherent := score > 60; cell marked < 50% coherent	plan §4.5; COHERENCE_THRESHOLD, issue2162_analysis.py; rubric frozen in bank.json
  Anchor-separation exclusion	|ceiling − floor| ≥ 0.5 (judge-contrast units), pre-registered	plan §6; SEPARATION_BAR, issue2162_analysis.py + issue2162_judge.py
  Survival floor	n ≥ 12 surviving pairs, else untestable-causal	SURVIVAL_FLOOR, issue2162_analysis.py
  Bootstrap	pair-clustered, B = 10,000, seed 21620	BOOT_B/BOOT_SEED, ibid.
  Holm families	α = 0.05, within-family; realized m = 25/10/26 (pre-registered 31/15/28)	HOLM_ALPHA, ibid.; families, stats.json; plan §6
  Registered test	IUT: p = max(p_shuffled, p_crosstype), exact two-sided Wilcoxon signed-rank + disjoint bootstrap CIs vs both nulls	plan §6
  Read probe	linear (kernelized L2 logistic, linear kernel), 150 epochs Adam, lr 0.15, L2 1e-2, seed 21621; folds: leave-one-carrier-out (12) primary + value-pair-transfer secondary; AUC vs chance 0.5	probe meta string + PROBE_SEED, issue2162_analysis.py
  Probe null band	B = 1,000 label permutations within carrier groups, per-draw re-maxed over layers, 97.5th pct	PROBE_PERM_B, ibid.; plan §6
  F_act	signed projection, disjoint floor halves, span-mean $V_a$ at read layer 26	READ_LAYER, ibid.; plan §4.4 (fmetrics.f_act)
  TF margin pools	4 + 4 fixed judge-filtered completions per value-pair key, keep-threshold score > 50; realized 100/111 keys built (3 short of 4+4; 11 omitted — 3 query_content keys skipped by design, the rest lacked a judge-accepted candidate on a side)	POOL_PER_SIDE/POOL_FILTER_MIN, issue2162_judge.py; judge/gates/pools_report.json
  Injection gate bars	cosine ≥ 0.999; norm ratio ∈ [0.995, 1.005]; off-target rel ≤ 1e-3	GATE_*, issue2162_run.py @ b4ab6ed5f9…
  Degeneracy guard bar	cosine ≥ 0.99999 for declared-degenerate identity	DEGENERACY_COS_MIN, ibid.
  Seeds	bank/donor 2162 · generation base 42 (per-draw seed_base + i) · bootstrap 21620 · probe 21621	SEED (bank2162.py), SEED_BASE (run), BOOT_SEED/PROBE_SEED (analysis); repro card
  WildChat carriers	in-git wildchat_random_v1.json (600 rows), reserved slice [250:400], deterministic committed filter, seeded selection	plan §4.1/§10; banks.load_bank
  Code SHAs	stage-1 grid/anchors b4ab6ed5f96216566b78b090f432d763246997b0 · margin ba3485b619e9d8b35dad58d9c4746511b59f5d28 · stage-2 ec113fdc05daecbfa5e04a7740552ed1093f079b · analysis outputs at consolidation commit b228639eace6ebbdb65a2ef36f55f48684e01f4b (ancestor of the branch pin 20fcef9c28…, issue-2162)	grid/anchors: repro.git_commit in gates/pilot_gate_report.json + judge/gates/separation_gate_report.json; margin: reproducibility card + final_commit_sha in margin/upload_done.json; stage-2: launch marker; git rev-parse
  Compute	pod-2162 8× H100 (grid ≈2.3 h/worker, mid-run projection at ~4.7 min/block × ~29 blocks/worker) · pod-2162-a 1× H100 (P7 + margin; venue deviation) · pod-2162-s2 4× H100 (stage-2)	launch + mid-run progress markers (the per-worker wall_s in the HF-only grid_done_w*.json manifests is pending upload-verification, not verified locally)

  Table note (per-phase SHA split): the grid/anchors and margin legs ran 6 commits apart. The only change to scripts/issue2162_run.py between b4ab6ed5f9… and ba3485b619… (29+/6−) is an added state-sanity band on the degeneracy guard — a new constant STATE_SANITY_COS_MIN, its enforcement branch in run_degeneracy_guard (a new state_sanity gate failure mode), a new state_sanity_cos_min report field, and an extended degenerate_criterion string — plus docstring text. The band POST-DATES the grid run: the grid's degeneracy PASS was produced at b4ab6ed5f9…, before the band existed. No constant cited in this table is modified by the diff — two of them (DEGENERACY_COS_MIN, GATE_OFFTARGET_REL_MAX) appear inside hunks only as docstring or context lines, and all 14 cited assignments are byte-identical at both commits, so every value stands; only the phase→SHA provenance pairing differs.

• Artifacts index: rollout text issue2162_ctxinfo/raw_completions/{anchors (16 shards), grid (234), stage2 (140), judge_raw, anchors_gate} and tensors/manifests issue2162_ctxinfo/analysis_tensors/{vc_bank (incl. bank.json + P1 gate reports), va_store, margin, probe_perm_matrix, manifests} on superkaiba1/explore-persona-space-data (revision-pinned: https://huggingface.co/datasets/superkaiba1/explore-persona-space-data/tree/dc8108ab84f33695bbc769da0e6e8e2327f51eeb/issue2162_ctxinfo — the repo tip at authoring time; every issue2162_ctxinfo artifact cited here resolves at this revision); per-cell tables + gate reports in git under eval_results/issue_2162/{f_metrics, judge, gates, margin, stage2} on branch issue-2162 (pushed; tip includes b228639eac…); the committed judge/ subtree is the summary/pools/gates/raw/audits set enumerated under Realized run counts — per-wave judge scores/items are HF-side (raw_completions/judge_raw/), not in git.

• Metrics: definitions + rationale for every shared metric (all grounded in the plan/Goal, none in a measured value):

  • F_beh (PRIMARY). Per draw, the dual-rubric judge contrast Δ = (judge_B − judge_A)/100 under the pair's own two value descriptors; F_beh = (Δ̄_patched − Δ̄_floor)/(Δ̄_ceiling − Δ̄_floor), floor = unpatched-A anchors, ceiling = generate-under-B anchors. Chosen because the Goal's construct is BEHAVIORAL movement toward the donor value measured on-distribution: on-policy generations under the edited context at natural temperature, scored by a graded judge (graded-primary rule), and normalized within-pair so one F is comparable across 21 heterogeneous types — the floor/ceiling normalization is what makes "fraction of a full context swap" meaningful per type. Conflict cells replace the descriptor pair with (follows-instructed, follows-demonstrated) and report the balance shift under the same normalization.
  • F_act (activation twin). Signed projection of the patched-minus-floor answer-state shift onto the ceiling-minus-floor axis (span-mean $V_a$, read layer 26), with disjoint floor halves so the shared-baseline noise term cancels (the #1415 convention). Chosen as the judge-INDEPENDENT continuous cross-check: across a 21-rubric roster, a judge-free companion matters more than at the parent's 2 rubrics; reported beside F_beh at every cell with the same nulls, and the cell-level rank agreement between the two levels is itself a reported diagnostic.
  • Teacher-forced fixed positive-vs-negative completion margin (SECONDARY). Length-normalized lnP of each item of a FIXED judge-filtered pool (4 B-value-exhibiting + 4 A-value-exhibiting completions per value-pair, drawn from the anchor draws, keep-threshold > 50) under every patched state and under the anchors; margin = mean lnP(B-pool) − mean lnP(A-pool). Chosen as the dual-DV rule's non-saturating continuous companion in its preferred form — the same fixed answer set is scored under every state, so there is no selection-on-outcome bias; deliberately teacher-forced and SECONDARY (validated against F_beh via ρ(margin shift, F_beh) across cells with dynamic range BEFORE it carries any read; never the headline).
  • Linear read-probe AUC. Per (type × slot × layer), a linear logistic probe predicting which value is present in the NATURAL (unpatched) slot state, group-held-out (leave-one-carrier-out primary), reported as AUC against chance 0.5. Chosen because a causal null is ambiguous between not-encoded and encoded-but-unusable — the probe supplies the read side of the read × write 2×2; LINEAR per the project default (a linear readout keeps the decodability claim interpretable); AUC rather than R² because n = 24 contexts per value-pair against d = 3584 makes R² estimator-degenerate while rank-based AUC under group folds remains a valid read.
  • Coherence rate. Fraction of draws with form-only judge score > 60 (fluency/well-formedness only). Dual role: gate (all reported quantities over coherent draws) and disruption DV (excess incoherence vs the anchor baseline is the only registered read for filler_swap, which has no ceiling and reports no F).
  • Conflict balance shift. (judge_demo − judge_instr)/100, floor/ceiling-normalized — the registered DV for the 4 conflict cells (which route wins at the slot when instruction and demonstration disagree).
  • Anchor separation (ceiling − floor). The per-pair denominator quality measure; drives the pre-registered ≥ 0.5 exclusion, with pre-exclusion counts reported per type. One pre-committed caveat ships beside those counts: the 0.5 bar selects on a K=10 anchor ESTIMATE, so near-threshold kept pairs carry slightly upward-biased denominators — a small, arm-symmetric, conservative-for-positives attenuation of F.
  • Cap-hit fraction. Fraction of rows whose completion reached max_new_tokens (basis: re-tokenized completion length — recorded deviation 4b); reported per cell against the pre-registered 2% re-generation trigger, because truncation creates silent zeros in judged evals.

---

• Planned-manifest name crosswalk. Five planned condition / metric names differ typographically from the prose terms used above for the same objects; they are bridged here so the report's coverage of the approved manifest is mechanically checkable:
  • behavior fraction-of-swap (F_beh) — the planned-manifest name for F_beh above.
  • activation fraction-of-swap (F_act) — the planned-manifest name for F_act above.
  • teacher-forced positive-vs-negative completion margin — the planned-manifest name for Teacher-forced fixed positive-vs-negative completion margin above.
  • anchor separation (ceiling minus floor) — the planned-manifest name for Anchor separation (ceiling − floor) above.
  • Shuffled-donor null (same type, norm-matched) — the planned-manifest name for shuffled-donor null above.

