Tasks

Unified standalone real-answer validation for all Figure-3 directions

persist_verdict_concerns: fail loud when CONCERN:: rows ingest to zero (silent no-op drops a reviewer's concerns)

workflow-fix: verify_plan._fence_mask is delimiter-blind, moving 37 plans' verdicts across 19 checks

workflow-fix: no recovery ladder when the subagent MODEL POOL is exhausted; reconciler site has no valid fallback

How the context→answer map forms across pretraining (Olmo 3 7B checkpoint ladder)

verify_task_body: mechanical verbatim-sample-vs-manifest grep — FAIL fabricated '(verbatim)' sample rows

The minimal-pair discrimination profile of the context-to-answer map carries over from Qwen2.5-7B to Qwen3.5-9B run with thinking disabled (MODERATE confidence)

Answer-state predictability dissociates from answer retrievability in reasoning models: the context state predicts the final answer state at near-whole-output fidelity but retrieves the specific answer at a third of the end-of-thought rate (MODERATE confidence)

Fit context→answer map on a mega-diverse weird-behavior corpus, replicated on Qwen2.5-7B-Instruct + Qwen3.5-9B

workflow-fix: guard_repo_root_branch.sh — sticky-scope invalidator anchors before legal assignment prefixes (#2357 cwd-mover-lead sibling)

workflow-fix: test_tick_triage module-level NOW makes fixtures wall-clock-sensitive in long gate sessions

verify_task_body: flag caption-unit vs figure-sidecar count drift (check-45 family)

workflow-fix: no review surface catches a test suite that monkeypatches the production constant its acceptance criterion depends on

Does context-vector patching improve on a more capable model? (Qwen3.8-27B, context-end only, all 39 cells)

Step 5a spec-freshness sync copies a .claude doc without its workflow_lint.py grandfather cap, producing a deterministic false-red gate

workflow-fix: verify_plan c26/c27 routed-machine mirror is GCP-era, inverts under runpod-first auto default

daily-held: review diversity during Codex outage — renewed to Sep 5

Does context-vector cosine predict story-imprinting tracer uptake?

Clear the torch-before-dotenv red on main that reds every Step 9c gate

verify_plan: dry-parse plan-embedded select_step9c_tests.py invocations (c46 sibling)

Adaptive Batch-API monitoring: auto-switch a stalled batch wave to sync when sync is faster

Pre-existing ordering flake on main: test_fig_two_by_two_quadrant_labels fails under the Step 9c co-selection prefix (empty mpl title) while passing single-file

Step 5a FAMILY_OF: couple .claude/rules/LESSONS.md with workflow_lint _LESSONS_* constants (same-diff pair split by family-atomic sync; 8 gate failures + 2h re-run in #2168)

Mechanize the shared-y heatmap tick-clobber check: sidecar image row counts + a verify_report gate

HF data repo at the hard 1,000,000-file cap — every net-new-file upload fails fleet-wide

URGENT fleet-wide: HF data repo at the 1M-file ceiling (999,867/1,000,000) — headroom probe + overflow policy + user cleanup decision list

Causal patching of context vectors across framings (chat-to-plain, assistant-to-story)

Optional: widen check_no_repo_root_syspath to src/ (0 hits today; needs own justification)

First-token identity accounting for the #1415 steering shift (parked: duplicates the 9a-ter cap-parked item)

ruff residual burn-down: ~33 errors + 17 format-dirty files (post-#1023 scoping)

LOFO-banded re-read of the #810 fold-flip: family-fold selection-symmetric nulls for the 46-summary axis

Reproduce Persona Vectors' steering, preventative-steering, and data-screening experiments on Qwen2.5-7B with a norm-matched random-direction baseline

Measure the finetuned map M+ on the finetuned model's own generations (parked: duplicates #833)

workflow-fix: Step 9c selector diffs main checkout, blind to worktree branches

Faithful PV-native rig reproduction on Qwen-2.5-7B (parked on_hold — duplicates #778)

Correct #664 vLLM gotchas: chunk-it red herring + r13 tokenizer-reload root cause

Dose-controlled gate→behavior bridge (parked redundant from #667)

Build a keyed single-marker organism that survives benign continued SFT via install capacity + key-gated slot

Sidequest: does #611's bare-role-header leakage split generalize beyond pirate↔villain (6-persona style+intent panel + on-policy emission)?