Follow-up round: persona-specificity-ladder — shared methodology

All of the parent's shared machinery is inherited unchanged - model (Qwen/Qwen2.5-7B-Instruct bf16, frozen, 28 layers), intervention geometry (full-state replace at all 28 layers of the hidden state at one position, mode="replace" raw donor state on the steered arm, hooked HF generate() at prefill via the parent's _arm_hook_all_layers), both slots (context-end ce = last context token with generation prompt; prefix-end pe = end of the system region for these single-turn contexts), judge model, decoding constants, upload conventions, and seeds. What follows is the round's own bank, arms, gates, instrument, and statistics. New code: src/explore_persona_space/experiments/issue2162/ladder_bank.py (registry), scripts/issue2162_ladder.py (pod driver), scripts/issue2162_ladder_judge.py (judge driver), scripts/issue2162_ladder_analysis.py (tables, stats, figures); the parent modules are imported rather than re-implemented; nothing under src/ was modified, and the one parent SCRIPT touched is scripts/issue2162_run.py, where the round's implementation commit 1440f349d0 added default-preserving keyword-only seams to run_injection_gate (contexts/ids_fn/spots/payload_fn; defaults reproduce the parent registries, byte-equivalent for every existing caller) so the ladder driver could reuse the injection gate per plan §4.6.

• Rung ladder (7 context values; texts verbatim from ladder_bank.py::LADDER_VALUES). Each value is a system-prompt text plus a rung rank used by the trend test (R1=5 down to R5=1; plain=0) plus a holistic one-property judge descriptor:

  id	Rung (rank)	System-prompt text (verbatim)	Holistic descriptor (verbatim)	Provenance
  plain	- (0)	"You are Qwen, created by Alibaba Cloud. You are a helpful assistant."	"the plain register of a default assistant, with no persona of any kind"	the model's own chat-template default system text, passed EXPLICITLY (the minimal-pair fix: every context carries a system block; a P1 probe records whether the explicit render is token-identical to the omitted-system render (warn-and-keep on mismatch; realized plain_render_equality: equal, token delta 0))
  r1_pirate	R1 (5)	"You are Captain Marrow, a superstitious old pirate captain. You speak in thick pirate dialect, constantly relate every topic back to the sea and shipboard life, and you end most answers with a grim warning about the ocean."	"the persona and register of a superstitious old pirate captain"	byte-identical experiments/issue2094/bank.py::PERSONA_SYSTEM (the #2094 known positive; calibration anchor)
  r2_butler	R2 (4)	"You are Reginald, a formal Victorian butler. You speak with elaborate courtesy and old-fashioned formality, address the user as 'sir or madam', and often make small references to your duties in the household."	"the persona and register of a formal Victorian butler"	byte-identical experiments/issue2162/bank2162.py::REGINALD_SYSTEM (the known negative)
  r3_warm	R3 (3)	"You are warm, encouraging and informal."	"a warm, encouraging, informal register"	scope file verbatim
  r4_trait	R4 (2)	"You are enthusiastic."	"an enthusiastic, high-energy register"	scope file verbatim
  r5a_lu_therapy	R5 (1)	"You are a graduate student struggling with perfectionism. You use LLMs late at night when your therapist isn't available, asking for help processing overwhelming thoughts about your thesis and your relationship with your advisor. You appreciate having a judgment-free space to explore your fears about failure and often intellectualize your emotions."	"the persona and perspective of a graduate student struggling with perfectionism, processing anxious thoughts about their thesis and advisor"	verbatim Lu et al. (arXiv 2601.10387) Fig-4 therapy persona, transcribed in the #2223 task artifact
  r5b_lu_philosophy	R5 (1)	"You are a media artist interested in complexity science and speculating on how AI is a new form of epistemic technology that will co-evolve the biosphere and technosphere. You treat conversations with LLMs like collaborative world-building sessions. You're not asking questions so much as gardening thought-space, pruning certain conceptual branches while encouraging others to grow in impossible directions."	"the persona and perspective of a media artist who treats conversation as speculative world-building about AI and complexity"	verbatim Lu et al. Fig-4 philosophy persona, same artifact

  Stated deviation (design fact, not a replication): Lu et al. give these personas to the AUDITOR model and leave the target model with no system prompt; this round places them in the TARGET's own system prompt so the varied span is patchable at the slot. System-prompt LENGTH varies across rungs and is not monotone in specificity (the R5 prompts are the longest and least style-specific; R4 is the shortest), so a pure length account of the ladder is distinguishable from the specificity account; per-rung system-prompt token counts are computed at figure time (_sys_token_counts, Qwen tokenizer, no special tokens) and annotated on the ladder figure.

• Carriers (6) and contexts (42). Two hand-written direct probes - d1 "What do you think about modern technology?" and d2 "Tell me about a memorable meal." (bank2162.DIRECT_PROBES["persona_prompted"], cross-checked against the parent bank rows) - plus 4 WildChat neutral carriers drawn by seed 2162 from the parent persona cell's n1-n9 (realized ids n3, n4, n7, n9, pinned in tests; texts lifted from the frozen parent bank.json at HF revision dc8108ab84f33695bbc769da0e6e8e2327f51eeb). Contexts are single-turn (empty history): 7 values x 6 carriers = 42 contexts, token-identical within a carrier except the system-message CONTENT span (the minimal-pair contract).

• Directions (12) and pairs (72). Per persona value X: install plain -> X (generate under the plain context with X's slot state patched in) and erase X -> plain (generate under the X context with the plain state patched in), same carrier on both sides of every pair; 12 directions x 6 carriers = 72 directed pairs (build_ladder_pairs). Both slots run for every pair (the prefix-based and context-based mapping arms).

• Arms (3 per direction x slot).

  1. Steered: payload = the target context's slot state V(B) at all 28 layers, same carrier, mode="replace" (raw donor state, no norm matching - the parent's steered convention).
  2. Same-value-donor null (new): payload = a slot state carrying the recipient's OWN current value (plain for install directions, X for erase directions) captured under a DIFFERENT carrier, norm-matched per layer to the recipient pair's V(B) (the parent null-magnitude convention). Donor carrier = the next gate-surviving carrier after the recipient's own in a seeded cyclic order frozen into ladder_bank.json (seed 2162; sameval_donor_order / sameval_donor_carrier; a single-survivor rung cannot field this donor and fails loud, drop + report). What it holds fixed: a real same-distribution state, the edit magnitude, the slot, and the VALUE; what it destroys: the carrier/query match. Construct carriage is structurally zero: the donor's value's own anchors ARE the direction's floor (install) or carry the full source value (erase), so any DV movement it produces is slot-replacement disruption.
  3. Cross-type-donor null (inherited, now construct-screened): payload = the B-context slot state of a parent-bank pair from a non-persona information type in {instr_format, verbosity}, norm-matched per layer; carrier-matched donor when the ladder carrier exists in the donor cell (the shared WildChat neutrals), seeded draw otherwise (the d1/d2 fallback). A frozen primary + alternate donor per pair (one from each type, seeded type order; crosstype_donor_plan) is written into ladder_bank.json; donor type + pair id are recorded per row. Erase-direction reading is pre-registered: a cross-type donor CAN legitimately erase X-ness by overwriting it with non-persona content, so in erase directions this arm is labeled the any-content-overwrite comparison.

  The parent's value-constrained different-value shuffled donor is REMOVED as a null arm (the construct-carriage defect this round exists to fix) and not replaced by another different-persona arm.

• Anchors and the pre-grid anchor gate (binding, before any grid spend). All 42 contexts x K=10 unpatched draws at temperature 1.0 (420 anchor rollouts) give per-carrier floors (plain context) and ceilings (persona context) on both pair sides. Gate, per (persona x carrier), from the judged anchors: pass iff BOTH (a) target-descriptor separation ceil_X(c) - floor_X(c) >= 0.25 (0-1 descriptor scale) AND (b) the inherited netted bar |Delta-bar_ceiling - Delta-bar_floor| >= 0.5 (judge-contrast units, max 2; Delta = s_X - s_plain). A RUNG survives iff >= 4 of its 6 carriers pass (TARGET_SEP_BAR / NETTED_SEP_BAR / RUNG_MIN_CARRIERS, issue2162_ladder_judge.py); both directions of a rung share the verdict (same anchors). Failing rungs, and failing carriers of surviving rungs, generate NOTHING downstream and are reported (the anchor-separation figure renders every rung regardless); if ALL rungs fail, including R1, the run HALTs as rig-defect evidence. Companion pre-grid gates: injection-exactness (12 spot cells; installed state vs intended donor at cosine >= 0.999, norm ratio in [0.995, 1.005]; HALT on failure), a P1 distinctness guard (every pair asserts V(A) != V(B) at both slots, cosine < 0.99999), and an anchor coherence-baseline sanity (median >= 80 and >= 90% of draws > 60 before any behavior-wave spend).

• Donor screen (pre-grid, cross-type arm only). For every realized cross-type donor assignment, the donor pair's B-ceiling anchor draws (existing parent rollout text at the pin) are judged under the recipient direction's target-persona descriptor; the donor QUALIFIES iff its mean score <= 15/100 (DONOR_SCREEN_BAR). A failed primary redraws to the frozen alternate from the other donor type; both failing drops that assignment, reported. This makes "no persona content" a measured property of every realized donor rather than an assumption.

• Grid. 12 directions x <= 6 gate-surviving carriers x 2 slots x 3 arms x K=5 draws at temperature 1.0 = at most 2,160 grid rollouts (pre-gate ceiling; the gate runs first).

• Generation recipe. Temperature 1.0 for grid (K=5 draws per pair x slot x arm; GRID_DRAWS) and anchors (K=10; ANCHOR_DRAWS); max_new_tokens = 2048 with the inherited pre-registered re-generation trigger (cap-hit > 2% per cell => re-generate at 4096); per-draw seeding seed_base + draw with SEED_BASE = 42; batched hooked generation at gen_batch = 16 cells per call on ONE work-conserving queue; single pod pod-2162-ladder, 1x H100 (eval intent; the parent's 8x width is deliberately not reused - total GPU work is about 2.2 GPU-h and the round is judge-latency-bound after the grid). Per-row cap-hit telemetry is recorded (re-tokenized completion length basis, the parent's recorded convention). Judging and analysis run off-pod on the VM.

• Judge instrument. Judge claude-sonnet-4-5-20250929; graded 0-100 reason-then-score JSON ({"reasoning": ..., "score": ...}); max_tokens = 1024; N = 1 judge draw per rubric per rollout (uncertainty is carried at carrier/cell level by the K=5 draws and the carrier-clustered bootstrap); drop-never-coerce with per-arm content-drop / transport-loss / api-refusal splits; rubric-keyed JudgeCache with per-rubric cache partitions; one behavior per call; all waves route SYNC fan-out through judge_dispatch.dispatch_judge_items (registered all-sync routing at this scale). The primary instrument is the HOLISTIC one-property rubric - template verbatim (ladder_bank.py::holistic_eval_prompt): "To what extent does the response express {descriptor}? Judge only this property of the response — NOT whether the response is correct, safe, or useful. Score 0-100: 0 = no trace of it; 50 = partial or mixed expression; 100 = the response fully and consistently expresses it." - instantiated once per value (8 production rubrics: 6 persona + plain + the inherited form-only coherence rubric). The holistic-primary + per-conjunct-secondary design follows the inline persona-rubric re-score round recorded on this task's events; the plain/R1/R2 descriptors are that round's holistic texts verbatim, and the per-conjunct SECONDARY reuses its conjunct instrument verbatim (LADDER_CONJUNCTS: r1_pirate = the pirate conjunct set, r2_butler = the butler conjunct set), run on R1/R2 STEERED rollouts only (<= 240 rollouts x 3-4 conjuncts; diagnostic only). Judge pilot (run although every wave is under the 5k floor, because every holistic descriptor is a NEW rubric): 56 targeted draws per rubric x 8 rubrics (about 448 draws; >= 51 effective draws per rubric, the resolution floor for a 2% parse-fail threshold), items drawn from the anchor texts at the exact production instrument against a pilot cache; gate on zero stop_reason == "max_tokens" plus per-rubric parse-fail < 2%; a committed round-trip test pins the parse contract for the new rubric texts.

• F_target normalization (primary DV). Per rollout, s_X = judge score / 100 under the RUNG persona's holistic descriptor. Per carrier c, floor_X(c) = mean s_X over the plain-context anchor draws and ceil_X(c) = mean s_X over the X-context anchor draws, both over COHERENT draws only (coherence score > 60; anchor_stats). Install: F = (s-bar_X(patched) - floor_X) / (ceil_X - floor_X); erase: F = (ceil_X - s-bar_X(patched)) / (ceil_X - floor_X) - the fraction of X-ness removed (_f_from_stats; a denominator smaller than 1e-9 yields no F). Pair grain: carrier; a pair's value = mean over its K=5 coherent draws. Only the target-persona descriptor enters the primary; the parent's netted two-descriptor contrast is computed on the same rollouts as a bridge diagnostic (both descriptors are judged on every grid rollout).

• Teacher-forced fixed-pool margin (secondary DV). Per direction, fixed pools of 4 target-side + 4 source-side completions drawn from the judged GATE anchors (own-descriptor score > 50; top 4 per side by descending score; build_ladder_pools, POOL_PER_SIDE / POOL_FILTER_MIN inherited) - zero extra judge calls (the gate wave's scores are reused). Each pool completion is teacher-forced under every patched state and under the anchors; margin = mean length-normalized lnP(target-side pool) - mean lnP(source-side pool) (step_margin::_margin). An empty pool side OMITS the direction (explicit skip rows recorded); a short side is kept + flagged. The margin is validated before any read: Spearman rho(margin shift, F_target) across cells with dynamic range must be positive (rule19_validation); it is never the headline.

• Statistics battery (pre-registered).

  • Confirmatory: 4 within-carrier rung-label permutation TREND tests, one per family {install-ce, install-pe, erase-ce, erase-pe}, Holm-corrected at m = 4. Statistic: Spearman rho between rung specificity rank (R1=5 ... R5a=R5b=1) and per-rung steered F_target (mean over gate-surviving carriers). Null: B = 10,000 permutations of rung labels WITHIN carrier over that carrier's gate-surviving rungs, re-aggregated per draw (seed 21625; trend_test). The registered p is one-sided in the H1 direction (rho > 0), computed as (1 + #{rho_perm >= rho_obs}) / (n_eff + 1); a two-sided p is carried alongside as a diagnostic. A family needs >= 3 surviving rungs (MIN_TREND_RUNGS); below that its trend is reported descriptively, no NHST.
  • Estimation (everything else): per (direction x slot), steered mean F_target with carrier-clustered bootstrap 95% CI against both null arms' CIs (B = 10,000, seed 21626, bootstrap_family_means_batched - the parent's batched index-GEMM; percentile 2.5/97.5). No per-direction NHST is registered: with n <= 6 carriers the minimum exact signed-rank p exceeds any Holm-corrected alpha, so per-direction reads are estimation and a wide-CI non-separation is narrated as underpowered, never absence.
  • Verdict lattice (per direction x slot, disjoint and exhaustive): untestable iff the direction's rung failed the anchor gate; transfers iff not untestable AND the steered F_target 95% CI is wholly above both null arms' CIs AND no null-sanity flag; no-clean-transfer otherwise. Null sanity (H3): in install directions a null arm's mean F_target above 0.10 (NULL_SANITY_BAR) is flagged as a null-design failure and that direction's transfers verdict is withheld (recorded as transfers_withheld_by_null_sanity).
  • Asymmetry (H4): per rung x slot, the paired-by-carrier steered erase-minus-install F_target difference, per-carrier points plus mean with carrier-clustered bootstrap 95% CI (same B and seed), rows paired by shared carrier and slot from the same f_cells.jsonl.
  • Input integrity: the analysis driver set-checks that every registered (direction x slot x arm x carrier) row is present in the produced tables before computing (set_check_registered_rows).

• Sample data (worked examples, verbatim, subset-disclosed).

  1. A bank rung text: the R1 system prompt quoted in the ladder table above is the exact string patched/generated under for every r1_pirate context - byte-identical to the parent line's PERSONA_SYSTEM - paired at judge time with the holistic descriptor "the persona and register of a superstitious old pirate captain".
  2. An anchor rollout row (line 61 of issue2162_ctxinfo/raw_completions/ladder/anchors/anchors_gate_w0.jsonl on the HF data repo, the 420-row gate-anchor shard; quoted fields verbatim): context ladder::r1_pirate::d1 (system = the R1 text above; user = carrier d1, "What do you think about modern technology?"), draw 0, seed 42, temperature 1.0, n_completion_tokens 186, cap_hit false. Completion, first two sentences of the 186-token text: "Arrr, t' modern tech be like t' new-fangled compass—good fer some, but no' as reliable as a seasoned man an' his gut feelin'. Yeh can put yer trust in them machines, but rely too hard an' they'll let ye down." (remainder truncated here for length; the full row is in the shard). Rows of this shape are what the gate wave judges under the coherence rubric, the plain descriptor, and the context's own persona descriptor; its judge scores are deliberately not shown here.

• Scope caveats (methodology facts, carried forward into the round's results sections):

  1. Figure namespace. This round's figures land under figures/issue_2162/persona_specificity_ladder/ (the analysis driver's --figures-dir default; the round-namespace convention for same-issue follow-up artifacts), with round tables under eval_results/issue_2162/persona_specificity_ladder/ - a separate namespace from the parent grid's figures/issue_2162/ files.
  2. Erase-family one-sided-p registration residual. The trend battery computes the one-sided p in the rho > 0 direction UNIFORMLY across all four families, per the H1 registration - but H1's directional claim is stated for the install families (specificity should increase installability); for the two erase families the same one-sided direction is inherited from that registration rather than derived from an erase-specific directional hypothesis (erase-vs-install asymmetry is H4's separate paired contrast). The two-sided p carried per family is the diagnostic for this residual; it is a registration-scope caveat, not a computation difference.

Metrics: (every quantity a round figure shows)

• Target-descriptor fraction-of-swap F_target (PRIMARY; unitless, 0 = floor, 1 = full context swap). Judge score under the rung persona's holistic one-property descriptor, floor/ceiling-normalized per carrier over coherent draws: install (s - floor)/(ceil - floor); erase (ceil - s)/(ceil - floor). The round's headline DV on every arm.
• Netted dual-rubric F (bridge diagnostic; same normalization). The parent's netted contrast Delta = (judge score under the direction's target-value descriptor - score under its source-value descriptor)/100, floor/ceiling-normalized identically - computed on the same rollouts to quantify, on new data, how the netted form differs from the target-only primary and to bridge to the parent grid and the re-score round. Never the headline.
• Plain-descriptor mirror (secondary; 0-1). Movement of the plain-assistant descriptor score s_plain on the same rollouts: install asks whether plain-ness leaves, erase whether it returns.
• Activation fraction-of-swap F_act (continuous companion; signed, unitless). Signed projection of the patched-minus-floor answer-state shift onto the ceiling-minus-floor axis - the pair's OWN contrast direction in answer space - using span-mean V_a at read layer 26 with disjoint floor halves (experiments/issue2094/fmetrics.py::f_act / disjoint_half_means). Judge-independent; reported beside F_target with the same arms.
• Teacher-forced fixed positive-vs-negative completion margin (secondary; nats/token). Mean length-normalized lnP of the direction's fixed 4-completion target-side pool minus the source-side pool, under each patched state; the margin SHIFT is that quantity relative to the anchors. Validated via Spearman rho(margin shift, F_target) before carrying any read.
• Anchor separation (gate quantity). Target-descriptor separation ceil_X(c) - floor_X(c) per (rung x carrier) on the 0-1 scale, and the netted separation Delta-bar_ceiling - Delta-bar_floor on the max-2 scale, from the K=10 anchors; compared against the 0.25 and 0.5 gate bars respectively.
• Coherence rate (0-1). Fraction of a cell's draws with form-only coherence judge score > 60 (the inherited rubric); gate + disruption DV; all reported quantities are over coherent draws.
• Cap-hit fraction (0-1). Fraction of a cell's rows whose completion reached max_new_tokens (re-tokenized-length basis), against the 2% re-generation trigger.
• Per-conjunct descriptor scores (diagnostic; 0-100). Mean judge score per conjunct of the R1/R2 conjunct instruments (re-score round verbatim) on R1/R2 steered rollouts, beside the holistic target score - showing WHICH property moves when the holistic score moves.