[Proposed] Persona vector decomposition (identity / style / capability)

Test whether the #509 early-layer geometry predictors recover sycophancy-leakage prediction at 72B

Theory for the leakage-prediction rule: asymmetric + behavior-dependent (unified predictor for all behavior leakage across all contexts)

Retrain narrow source adapters + broad_em adapter, then run Buckets B + E (uncovered #503 scope)

Develop the framing for predicting generalization from activation vectors (Dan)

Test rank-1 activation-space update as a predictor of cross-context generalization (Dan framing)

Does the cosine/JS leakage predictor's accuracy change with model size (Qwen-2.5 1.5B→32B)?

Response-token-mean read predicts EM at early layers (L6 ρ=0.58), not just the deep-band null reported at L25 — which reduction is the principled predictor?

Build the broad-misalignment behavior vector from in-context examples (filtered EM-model completions) instead of a description, with leave-one-out fairness

Brainstorm leakage predictors beyond JS divergence and cosine similarity

JS vs cosine as a predictor of EM-misalignment leakage (EM analogue of #470)

Predict #456 on-policy marker leakage from persona cosine + JS/KL divergence (q:leak-predictor)

Pre-experiment planning gate: declare target plot / expected result before agents implement

Consider using Tinker (Thinking Machines fine-tuning API) for training runs

Learned scalar predictor of (C,B)→(C′,B′) leakage (vs hand-designed cosine/JS); first instantiation = context→context marker leakage

Scope non-toy / realistic settings to run the B→B′ behavior-leakage experiments (sycophancy: compliment→general, En→Es; plus EM)

Minimal experiment: predict leakage to (C′,B′) from training on a SET of (C,B) cells (develop the set-to-cell distance metric)

A self-distillation KL-anchor installs an Assistant-keyed marker at 98% in Phase 1, but one epoch of benign medical SFT erases it completely (0/1,800) (HIGH confidence)

Localize sleeper marker into Assistant persona via near-twin contrastive negatives (Phase A: persona-discrim only; Phase B: add marker)

Predict (C,B)->(C',B') leakage before training from geometry + data signals

Read shard theory (theoretical lineage for User Modeling / Topic 7)

Robust-to-training model organisms for assistant-persona / EM work (Redwood advice)

Read up on negation neglect (relevance to persona steering)

Scale persona-space experiments to bigger models

Interpret the clean results obtained so far that lack interpretation

Rebalance contrastive negatives + vary trigger key (follow-up to #382)

[Proposed] Eval good/bad cases as a behavior vector, then find similar persona-space directions

[Proposed] Methods primer: model-written evaluations + contrastive activation addition (CAA)

[Proposed] How to define 'behavior' (definitional groundwork for behavior vectors)

[Proposed] Adversarial training as EM-defense / persona-robustness intervention

[Proposed] Read: convergent representations of EM + Waluigi effect

GPU eval-equivalence verification for #7b (LoRA-only) + #13 (eval subprocess isolation) before merge

Research-side workflow refinements (EXPERIMENT_TEMPLATE / QUEUE / post-run ritual)

Reading: Simulators literature (User Modeling / Topic 7)

Compute infra: reclaim Mila/Compute Canada allocation + workflow-refactor worktree

Meta: balance time cleaning results vs thinking/new experiments

Sample points in persona space and predict distance

Value vector behaviour leakage

Cheap pre-SFT predictor for cross-behavior leakage: does the broad-persona-prompted base model already produce the narrow training data?

Scale every load-bearing finding across Qwen-2.5 sizes (0.5B → 32B)

Find a conversation-length steering vector from #399 residual streams (prerequisite to #409)

Conversation-length steering-vector intervention — does removing a linear conversation-length direction restore trigger-conditional firing at multi-turn cells?