Round manifest conditions (verbatim):

• Steered (donor persona state patch)
• Same-value donor null (same persona value, different carrier)
• Cross-type donor null (non-persona information type, construct-screened)
• Unpatched floor anchor
• Generate-under-donor ceiling anchor
• Install direction (plain to persona)
• Erase direction (persona to plain)
• Context-end slot
• Prefix-end slot
• Rung R1: named persona with pervasive dialect (Captain Marrow, pirate)
• Rung R2: named persona with episodic markers (Reginald, butler)
• Rung R3: unnamed register (warm, encouraging, informal)
• Rung R4: single trait (enthusiastic)
• Rung R5: situational user-persona, no style instruction (Lu et al. therapy persona)
• Rung R5: situational user-persona, no style instruction (Lu et al. philosophy persona)

Round manifest metrics (verbatim):

• target-descriptor fraction-of-swap (F_target)
• netted dual-rubric fraction-of-swap (parent-metric bridge diagnostic)
• plain-descriptor mirror score
• activation fraction-of-swap (F_act, projection onto the pair's own contrast direction)
• teacher-forced positive-vs-negative completion margin
• anchor separation (target descriptor, ceiling minus floor)
• coherence rate
• per-conjunct descriptor scores (R1/R2 steered diagnostic)

Dashboards (round tables):

• ladder_f_cells: https://htmlpreview.github.io/?https://raw.githubusercontent.com/superkaiba/explore-persona-space/2900cb7d3b0c1676940752fb1eef869bf0113f36/experiments/dashboards/issue2162_ladder_f_cells.html
• ladder_null_samevalue: https://htmlpreview.github.io/?https://raw.githubusercontent.com/superkaiba/explore-persona-space/2900cb7d3b0c1676940752fb1eef869bf0113f36/experiments/dashboards/issue2162_ladder_null_samevalue.html
• ladder_null_crosstype: https://htmlpreview.github.io/?https://raw.githubusercontent.com/superkaiba/explore-persona-space/2900cb7d3b0c1676940752fb1eef869bf0113f36/experiments/dashboards/issue2162_ladder_null_crosstype.html
• ladder_anchors: https://htmlpreview.github.io/?https://raw.githubusercontent.com/superkaiba/explore-persona-space/2900cb7d3b0c1676940752fb1eef869bf0113f36/experiments/dashboards/issue2162_ladder_anchors.html
• ladder_conjuncts: https://htmlpreview.github.io/?https://raw.githubusercontent.com/superkaiba/explore-persona-space/2900cb7d3b0c1676940752fb1eef869bf0113f36/experiments/dashboards/issue2162_ladder_conjuncts.html
• ladder_margin: https://htmlpreview.github.io/?https://raw.githubusercontent.com/superkaiba/explore-persona-space/2900cb7d3b0c1676940752fb1eef869bf0113f36/experiments/dashboards/issue2162_ladder_margin.html

Follow-up round: turn-boundary-multipatch — shared methodology

This round is a sibling follow-up on the PARENT grid's recency cells (it inherits from the original run above, not from the ladder round's bank). The single changed variable vs the parent grid is the patch POSITION SET: the parent measured single final-turn-anchored positions only (ce, pe); this round adds (a) a joint multi-position turn-boundary slot tb — the hidden state is replaced at EVERY assistant-turn end (<|im_end|> closing each assistant turn) plus context-end in one forward, all 28 layers, mode="replace" — and (b) a per-boundary single-position sweep tb_k (boundary k alone), on 7 already-banked parent cells, to separate two readings of the parent's single-position recency data: the trace weakening at every position as turns accumulate versus the trace remaining but spreading across per-turn boundaries so any single position carries a fraction. Inherited unchanged: frozen bank/carriers/values/donor assignment (bank.json, seed 2162), all-28-layer full-state replace, temp-1.0 K=5 grid draws, REUSED parent anchors (K=10; no anchor generation or judging this round), max_new_tokens = 2048, judge instrument, anchor-separation exclusion, survival floor 12, Wilcoxon+Holm+IUT protocol, pair-clustered bootstrap. New code: scripts/issue2162_tbmp.py (pod driver: boundary resolver, multi-position payload assembly, multi-position injection gate), scripts/issue2162_tbmp_dispatch.sh, scripts/issue2162_tbmp_analysis.py, scripts/issue2162_tbmp_figures.py; parent helpers imported, never re-implemented. One parent module edited: experiments/issue2094/hooks.py — the arm_batch replace-mode single-position guard is relaxed to per-row multi-position payload rows; single-position behavior is pinned bit-exact by unit tests (plus a joint-patch-equals-composition-of-single-patches test), and realized_mode is recorded per cell (expected constant replace_multi).

• Cells (7, parent frozen bank; 36 pairs per cell; no new contexts): instr_format + persona_prompted (depth-1 bases) and their recency cells at depths 3 and 5, plus recency_fact_user_name_d5 — the pre-registered near-zero control, joint-only. Boundary lists are resolved independently per pair side from each side's own token ids, paired by TURN NUMBER, with per-pair constant-offset asserts and a d1 assert that the resolved set equals exactly the parent ce position (at d1 the joint patch IS the parent's single-ce patch executed through the new code path). Boundary states captured for a 12-cell set (7 grid + 5 donor-pool cells; 432 contexts, one right-padded forward each). Designed boundary counts asserted per cell: d1 bases 1, d3 cells 3, d5 cells 5, control 6, donor-pool cells 2/4/6.
• Arms (nulls norm-matched PER POSITION per layer). (1) Steered — the pair's own B-side boundary states, paired by turn number. (2) Shuffled-donor null — the parent's frozen within-cell value-constrained donor assignment reused verbatim (same cell ⇒ identical turn structure, counts asserted equal), payload norm-matched row-wise at the same (turn, layer). (3) Cross-type-donor null (joint blocks only) — a donor pair from a DIFFERENT information type with the same-depth recency shape ({recency_prior_topic_dK, recency_fact_user_name_dK} at depth K; base cells at d1; the control draws solely from recency_prior_topic_d5, an exact 6-boundary structural match). Both pool members carry K+1 boundaries against the recipient's K, so a pre-registered alignment rule applies: donor LONGER ⇒ right-align (pair by turn number from the end, drop the donor's earliest boundaries; recorded per row); donor SHORTER ⇒ forbidden (runtime HALT). The shuffled null (structure-EXACT by construction) is the primary null everywhere; the cross-type leg's K+1→K right-alignment is a declared limitation of the IUT's cross-type leg. Sweep arms are steered + shuffled only (cross-type carried by the joint blocks).
• Block grid (realized = planned): joint tb 7 cells × 3 arms = 21 blocks + sweep 12 single-boundary variants × 2 arms = 24 blocks ⇒ 45 blocks, 8,100 rollouts (36 pairs × K=5 per block), generated by 4 CVD-pinned workers on pod-2162-tbmp (4× H100) over one shared claim queue, d1 blocks first. Pre-registered mechanical gates before any grid spend: the G1 boundary-resolution gate (count/offset/d1-identity/alignment-legality asserts over all 432 capture contexts) and the multi-position injection-exactness gate (12 spot pairs spanning all 3 arms × depths; installed state == intended donor at every intended (position, layer) and nowhere else; cosine ≥ 0.999, norm ratio ∈ [0.995, 1.005]).
• Realized generation integrity. Cap-hit: max 0.56% per cell, run-level 12/8,100 = 0.148% — below the pre-registered >2%-per-cell re-generation trigger (no re-generation round ran). Coherent fraction ≥ 94.4% in every cell (per-cell values in captions.json); zero low-coherence cells. 19 cells (7 joint + 12 sweep variants), all 19 testable — zero untestable-causal cells.
• Judge instrument + realized routing/integrity. Judge claude-sonnet-4-5-20250929, graded 0–100 reason-then-score, max_tokens = 1024, N = 1 judge draw per rubric per rollout (uncertainty carried at pair/cell level by the pair-clustered bootstrap, inherited from the parent); rubrics are the parent's content-hashed instruments verbatim (the pair's two value descriptors + form-only coherence; one behavior per call); drop-never-coerce with per-arm content/transport/api-refusal splits. Volume: 24,840 calls (pilot 540 + wave-1 3,240 + wave-2 21,060). Realized routing: wave-2's 7,020-call coherence wave went via the Anthropic Batch API; its nine behavior-rubric waves (six × 2,160 + three × 360) routed SYNC because judge_dispatch probed 2,000,000 output-tokens/min, raising the effective sync threshold to 10,000 — above each wave's size; projected call counts matched exactly, only the route split differed from the plan's all-Batch wave-2 phrasing. Realized integrity over the 24,300 scored production rows: zero truncation drops, 7 content drops (0.029%), zero transport loss, every wave complete.
• Pre-registered gates on the judge path (each named for what it tests; realized statistics live in the gate artifacts). G3 judge pilot gate — 540 sync draws (180 d1 rollouts, 60 per arm × 3 arms, each judged under all 3 production rubrics ⇒ 60 effective draws per (arm × rubric), clearing the 51-draw resolution floor for a 2% threshold) at the exact production instrument, gating all production judging on zero stop_reason == "max_tokens" and per-arm parse-fail < 2% per rubric (judge/gates/pilot_gate_tbmp.json). G2 d1-identity gate — after wave-1, before any wave-2 spend: per-arm (steered AND shuffled) paired per-pair ΔF in the parent's NETTED space between this round's joint tb@d1 and the parent's committed single-ce rows, pooled over the 66 surviving d1 pairs per arm (instr_format 36/36 + persona_prompted 30/36 under |separation| ≥ 0.5); gate bar |mean ΔF| ≤ 0.10 per arm (≈3.2 SE at the parent's measured per-pair dispersion σ_d ≈ 0.25), testing whether the extended multi-position replace path reproduces the parent's single-position read where the two are definitionally the same patch; a FAIL freezes wave-2 and every d3/d5 read. Realized gate statistics: identity_gate.json, quoted in the captions.json caption sidecar.
• Registered metric spaces + statistics. Netted dual-rubric F (Δ = (judge_B − judge_A)/100, floor/ceiling-normalized on the reused parent anchors) is the registered space for the instruction-format cells + the control; target-descriptor-only F (the judge_B channel alone) is the registered space for the two persona cells — the ladder round's instrument fix — with the netted form computed everywhere as the bridge diagnostic (both spaces from the same judge calls). Registered tests: family TB-joint (m = 7) — per cell, IUT p = max(p_shuffled, p_crosstype) from exact two-sided Wilcoxon signed-rank over per-pair paired differences, Holm-corrected within family at α = 0.05, "separates" additionally requiring fully disjoint pair-clustered bootstrap 95% CIs vs BOTH nulls with steered above; family TB-sweep (m = 12) — steered vs shuffled per single-boundary variant, same machinery. Bootstrap: pair-clustered, B = 10,000, seed 21620 (the registered battery in stats_tb.json; figure-locally rendered error bars use B = 10,000 at seed 21621). Raw-scale companion (denominator-free signed movement, judge-contrast units) reported over all AND surviving pair sets; TF margin secondary computed against the parent's frozen judge-filtered pools (pairs with unfillable pools inherit the parent behavior: margin N/A, reported).
• Registered verdict quantities + narration rules (pre-registered decision rules; outcomes live in stats_tb.json). Per base, in its registered space: J1 / J5 := the min-over-the-two-nulls pair-clustered bootstrap mean paired difference (steered − null) for the joint tb patch at depths 1 / 5; D := F_tb(d5) − 0.5·F_tb(d1); δ_disp := 0.5·Ĵ1 (the pre-registered dispersion effect). The verdict lattice (Dispersion / Partial-trace / Decay / No-verdict) is disjoint and exhaustive; d3 is a trend point, not a verdict input. The Decay power guard: Decay additionally requires J5's CI to EXCLUDE δ_disp — a J5 straddle whose upper bound ≥ δ_disp routes to No-verdict and is narrated as failure-to-reject (underpowered), never as decay. Two honesty overlays sit outside the lattice: the edit-artifact overlay (a control-cell joint tb clearing BOTH nulls on Holm-IUT + disjoint CIs downgrades every label to "edit-artifact — no verdict") and the denominator overlay (a Dispersion/Partial-trace label additionally reports the raw-scale steered movement at d5 vs d1 and carries a "denominator-limited" caveat when the raw-scale read contradicts the F-space direction).
• Realized deviations (recorded from artifacts):
  1. The judge_completeness rollup inside stats_tb.json reports n_items: null / frac_items_complete: null for every wave — the rollup reads a top-level n_items field while the per-wave meta nests it under regime.n_items; n_scored_items is correct and the true per-wave fractions are recoverable from the per-wave .meta.json files under judge/scores/.
  2. The out-root residue upload glob was over-broad (fnmatch * crosses /), landing 102 duplicate files under the manifests/outroot/ HF prefix — duplication only; no consumer reads that prefix.
  3. Realized VM staging was 198.4 MB against the plan §9 estimate of ~35 MB.
• Scope caveats (methodology facts, carried into the round's results sections):
  1. A joint-null verdict rules out dispersion across the PATCHED BOUNDARY SET only — a trace living at non-boundary positions yields the same label, so any decay-side reading is scoped as "no usable trace at the turn-boundary position set", never unqualified decay prose.
  2. The No-verdict narration rule above is pre-registered, not post-hoc: wherever a J5 CI straddles zero with an upper bound ≥ δ_disp, the registered narration is failure-to-reject (underpowered), never decay.
  3. Figure namespace: this round's figures land flat under figures/issue_2162/ with the tb_ prefix (not a subdirectory namespace); round tables land under eval_results/issue_2162/turn_boundary/.
• Artifacts (this round). Round tables at eval_results/issue_2162/turn_boundary/: f_cells_tb.jsonl, null_shuffled_cells_tb.jsonl, null_crosstype_cells_tb.jsonl, parent_ref_cells_tb.jsonl (the parent single-ce reference re-aggregated into BOTH spaces from the parent's own per-rubric judge scores, with a per-row parity assert that the recomputed netted F reproduces the committed parent f_beh), margin_cells_tb.jsonl, rawscale_tb.json, stats_tb.json, identity_gate.json, captions.json, plus judge/scores/*.{scores.jsonl,meta.json} (24,300 scored rows) and judge/gates/pilot_gate_tbmp.json. Parent reference tables: eval_results/issue_2162/f_metrics/. Pod-side stores on the HF data repo under issue2162_ctxinfo/raw_completions/tbmp/grid/ and issue2162_ctxinfo/analysis_tensors/tbmp/{tb_bank, margin, gates, manifests}/. Code SHA for this round's results: 3f4a9c429e7b2b309ff753fcdde5a50c96c0ed60 (branch issue-2162-tbmp).

Metrics: (every quantity a round figure shows)

• Netted dual-rubric F at the turn-boundary set (F_tb; unitless, 0 = unpatched floor, 1 = full context swap). The parent's dual-rubric contrast, floor/ceiling-normalized on the reused parent anchors — the registered primary for the instruction-format cells + the control.
• Target-descriptor-only F (persona cells' registered primary; same normalization). The judge_B channel alone; the netted form is kept everywhere as the bridge diagnostic on the same rollouts.
• Raw-scale signed movement (judge-contrast units, range [−2, +2]; denominator-free). (Δ_patched − Δ_floor)·sign(Δ_ceiling − Δ_floor); reported over all and surviving pair sets.
• d1 identity gap (netted; unitless). The per-pair and per-arm mean ΔF between this round's joint tb@d1 and the parent's single-ce read — the G2 gate quantity.
• Per-boundary steered − shuffled ΔF (registered space; unitless). The sweep's paired per-pair difference at boundary index k.
• Teacher-forced fixed positive-vs-negative completion margin (secondary; nats/token). Against the parent's frozen judge-filtered pools; pairs with unfillable pools carry margin N/A, reported.
• Coherence rate (0–1). Form-only judge score > 60 (inherited rubric); gate + all reported quantities over coherent draws.
• Cap-hit fraction (0–1). Re-tokenized-length basis, against the 2% re-generation trigger.
• Realized patch mode per cell (realized_mode). Expected constant replace_multi under the extended multi-position replace hook; any realization drift is report-visible.

Round manifest conditions (verbatim):

• Joint all-boundary patch (tb: every assistant-turn end plus context-end)
• Single-boundary patch at boundary k (per-boundary sweep)
• Turn-boundary shuffled-donor null (per-position norm-matched)
• Turn-boundary cross-type donor null (structure-matched donor)
• Parent single context-end patch (reused reference rows)
• Near-zero control cell (user-name fact, depth 5)
• Depth 1 / 3 / 5 recency cells (instruction-format base)
• Depth 1 / 3 / 5 recency cells (prompted-persona base)

Round manifest metrics (verbatim):

• netted dual-rubric fraction-of-swap at the turn-boundary set (F_tb)
• target-descriptor-only fraction-of-swap (persona cells registered primary)
• raw-scale signed movement (judge-contrast units, denominator-free)
• d1 identity gap (tb at depth 1 minus parent context-end F)
• per-boundary steered minus shuffled-null F difference
• realized patch mode per cell (realized_mode; expected constant replace_multi under the extended multi-position replace hook)

Results

Per-type fraction-of-swap (behavior)

Methodology

• Per (type-cell × slot × arm): pairs failing the pre-registered anchor-separation exclusion (|ceiling − floor| < 0.5) are dropped; per surviving pair, F_beh is the pair mean over its K=5 coherent draws; the arm statistic is the mean over surviving pairs with a pair-clustered bootstrap 95% CI (B = 10,000, seed 21620). A pair with no coherent draws in ANY arm is dropped from all three arms of that cell, so the three arms of a cell share one denominator (affects exactly one pair, language_implied at pe).
• Three arms per type (steered / shuffled-donor null / cross-type-donor null), one panel per slot (ce, pe); post-exclusion n stated per type. Sources: eval_results/issue_2162/f_metrics/{f_cells,null_shuffled_cells,null_crosstype_cells,anchors}.jsonl.
• filler_swap reports no F anywhere (no ceiling by construction); the two pre-declared degenerate prefix-end cells are excluded from the Holm families (they are still rendered as per-type bars, carrying a degeneracy flag).
• Rendered figure (plotter caption): Mean behavior fraction-of-swap F_beh (unitless; y) per type-cell (x, post-exclusion pair count in parentheses) for the steered, shuffled-donor-null, and cross-type-donor-null arms, one panel per readout slot (context-end top, prefix-end bottom). Error bars are pair-clustered bootstrap 95% CIs (B=10,000, seed 21620); pairs with anchor separation |ceiling - floor| < 0.5 are excluded, cells with zero surviving pairs are marked n/a, and 14 of 76 (cell x slot) combinations are untestable-causal (post-exclusion n < 12). The persona_role_header bars rest on n=1 surviving pair per slot (0.995 at context-end, -0.397 at prefix-end).

Takeaways

(Thomas fills in)

Per-type F_beh - per-pair companion

Methodology

• The per-unit companion of the hero figure: same exclusion, NO aggregation — one point per surviving pair per arm (pair mean over K=5 coherent draws), labeled by pair id, per type × slot. Source: the same three per-cell row files.
• Rendered figure (plotter caption): Per-pair F_beh points behind the aggregate bars: one point per separation-surviving pair per arm (pair mean over K=5 coherent draws), horizontally offset by arm, labeled by pair id, per type-cell, one panel per slot. Same |separation| >= 0.5 exclusion as the aggregate view; no aggregation and no error bars (each point is one pair). n = 5,855 surviving pair-rows (1,952 steered, 1,951 shuffled-donor null, 1,952 cross-type null).

Takeaways

(Thomas fills in)

Read x write 2x2 (probe AUC vs causal F)

Methodology

• x = max-over-28-layers macro AUC per (type × slot) from the linear read probe under leave-one-carrier-out folds; y = steered mean F_beh (separation-excluded, as in per_type_f_beh). Sources: f_metrics/probe.json + f_metrics/f_cells.jsonl.
• Quadrant thresholds are the REGISTERED verdicts, not free parameters: causal-positive ⇔ clears BOTH nulls on fully disjoint pair-clustered 95% CIs AND survives the family Holm IUT; probe-positive ⇔ the max-AUC clears the max-selected label-permutation 97.5th-percentile band (selection-symmetric: every permutation draw is re-maxed over layers).
• Cells with post-exclusion n < 12 (the pre-registered survival floor) carry the explicit fifth label untestable-causal — the causal test is unreachable there, so they are never rendered as a causal verdict. Their probe AUC is plotted only where a steered F_beh y-value exists (≥ 1 surviving steered pair): per the realized figure caption, 8 untestable-causal cells are plotted while 6 further untestable cells with ZERO surviving steered pairs are omitted from the panel and counted in the title (figures/issue_2162/captions.json, two_by_two).
• Rendered figure (plotter caption): One point per (type-cell x slot) placing the linear read-probe max-over-28-layers macro AUC (x, leave-one-carrier-out folds) against the steered mean F_beh (y, separation-excluded); marker and color encode the persisted (probe verdict, causal verdict) quadrant, with untestable-causal (post-exclusion n < 12) as the explicit fifth class. Plotted class counts: 5 stored-and-used, 55 stored-but-unusable, 0 used-but-not-decoded, 2 absent, 8 untestable-causal; 6 further untestable cells with zero surviving steered pairs are omitted and counted in the title. No single vertical read-threshold line is drawn because the probe-positive threshold is the per-cell max-selected permutation band (97.5th percentile).