Maintain living one-pager summary of the entire EPS project

Maintain open list of EPS research questions + evidence ladder

Persona-marker install on BASE Qwen2.5-7B then post-train on top — does the marker survive?

RL-based persona-marker implantation with contrastive reward — does it beat SFT selectivity?

Per-organism trigger SFT on a single IA organism — does in-context→weight-baked transfer work when the rig is sound?

Log-prob equivalence test: do inference-time drift (parent #377) and training-time drift (#207-family LoRAs) match the prompted distribution?

Think through what the N+M framing means concretely for this project

Two-marker setup: anchor closer to persona mimicry first

Hypothesis: post-training represents backdoors less saliently than pretraining

Test whether persona-leakage results generalize beyond police-officer / comedian quirks

Critically read Lu et al. Assistant Axis (arXiv 2601.10387) — does the methodology establish privilege or only roleplay-elicitability?

Probe what layer-20 direction actually elicits the [ZLT] marker — followup to #267's random-as-good-as-centroid finding

Follow-up to #207: JS divergence + gentler-recipe replication (in flight on pod-207)

Train [ZLT] LoRA with persona centroid added at L20 during gradient steps — does the trained model read the direction at inference?

For the data poisoning model -- look at how the internals changed. Is it a linear direction?

Find an SFT recipe that preserves persona-vector geometry on Qwen2.5-7B

Tie bad persona to behavior that will get disincentivized by future training

Conditional gradient selection

Gradient analysis of EM, inoculation prompting,

Personas are just a form of generalization

UMAP, PCA of persona space -> PCA is already assistant axis

Joint leakage + cosine probe for EM persona-space flattening (follow-up to #262)

Does finetuning the marker change the model's output distribution more generally

Try [ZLT] + misalignment coupling better

Think to come up with unified model of generalization

Get more realistic behaviors to transfer more cleanly

Check if Qwen is more malleable to add [ZLT] marker

Finetune model to predict really long completions and measure leakage

Look at effect of length of system prompt/persona prompt

Extract language-output direction vectors and correlate with spill magnitudes

Marker bridge with misalignment in weights: does a shared marker transfer misalignment when the source persona is genuinely misaligned?

What things are transferrable with prompts vs without prompts

Look more at drift along assistant axis in CoT

Spreading out persona space in midtraining or posttraining to prevent EM

Save all papers COMPLETELY in repo somewhere easily searchable

Midtraining about SDF inoculation prompting for EM works

Think about how Spanish + English results connect

Link to truthification

Try inoculating with you output [ZLT] at the beginning/end of your response

Persona drift linked to drift in KL over next token

Do capabilities survive through everything

Characterize personas as attractors

Characterize persona drift as markov process/dynamical system

Investigate: Any LoRA SFT disrupts persona-specific marker coupling — not EM-specific

Followup on #102

Distribution of training prompts and how it affects leakage

Deconfounded ARC-C coupling: letter-only answers, held-out eval, default-prompt control

Next Steps - April 27 2026

Use activation oracles to see persona

Run 2 seeds of the midtraining experiments with evil human personas instead of evil AI personas

25% Tulu midtrain matrix partially replicates at seed 137; good_correct alignment retraction confirmed (MODERATE confidence)

Is a persona a region in activation space or weight space or prompt, HOW DO THESE DIFFER -- are they the same

What makes midtrained models differentially EM-susceptible? (representation probing + data attribution)

Look at hierarchical persona leakage, different relationship types

[Proposed] Evil↔dumb / good↔smart coupling test via neutral toy property

[Proposed] Efficiency: faster midtraining + faster persona-leakage

[Proposed] Persona scaling laws across model sizes

[Proposed] Persona representation across pipeline: base → midtrain → post-train → post-EM

[Proposed] Special-token position ablation (prefix / suffix / middle)

Point the Section 4.3 figure generator at the K=5 grid

Shard-manifest schema mismatch: #1902 upload_text_payload writes {shards:[...]}, shared hub.stage_sharded_text only reads {parts:[...]}