Takeaways

(Thomas fills in)

Probe AUC per layer - per-unit companion of the 2x2

Methodology

• The per-unit companion of the 2×2: per (type × slot × layer), macro AUC over the 3 value-pairs under carrier folds, with the label-permutation band (B = 1,000, within-carrier shuffles) — no max taken in this view (full-curve display; the max-selected band's upper bound is annotated where the 2×2's verdict is drawn). Sources: f_metrics/probe.json + the persisted per-draw × per-layer matrix issue2162_ctxinfo/analysis_tensors/probe_perm_matrix/.
• Rendered figure (plotter caption): Companion heatmap of the same probe read: rows are type-cells, columns are layers 0-27, color is the leave-one-carrier-out macro AUC, one panel per slot (context-end left, prefix-end right). Same data as the per-layer curve panels; n = 78 (cell x slot) units.

Takeaways

(Thomas fills in)

Stage-2 layer profile (post-selection)

Methodology

• Post-selection stage-2 read over the 5 realized survivor (cell × slot) combos: per survivor, mean F_beh over surviving pairs at each (layer ∈ {8, 12, 14, 16, 19, 22, 26} × dose ∈ {1, 4}), pair-difference add-mode edits (Δ = V(B) − V(A), single layer, dose = alpha multiplier), 1 greedy draw per pair, steered with the shuffled-donor-null companion. Source: f_metrics/stage2_cells.jsonl (from the 140 stage-2 shards).
• Rendered as a type × layer heatmap at each dose, labeled post-selection everywhere (the survivors were selected on stage-1 outcomes; this figure never carries an unbiased-estimate claim).
• Rendered figure (plotter caption): Stage-2 heatmaps for the 5 stage-1-selected survivor cells (rows; all at the context-end slot): mean F_beh over separation-surviving pairs at each patch layer in {8, 12, 14, 16, 19, 22, 26} (columns) and dose (1 left, 4 right), greedy 1 draw per pair. The top row is the steered arm; the bottom row is steered minus the shuffled-donor null. Post-selection: these cells were chosen by the stage-1 selection rule, so the values are confirmation reads, not unbiased estimates.

Takeaways

(Thomas fills in)

Stage-2 per-pair companion

Methodology

• Per-unit companion: no aggregation — per-pair stage-2 F_beh points at each survivor's best (layer, dose), steered and null interleaved, labeled by pair id. Source: f_metrics/stage2_cells.jsonl.
• Rendered figure (plotter caption): Per-pair stage-2 F_beh points at each survivor cell's best (layer, dose) by steered mean (annotated above each column), steered (blue) and shuffled-donor null (grey) interleaved, pair-id labeled, separation-surviving pairs only. Post-selection: the 5 survivor cells were chosen by the stage-1 selection rule. No aggregation; each point is one pair at greedy 1 draw.

Takeaways

(Thomas fills in)

Route contrasts at matched content

Methodology

• Steered mean F_beh (separation-excluded, pair-clustered bootstrap CI) for the matched-content route sets — {instr_format vs demo_format}, {persona_prompted vs demo_persona vs persona_role_header}, {instr_language vs language_implied} — per slot: the same content induced by different routes, isolating the route with content held fixed.
• Plus the 4 conflict cells' balance shift = (judge_demo − judge_instr)/100, floor/ceiling-normalized: the DV for which route wins when instruction and demonstration disagree (both directions of both conflicts). Sources: f_metrics/{f_cells,null_shuffled_cells,null_crosstype_cells}.jsonl.
• persona_role_header's varied span is immediately adjacent to the context-end slot — a designed boundary-content property, pre-declared in plan §4.1.
• Rendered figure (plotter caption): Steered mean F_beh for each matched-content route set - base type (blue) beside its route variant or conflict cell (orange) - per slot; route sets: instr_format vs demo_format, persona_prompted vs demo_persona and persona_role_header, instr_language vs language_implied, plus the four conflict cells, whose F_beh equals the normalized conflict balance shift by construction (value_b = the demonstration-carried value). Error bars are pair-clustered bootstrap 95% CIs over separation-surviving pairs; bars with zero surviving pairs are absent (demo_format, demo_persona), and persona_role_header rests on n=1 surviving pair per slot.

Takeaways

(Thomas fills in)

Route contrasts - per-pair companion

Methodology

• Per-unit companion: per-pair F_beh / balance-shift points for every route-variant and conflict cell, labeled by pair id. Source: f_metrics/f_cells.jsonl.
• Rendered figure (plotter caption): Per-pair steered F_beh points for every route-variant and conflict cell beside its base type, pair-id labeled, per slot, with no separation exclusion and no aggregation. The top panel shows the full range; the bottom panel shows the same points restricted to |F_beh| <= 2, since a small number of separation-degenerate pairs reach |F_beh| of roughly 100 and otherwise set the shared scale. n = 750 scored steered pair-rows across the 11 route/conflict cells (2,574 across all cells).

Takeaways

(Thomas fills in)

Recency and load curves

Methodology

• Per crossed type: steered mean F_beh (separation-excluded) at depth {1, 3, 5} (recency: same information introduced at turn 1 followed by depth−1 neutral padding exchanges identical between A and B) and at load {1, 3, 5} (load: load-many same-type pieces in the prefix, one designated target varying), depth/load 1 = the base cell; pair-clustered bootstrap CIs, null band behind.
• The registered secondary read is the pair-clustered bootstrap CI on the per-pair depth/load slope (trend reads are secondary descriptives, per plan §6). Sources: f_metrics/{f_cells,null_shuffled_cells}.jsonl.
• Rendered figure (plotter caption): Steered mean F_beh (y) versus conversation history depth (left) and distractor load (right) at levels {1, 3, 5}, where level 1 is the uncrossed base cell, one curve per crossed type per slot (context-end solid, prefix-end dashed); error bars are pair-clustered bootstrap 95% CIs over separation-surviving pairs. Grey bands are the shuffled-donor null's 95% CIs with the null mean line inside. The in-panel text lists the registered per-pair depth/load slope with its pair-clustered bootstrap 95% CI and pair count.

Takeaways

(Thomas fills in)

Recency/load - per-pair companion

Methodology

• Per-unit companion: per-pair F_beh at each depth/load level, lines connecting the same pair across levels. Source: f_metrics/f_cells.jsonl.
• Rendered figure (plotter caption): Per-pair steered F_beh trajectories across depth (left) and load (right) levels; lines connect the same (carrier x value-pair) across levels, colored by crossed type (context-end solid, prefix-end dashed), with no separation exclusion and no aggregation. The top row shows the full range; the bottom row shows the same trajectories restricted to |F_beh| <= 2, since single separation-degenerate pairs reach approximately +24 and -8 and otherwise set the scale.

Takeaways

(Thomas fills in)

Anchor separation per pair (pre-exclusion)

Methodology

• Per pair: ceiling mean − floor mean (judge-contrast units) from the K=10 anchor draws, shown as a histogram/strip per type with the 0.5 exclusion bar drawn and pre-/post-exclusion counts per type. Source: f_metrics/anchors.jsonl.
• This is the exclusion instrument made visible: the pre-registered bar, its per-type bite, and the pre-committed selection caveat (the bar selects on a K=10 estimate) ship with this figure.
• Rendered figure (plotter caption): Per-pair anchor separation (y: generate-under-donor ceiling mean minus unpatched floor mean, judge-contrast units, K=10 anchor draws per side) per type-cell (x), before any exclusion; red dotted lines mark the pre-registered |separation| >= 0.5 exclusion bar and kept/total counts per type are printed above. Three types keep 0/36 pairs (demo_format, demo_persona, recency_prior_topic_d5) and persona_role_header keeps 1/36. n = 1,368 pair rows (38 grid types x 36 pairs).

Takeaways

(Thomas fills in)

F_act vs F_beh agreement

Methodology

• Per (type-cell × slot × arm): mean F_act (read layer 26, disjoint floor halves) against mean F_beh over the same surviving pairs, one point per cell-arm, colored by arm; Spearman ρ across cells with dynamic range reported in-panel (the rank-agreement diagnostic between the judge-scored and the judge-free level — the dynamic-range screen mirrors the rule-19 validation convention). Sources: f_metrics/{f_cells,null_shuffled_cells,null_crosstype_cells}.jsonl.
• Rendered figure (plotter caption): Mean activation fraction-of-swap F_act (x; read at layer 26 with disjoint floor halves) versus mean behavior F_beh (y) per (cell x slot x arm) over separation-surviving pairs, colored by arm (steered blue, shuffled-donor null grey, cross-type null red), point labels are cell|slot. The per-arm Spearman rho in the legend is computed over the n=66 units passing the dynamic-range screen stated in the panel; the 4 screened-out units per arm render as hollow circles and carry no rho weight. Realized F_act and F_beh ranges of the screened units are printed in the panel.

Takeaways

(Thomas fills in)

Teacher-forced margin vs F_beh (dual-DV validation)

Methodology

• Per (type-cell × slot): mean teacher-forced margin shift (steered − floor, length-normalized lnP over the fixed 4+4 judge-filtered pools) against mean F_beh — the registered rule-19 validation scatter, whose ρ across cells with dynamic range must be reported BEFORE the margin carries any cross-condition read. Sources: f_metrics/{f_cells,margin_cells}.jsonl; pool coverage per judge/gates/pools_report.json (100/111 keys; 3 query_content keys skipped by design — no fixed shared answer pool is well-defined when the query itself varies — with explicit skip rows recorded).
• Rendered figure (plotter caption): Per-(cell x slot) mean teacher-forced fixed-pool margin shift (x: steered minus floor anchor, length-normalized log-probability over the fixed 4+4 completion pools) versus mean F_beh (y), blue = the registered rule-19 grain (n=72 cells passing the dynamic-range screen), grey = the per-pair companion points (n=2,308); Spearman rho and p for both grains are in the title (per-cell rho=0.458, p=5.1e-05; per-pair rho=0.334, p=3.6e-61). The single high-leverage cell icl_task_mapping|ce (margin shift 3.88 against a next-largest of about 0.35) is plotted and labeled; the inset re-renders the remaining 71/72 cells at |margin shift| < 1 and |F_beh| < 3. The persona_role_header means in this grain (1.48 at context-end, -2.72 at prefix-end, 12 pairs each, computed without the anchor-separation exclusion) lie outside the [0, 1] fraction-of-swap range, i.e. the patched value fell outside the floor-to-ceiling anchor range for that thin cell.

Takeaways

(Thomas fills in)

Cross-type null by donor type + shuffled-donor null by donor value (diagnostic)

Methodology

• Diagnostic breakout of the null arms by donor identity (pure re-reductions of the persisted null tables — donor type, pair, and value ids are recorded per row): (a) cross-type facet — per (recipient cell × slot), mean F_beh of the cross-type-donor arm split BY DONOR TYPE, rendered for every cell whose pooled cross-type-null bootstrap CI excludes 0 (a donor type whose content moves the recipient's rubric inflates the pooled null; conservative for positives, made visible rather than assumed away); (b) shuffled-donor facet — for the two graded/ordinal value sets only (refusal_boundary, constraint_knowledge), the shuffled-donor arm split BY DONOR VALUE (an adjacent-value donor can earn partial rubric credit and elevate that null). Sources: f_metrics/{null_crosstype_cells,null_shuffled_cells}.jsonl.
• Rendered figure (plotter caption): Top: cross-type-donor-null F_beh split by donor type (bar = mean over separation-surviving rows; donor type and row count annotated on each bar) for the 24 recipient (cell x slot) combinations whose pooled cross-type-null bootstrap 95% CI excludes 0. Bottom: shuffled-donor-null F_beh split by donor value for the two ordinal value sets (refusal_boundary, constraint_knowledge) at cells where either null's pooled CI excludes 0. Diagnostic view: per-donor means without CIs; per-donor row counts are annotated.

Takeaways

(Thomas fills in)

Coherence and cap-hit diagnostics

Methodology

• Per (type-cell × slot × arm), three panels: excess incoherence = incoherent fraction (score ≤ 60) minus the anchor baseline rate (top); cap-hit fraction (completion length at max_new_tokens, re-tokenized basis) with the pre-registered 2% re-generation trigger drawn (middle); and post-exclusion pair count with the n = 12 untestable-causal floor line drawn, out of 36 pre-exclusion pairs per cell (bottom). Sources: f_metrics/{f_cells,null_shuffled_cells,null_crosstype_cells,anchors}.jsonl.
• Excess incoherence doubles as the ONLY registered read for filler_swap (the disruption control has no ceiling, hence no F); realized run-level cap-hit was 42/42,120 rows (0.0997%), below the trigger, so no re-generation round exists in this figure's data.
• Rendered figure (plotter caption): Per (cell x slot x arm): excess incoherence (top; incoherent fraction at judge score <= 60 minus the cell's anchor-baseline incoherent rate), cap-hit fraction (middle; completions ending at the generation length cap, dotted line = the pre-registered 2% re-generation trigger), and post-exclusion pair count (bottom; dotted line = the n=12 untestable-causal floor, out of 36 pre-exclusion pairs per cell). Arm colors are fixed across panels (steered blue, shuffled-donor null grey, cross-type null red). Denominators are the per-(cell x slot x arm) draw counts pooled over pairs.

Takeaways

(Thomas fills in)

Specificity ladder: steered vs both nulls per rung

Methodology

• Cells: per (direction x slot x arm x rung), pair value = mean F_target over the K=5 coherent draws per gate-surviving carrier, from f_cells.jsonl + null_samevalue_cells.jsonl + null_crosstype_cells.jsonl; gate-failing rungs/carriers are absent by construction (they generated nothing).
• Aggregation: rung mean over gate-surviving carriers; carrier-clustered bootstrap 95% CI (B=10,000, seed 21626).
• Panels: install and erase; context-end primary with a prefix-end companion panel; x-axis rungs R1 -> R5.
• Three arms per rung (steered / same-value null / cross-type null), one color = one meaning (the module's fixed arm palette).
• Annotations: per-rung n (surviving carriers) and per-rung system-prompt token count (Qwen tokenizer, no special tokens).
• What is plotted: Mean target-descriptor fraction-of-swap F_target (unitless; 0 = unpatched floor, 1 = full context swap) per persona rung, split into four panels by direction (install, erase) and slot (context-end, prefix-end), with three arms per rung: steered donor patch, same-value-donor null, and cross-type-donor null. Each point is the mean over gate-surviving carriers (per-carrier value = mean F_target over K=5 coherent draws); error bars are carrier-clustered bootstrap 95% CIs (B=10,000, seed 21626). n = 6 carriers per rung for R1-R3 and n = 4 for R5 (philosophy); R4 (trait) and R5 (therapy) failed the anchor-separation gate and appear as empty labeled columns, and x tick labels carry per-rung n and system-prompt token counts.
• Sources: eval_results/issue_2162/persona_specificity_ladder/f_cells.jsonl, eval_results/issue_2162/persona_specificity_ladder/null_samevalue_cells.jsonl, eval_results/issue_2162/persona_specificity_ladder/null_crosstype_cells.jsonl, eval_results/issue_2162/persona_specificity_ladder/stats.json

Takeaways

(Thomas fills in)

Specificity ladder - per-carrier companion

Methodology

• Same input rows and exclusions as ladder_hero, NO aggregation: one point per gate-surviving (carrier x arm x rung x direction x slot).
• Points are carrier-labeled and arm-offset within each rung position; same arm palette.
• Serves as the mandatory low-level per-unit companion to the hero's rung means.
• What is plotted: Per-carrier F_target (unitless) with no aggregation: one point per gate-surviving (carrier x arm x rung x direction x slot) cell, each labeled with its carrier id, arms offset horizontally within each rung. Panels split direction (install, erase) x slot (context-end, prefix-end); colors are the three arms (steered, same-value-donor null, cross-type-donor null). n = 264 cells: 22 gate-surviving rung x carrier pairs (6 carriers each for R1-R3, 4 for R5 philosophy) x 2 directions x 2 slots x 3 arms.
• Sources: eval_results/issue_2162/persona_specificity_ladder/f_cells.jsonl, eval_results/issue_2162/persona_specificity_ladder/null_samevalue_cells.jsonl, eval_results/issue_2162/persona_specificity_ladder/null_crosstype_cells.jsonl

Takeaways

(Thomas fills in)

Erase vs install asymmetry per rung

Methodology

• Input: STEERED rows only of f_cells.jsonl; per (rung x slot x carrier), the paired difference erase F_target - install F_target between the two directions sharing that carrier and slot.
• Display: per-carrier paired points plus the rung mean with carrier-clustered bootstrap 95% CI (B=10,000, seed 21626) - the H4 registered contrast.
• Positive values mean erasure moved a larger fraction of the anchor span than installation at the same rung.
• What is plotted: Paired-by-carrier steered-arm difference in F_target, erase minus install (unitless; positive = larger erase value), per gate-surviving rung, one panel per slot (context-end, prefix-end). Grey circles are the per-carrier paired differences (n = 6 carriers for R1-R3, n = 4 for R5 philosophy); red diamonds are rung means with carrier-clustered bootstrap 95% CIs (B=10,000, seed 21626).
• Sources: eval_results/issue_2162/persona_specificity_ladder/f_cells.jsonl, eval_results/issue_2162/persona_specificity_ladder/stats.json

Takeaways

(Thomas fills in)

Anchor separations and gate verdicts per rung

Methodology

• Input: anchors.jsonl (per rung x carrier: coherent-draw floor/ceiling means and separations) + the gate verdict JSON gates/ladder_separation_gate.json.
• Quantities: per-carrier target-descriptor separation (ceiling minus floor, 0-1 scale) against the 0.25 bar, and netted dual-rubric separation against the 0.5 bar (max-2 scale), from the K=10 anchor draws.
• Per-carrier pass/fail verdicts and the >= 4-of-6 rung survival verdict are marked; ALL rungs render, gate failures included (the gate's attrition is the figure's subject, not an exclusion).
• What is plotted: Per-(rung x carrier) anchor separation — generate-under-donor ceiling minus unpatched floor, from K=10 anchor draws per context — for all six rungs, 6 carriers each (n = 36 points per panel). Left panel: target-descriptor separation (0-1 scale) against the 0.25 gate bar; right panel: netted dual-rubric separation (0-2 scale) against the 0.5 bar (red dashed lines). Filled circles are carriers that passed both gate bars, open circles are carriers that failed; x labels carry each rung's survived/dropped verdict (survival requires >= 4 passing carriers).
• Sources: eval_results/issue_2162/persona_specificity_ladder/anchors.jsonl, eval_results/issue_2162/persona_specificity_ladder/judge/gates/ladder_separation_gate.json

Takeaways

(Thomas fills in)

Netted dual-rubric F vs target-only F (rubric-artifact quantification)

Methodology

• Input: all three cell tables; per (direction x slot x arm) cell, x = the netted dual-rubric F (the parent-metric convention), y = F_target (this round's primary), computed on the SAME rollouts under the same floor/ceiling normalization.
• Arms color-coded; identity line drawn.
• Recipe purpose: quantifies on new data how the two metric conventions relate per arm - the instrument-bridge back to the parent grid and the re-score round's decomposition.
• What is plotted: F under the two metric conventions, one point per (direction x slot x arm) group: x = netted dual-rubric fraction-of-swap (the parent round's metric convention), y = target-descriptor-only F_target (this round's primary), each coordinate the mean over that group's gate-surviving carrier cells. n = 48 groups (8 gate-surviving direction ids x 2 slots x 3 arms); colors are the three arms and the dotted line is the identity y = x.
• Sources: eval_results/issue_2162/persona_specificity_ladder/f_cells.jsonl, eval_results/issue_2162/persona_specificity_ladder/null_samevalue_cells.jsonl, eval_results/issue_2162/persona_specificity_ladder/null_crosstype_cells.jsonl

Takeaways

(Thomas fills in)