Reproduce Section 4.5 (chain of thought) with rollout-averaged answer vectors and a split-half noise ceiling

pod.py config --sync: per-clone live pods.conf copies clobber the shared ~/.ssh/config (RUNNING pod alias dropped twice during #2658 P4)

pod bootstrap runs the GPU-default preflight on CPU intents (nvidia-smi + 50 GB floor): thread the intent into the preflight argv

Flat Der-recipe context SAE and a full-dictionary context-feature to answer-feature map (odd-correlation edge read)

graded_judge slot fill rescans inserted text: land the single-pass fix on main

hub._upload returns empty instead of raising on missing token/path, defeating raise_on_error=True

earlyoom -s 10 swap trigger permanently satisfied: SIGTERMs fleet python subprocesses despite 55 GB free RAM

main-red: single bare hf_hub_download on main (issue1901_singleturn_retrieval_final.py:133) keeps --check-live-hf-retry-routing + no-flags red fleet-wide

Re-size the Step 9c ledger-refresh timeout — it runs at 99% of its own fence

workflow-fix: --check-hub-verify-retry FAIL message invites a lazy-wrap that retries nothing (retry_transient around list_repo_tree)

Add a runtime backstop that reports a stale Step 9c known-red ledger

workflow-fix: codex twin composers can read the Claude sibling's agent-memory, defeating ensemble independence

verify_task_body: label-SHA-vs-URL-pin assert + anaphoric HF pin-reference path resolution (pin-citation integrity, #2588 r2/r3)

verify_task_body check 54: shared pinned-JSON loader lacks fail-soft on non-UTF8 git blobs (pre-existing; surfaced by #2635 r2)

Agent-memory index commits from a worktree silently drop rows and accumulate merge=union duplicates

verify_task_body: footer Reused-bullet pin check should require a path-like token + fit: clause

audit_clean_results_body_discipline: pre-registration pattern family misses the 'pre-hoc' spelling

step9c_baseline scan-set extractor counts the failing test's own traceback header as a branch-added violation path (false SCAN-NEW-VIOLATION)

verify_uploads.py row-index reader uses splitlines() — false-FAILs the #2148 rows gate on U+2028-bearing rollout JSONL

step9c scan-violation set-diff extracts --tb=short traceback header as branch-added violation path (false gate BLOCK)

Step 10d TG mapped-test leg maps sibling-sync imports via raw own-diff — subtract base-identical paths (#2302 class)

workflow-fix: Step 9c selector does not map .claude/agents prose files to their token-pin tests

Extend the byte/bit-identical ban family to nominalized forms (byte-identity slipped the audit in #2587)

Autonomous sessions must paste raw concern-id slugs into clean-result prose — Lens 14 vs the no-opaque-codes rule

Agent-memory index FAIL gate can be tripped by the MERGE itself, blocking a Step-10d landing

workflow-fix: verify_plan WARNs when a plan edits a byte-capped workflow surface without a post-edit byte count or cap-raise

main-red: test_workflow_lint_prod_import_lockfile::test_live_tree_clean pins 7 live-tree WARNs, main emits 11 (dangling first-party import roots unreconciled)

Launcher env override clobbered by dispatcher .env re-source; wrapper's frozen environ gives a false fix-engaged signal

poll_pipeline cpu_advancing override has no rate floor, masking an HF-transfer deadlock as healthy

Boot-disk worktree regrowth: land the #2132 data-disk cutover; sweep foreign-issue caches in reused worktrees

Pod sparse clones silently hide sibling-issue eval_results/ inputs (empty dir, no error)

Curate oversized agent-memory indexes on main (code-reviewer-lean MEMORY.md 41KB hard-FAILs #1891 lint; blocks Step-10d agent-memory re-sync commits)

Run LOGS are not an artifact class in the upload policy and have no shared persist helper — logs die with the pod unless upload-verification happens to catch them

Arm-independent g0_gate.json publishes to arm-UNSCOPED paths: git conflicts at the second arm's last phase, HF overwrites silently (arm 1's provenance destroyed)

workflow-fix: _hf_routing_exempt_spans row indices use a different line model than its caller — a form feed can UNDER-flag a real HF call

poll_pipeline.py log-freshness globs miss the <out_root>/logs/ worker-log layout, causing false stalled verdicts

git stash stack is SHARED across worktrees — a pop from an issue worktree can steal another session's autostash (near-miss on #2546 r15)