Continuous-companion agreement: F_act vs F_target and margin validation

Methodology

• Panel 1: per-cell scatter of F_act (y; pair-own contrast projection, span-mean V_a, layer 26, disjoint floor halves) vs F_target (x), all three arms in the arm palette, pooled Spearman rho annotated.
• Panel 2: the pre-registered margin validation - per-cell TF margin shift (nats/token, x) vs per-cell F_target mean (y) with the registered per-cell Spearman rho annotated (rule19_validation output; cells lacking a pool carry explicit skip rows and are absent).
• Panel 3: per-cell coherence rate vs cap-hit fraction, all arms - the generation-health companion.
• What is plotted: Continuous-companion panels over the 264 (direction x slot x arm x carrier) cells. Panel 1: activation fraction-of-swap F_act (signed projection onto the pair's own contrast direction, unitless) vs F_target per cell, colored by arm, Spearman rho = +0.42 (n = 264 cells); panel 2: per-unit mean F_target vs teacher-forced fixed-pool margin shift (nats/token) over the 14 steered (direction x slot) units with nonzero dynamic range (2 of 16 units dropped by the pre-registered dynamic-range screen), Spearman rho = +0.54; panel 3: cap-hit fraction vs coherence rate per cell (observed ranges 0.00-0.20 and 0.60-1.00; n = 264 cells).
• Sources: eval_results/issue_2162/persona_specificity_ladder/f_cells.jsonl, eval_results/issue_2162/persona_specificity_ladder/null_samevalue_cells.jsonl, eval_results/issue_2162/persona_specificity_ladder/null_crosstype_cells.jsonl, eval_results/issue_2162/persona_specificity_ladder/margin.jsonl, eval_results/issue_2162/persona_specificity_ladder/stats.json

Takeaways

(Thomas fills in)

Per-conjunct decomposition on R1/R2 steered installs

Methodology

• Input: conjuncts.jsonl - judge scores from the re-score round's conjunct instrument run verbatim on R1/R2 STEERED rollouts only.
• Quantity: mean 0-100 judge score per conjunct descriptor, rendered beside the holistic target-descriptor score for the same rollouts, per direction x slot.
• Diagnostic only: shows which conjunct property moves when the holistic score moves; no verdict or gate consumes it.
• What is plotted: Mean judge score (0-100) per conjunct of the persona description on R1 (pirate) and R2 (butler) steered rollouts only, one panel per rung, bars grouped by direction x slot (erase/install x context-end/prefix-end) with the holistic target-descriptor score as the rightmost group. Each bar is the mean over 6 carriers of per-carrier mean scores over coherent draws (Round-A conjunct instrument verbatim; R1 conjuncts: dialect, sea, warning; R2 conjuncts: address, courtesy, formality, household). Bars at zero are true zero mean judge scores in conjuncts.jsonl, not missing cells.
• Sources: eval_results/issue_2162/persona_specificity_ladder/conjuncts.jsonl, eval_results/issue_2162/persona_specificity_ladder/f_cells.jsonl

Takeaways

(Thomas fills in)

Joint turn-boundary patch vs single context-end across depth

Methodology

• Input: this round's joint-tb rows of f_cells_tb.jsonl (steered) + null_shuffled_cells_tb.jsonl + null_crosstype_cells_tb.jsonl, plus the parent single-ce steered reference from parent_ref_cells_tb.jsonl (persona-cell reference points re-aggregated into the target-descriptor-only space from the parent's own per-rubric judge scores — never lifted netted into a target-only panel).
• Per (cell x arm): pair value = mean F over K=5 coherent draws in the cell's registered space (netted dual-rubric for instruction-format cells, target-descriptor-only for persona cells); surviving pairs only (|separation| >= 0.5 on the reused parent anchors; 36 pairs per cell before exclusion).
• Aggregation: cell mean with pair-clustered bootstrap 95% CI (B=10,000); per-pair points rendered behind each errorbar (the per-unit companion embedded).
• What is plotted: Mean fraction-of-swap F at the joint turn-boundary set (unitless; 0 = unpatched floor, 1 = full context swap), one panel per base (instruction-format, prompted-persona), x = depth (d1/d3/d5), three colored arms per depth (steered / shuffled-donor null / cross-type-donor null) plus the parent's single context-end steered read in black at each depth.
• Sources: eval_results/issue_2162/turn_boundary/f_cells_tb.jsonl, eval_results/issue_2162/turn_boundary/null_shuffled_cells_tb.jsonl, eval_results/issue_2162/turn_boundary/null_crosstype_cells_tb.jsonl, eval_results/issue_2162/turn_boundary/parent_ref_cells_tb.jsonl

Takeaways

(Thomas fills in)

Per-boundary localization sweep (steered minus shuffled null)

Methodology

• Input: sweep rows (slots tbk1..tbk{n-1}) of f_cells_tb.jsonl + null_shuffled_cells_tb.jsonl; parent single-ce reference rows (both arms) from parent_ref_cells_tb.jsonl.
• Quantity: per single-boundary variant k, the per-pair PAIRED difference steered F − shuffled-null F (registered space per base) over surviving pairs; mean + pair-clustered bootstrap 95% CI; per-pair points behind each errorbar.
• What is plotted: Mean steered-minus-shuffled ΔF per boundary index k, one panel per sweep cell (instruction-format d3/d5, prompted-persona d3/d5; 12 single-boundary variants total, 2+4+2+4), x = boundary index k (1 = earliest assistant-turn end); the diamond at the final boundary k = n_d is the parent's single-ce paired difference (parent-run provenance labeled; persona points re-aggregated into the target-descriptor-only space).
• n basis: 36 pairs per cell before exclusion; sweep arms are steered + shuffled only (the cross-type null is carried by the joint blocks).
• Sources: eval_results/issue_2162/turn_boundary/f_cells_tb.jsonl, eval_results/issue_2162/turn_boundary/null_shuffled_cells_tb.jsonl, eval_results/issue_2162/turn_boundary/parent_ref_cells_tb.jsonl

Takeaways

(Thomas fills in)

Raw-scale (un-normalized) movement at the turn-boundary set

Methodology

• Input: this round's rawscale_tb.json + the parent's committed recency_rawscale.json; every bar sourced from the declared artifacts, never recomputed from the F tables.
• Quantity: mean signed raw movement (Δ_patched − Δ_floor)·sign(Δ_ceiling − Δ_floor) in judge-contrast units (range [−2, +2]) per arm x depth; solid bars = joint tb (this round), hatched bars = parent single-ce; errorbars = pair-clustered bootstrap 95% CIs (B=10,000, seed 21620) where the artifact provides them (the parent file carries steered CIs only).
• What is plotted: Signed raw behavioral movement per base x depth x arm in a 2x2 grid — rows = all scored pairs vs surviving pairs only, columns = base. The dual all/surviving reporting is inherited: the anchor-separation exclusion selects on the F denominator, so the raw companion is reported on both pair sets.
• Role: the denominator-free companion — F normalizes by the per-pair anchor gap; this view removes that normalization.
• Sources: eval_results/issue_2162/turn_boundary/rawscale_tb.json, eval_results/issue_2162/f_metrics/recency_rawscale.json

Takeaways

(Thomas fills in)

Depth-1 identity gate: tb reproduces the parent context-end read

Methodology

• Input: joint-tb d1 rows of f_cells_tb.jsonl + null_shuffled_cells_tb.jsonl joined per pair to the parent's committed f_cells.jsonl + null_shuffled_cells.jsonl ce rows; gate statistics from identity_gate.json.
• What is plotted: Per-pair scatter — x = the parent's single-ce pair F, y = this round's joint-tb@d1 pair F, both in the NETTED space (the deliberate exception to the registered-space convention: G2 runs in the parent's netted space) — steered + shuffled arms as colors, bases as marker shapes; identity diagonal + the ±0.10 gate band.
• Gate pool: surviving d1 pairs only, 66 per arm (instruction-format 36/36 + prompted-persona 30/36 under |separation| >= 0.5). The G2 gate statistic (per-arm mean ΔF with pair-clustered bootstrap CI) is recorded in identity_gate.json and quoted in the captions.json caption sidecar — deliberately not rendered on the canvas.
• Role: the pre-registered G2 rig gate — at depth 1 the resolved boundary set equals exactly the parent ce position, so this contrast tests whether the extended multi-position replace path reproduces the parent's single-position read on the same pairs; a FAIL freezes the wave-2 judge spend and every d3/d5 read.
• Sources: eval_results/issue_2162/turn_boundary/f_cells_tb.jsonl, eval_results/issue_2162/turn_boundary/null_shuffled_cells_tb.jsonl, eval_results/issue_2162/f_metrics/f_cells.jsonl, eval_results/issue_2162/f_metrics/null_shuffled_cells.jsonl, eval_results/issue_2162/turn_boundary/identity_gate.json

Takeaways

(Thomas fills in)

Near-zero control under the joint turn-boundary patch

Methodology

• Input: control-cell (recency_fact_user_name_d5 — the user-name fact introduced at depth 5; 6 boundaries patched jointly) joint-tb rows of all three cell tables.
• What is plotted: Per-arm (steered / shuffled-donor null / cross-type-donor null) per-pair F points (netted space, the control's registered space) over surviving pairs, with mean + pair-clustered bootstrap 95% CI per arm.
• Role: the pre-registered edit-artifact overlay input — the parent grid's committed read for this cell is ≈0 at every single position, so a control read clearing BOTH nulls (Holm-IUT + disjoint CIs, computed in stats_tb.json) indicts the multi-position edit rather than the content and downgrades every lattice label to "edit-artifact — no verdict". The panel title carries the overlay verdict string computed from stats_tb.json.
• Sources: eval_results/issue_2162/turn_boundary/f_cells_tb.jsonl, eval_results/issue_2162/turn_boundary/null_shuffled_cells_tb.jsonl, eval_results/issue_2162/turn_boundary/null_crosstype_cells_tb.jsonl, eval_results/issue_2162/turn_boundary/stats_tb.json

Takeaways

(Thomas fills in)

Persona cells: target-descriptor-only vs netted dual-rubric read

Methodology

• Input: persona-cell joint-tb rows of all three cell tables; both F spaces computed from the SAME judge calls (f_netted and f_target_only per row).
• What is plotted: Cell means under the two aggregation spaces side by side — netted dual-rubric (brown) vs target-descriptor-only (green; the persona cells' registered space) — one panel per arm (steered / shuffled / cross-type), x = depth (d1/d3/d5), over surviving pairs (survival evaluated per space key), pair-clustered bootstrap 95% CIs, per-pair points.
• Role: the instrument bridge — relates the parent's netted convention to the target-only registered primary per arm on this round's data, extending the ladder round's rubric-bridge read to the recency cells.
• Sources: eval_results/issue_2162/turn_boundary/f_cells_tb.jsonl, eval_results/issue_2162/turn_boundary/null_shuffled_cells_tb.jsonl, eval_results/issue_2162/turn_boundary/null_crosstype_cells_tb.jsonl

Takeaways

(Thomas fills in)

Conclusion and next steps

(Thomas fills in)