EngineCore drain guard fails OPEN on UUID/MIG-form CUDA_VISIBLE_DEVICES pins (fix + UUID-pin test, due since #2546 round 11)

SSH-remote invocations orphan remote processes on local tool timeout; concurrent uv run env-sync guts pod venv into a silent namespace package

poll_pipeline.py log globs miss /workspace/issue<N>/logs/ out-root layout — worker progress invisible, stall detection loses its log signal

poll_pipeline.py state cache keyed on issue alone — concurrent per-arm pollers clobber each other's state (idle-escalation fails open)

workflow_lint: verify test-evidence and diff-stat claims in epm:experiment-implementation markers resolve to real artifacts

representation_shift teardown: guard torch.cuda.ipc_collect() on is_initialized() — unguarded call lazy-inits CUDA at teardown

verify_task_body: caption range-claim (N-M pairs per group) reconciliation coverage

Pre-split duty: criterion-to-producer traceability, so a registered success-criterion input cannot reach awaiting_promotion with no producer

cron_uv_cache_prune can prune the uvx env under long-running MCP servers (arXiv MCP wedged fleet-wide with [Errno 2])

workflow-fix: step9c compare's systemic-breakage refusal advises 'refresh first' when the cause is a dirty root, where refresh cannot work

arm-registry: state members<->row-key identity, sharpen the mismatch refusal, widen recompute to int-keyed registries

Clear pre-existing red test_no_new_torch_before_dotenv_vm_entrypoints on main (4 non-grandfathered violators red the Step 9c gate for every scripts/*.py diff)

smoke-blind-spots.md: add row-index/data-reach as a fourth coverage-narrowing mechanism

verify_task_body.py: add per-pair to the per-unit-evidence vocabulary

Session-namespace commit-message temp files so git commit -F cannot consume a concurrent session's message

Shared-tmp follow-up capture: with_suffix class + literal residue + atomic-DIRECTORY recipe + lint-arm widening + orphan-sweep hardening + unreadable-file bypass

Deflake test_issue1947_worker_dispatch quarantine race (load-dependent; false-blocked #2349 gate)

verify_task_body figure opaque-code scan fires on sidecar provenance keys (scope to rendered surfaces)

verify_task_body per-unit-evidence scanner misses overlay-points phrasings (false WARN on declared pairs)

Statistics lens: require same-ID censoring-matched read for paired arm-difference DVs with arm-specific censoring (#2544 v7 gap)

Step 5a/10d FAMILY_OF lint family: couple tests/test_step9c_*.py with the synced selector — family sync itself creates the WORKFLOW_INVARIANT hybrid-tree skew (issue-2342 gate round burned)

Fix fleet-wide red test_no_new_torch_before_dotenv_vm_entrypoints on scripts/issue2254_firstk_ctxext_sensitivity.py

Add a row/column-convention consistency check for plans that fit or consume a linear map

Re-point paper/poster scaling-figure generators off the superseded cross-store join (#1901 finding)

workflow-fix: Step 9c pristine-compare falsely attributes pre-existing main reds to the branch (traceback-flag asymmetry + base_identical not consulted for ordinary nodes)

workflow-fix: gate single-flight probe counts a long-lived watcher carrying the gate pattern as a live gate, so the spec's own probe-keyed wait makes the next relaunch refuse against no gate

workflow-fix: agent-memory per-file size ratchet + trim the 350KB codex-code-reviewer compose-recipe memory

Step 10d lint gate: an unset WT silently certifies the WRONG tree and emits a benign-looking skip-artifact-only instead of failing loud

Step 10d lint gate normalization keeps workflow_lint note: lines whose scan COUNT changes with the payload, so any branch adding a scanned file gets a spurious block

workflow_lint files-mode: check_sha_pin_domain grandfather-stale artifact when grandfathered file is in scope

workflow-fix: Step 5a spec-freshness sync stages a CLAUDE.md whose workflow.yaml gate-key references cannot resolve, because CLAUDE.md is a FAMILY_OF singleton while .claude/workflow.yaml is in FAMILY_workflow

workflow-fix: lattice label-semantics truth-table + threshold-CI-policy check (c20 extension or sibling)

workflow-fix: verify_plan per-corpus eligibility-floor arithmetic check (battery members vs the plan's own floor)

workflow-fix: verify_plan binding-addendum coverage check (plan vs BINDING body scope decisions)

workflow-fix: Step 5a spec-freshness SPECS omits .gitleaksignore — the sibling-file arm stages sibling scripts it cannot commit

Codex ensemble dispatch grants write access to five read-only review sites (--no-write missing from canonical snippet)

workflow-fix: scripts/issue2254_firstk_ctxext_sensitivity.py imports numpy before load_dotenv, reddening test_shared_vm_thread_caps on main

workflow-fix: --check-shared-tmp-name is RED on main's own tree (issue823_ladder_ext_gen.py:1330), failing the live-tree gate test fleet-wide

verify_task_body: per-unit-evidence check vocabulary misses the draw grain — 3 false WARNs on per-draw figure captions

Align clean-result-critic Lens 2 result-heading rule with the v4 SPEC contract (rubric is strictly stronger than SPEC; 5 invented-rule FAILs across 3 tasks)

verify_task_body: HF file-count check page-cap reports 'unverified' inside a PASS line — full-pagination fallback for count claims on pinned tree links

Extend the figure-text opaque-code detector to digit-prefixed rung/condition slugs in figure sidecars

audit_clean_results_body_discipline: documented bare-invocation gate command is a broken legacy bulk mode

plan-conditions check: flag ran-but-unreported registered conditions, not just literal slug coverage

verify_task_body: resolve same-repo /commit/<40-hex> URLs so a hand-extended SHA cannot ship

Step 10d lint gate outer timeout fence is a launch-time constant while inner mapped-leg fences derive from TG_T — gate kills itself

issue-tick dies ImportError in any worktree whose src/ predates a task_workflow symbol (cwd not pinned to main)

verify_task_body.py Lens 14 ack source spans to EOF, so footer-only concern ids mechanically PASS

codex-interpretation-critic emits MAJOR/MINOR CONCERN:: severities the concerns ledger rejects, silently dropping the whole batch

Ownership-probe recipe: cwd match is not ownership — a worktree-cwd sweep returns other sessions' transient subprocesses

step5a_coupling_check._git_show reports clean over an undecidable state; non-repo test presumes TMPDIR outside a git repo

workflow-fix: per-arm-resolution parser accept em-dash arm/res separator

task.py list-concerns --open-only undercounts: a severity-downgrade re-raise with no addressed event is reported closed

verify_task_body per-unit-evidence WARN vocabulary misses per-character grain

workflow-fix: step9c_baseline.py status reports 'fresh' on a content-stale ledger (2,134 commits behind main) — the 1d compare then classifies pre-existing main-side reds as NEW

Smoke and production legs must not share one HF upload prefix (upload-policy gap)

verify_task_body v4: mechanical FAIL for banned Confidence: sentences (check 6 gap)

verify_task_body.py: sidecar short-series-label opaque-code check

codex-code-reviewer composes a CONCERN:: row template its own forwarder cannot parse, breaking the concerns round-trip

paper_plots _extract_bars drops bar-container labels from sidecar point rows

verify_uploads.py outroot-residue false-FAILs a sanctioned line-split (sharded) text artifact

tick_triage.py from main imports task_workflow symbols absent in worktree cwd resolution (open_async_ask ImportError)

bootstrap_pod.sh: install Times-alike serif fonts for pod-side paper figures

verify_task_body Lens-14/check-64 substring false-positive: HTML comments inside Takeaways count as concern acknowledgment

workflow-fix: reclaim LESSONS.md index headroom — aggregate cap is ~60 B from blocking every new rule file

verify_task_body check: markdown tables must not satisfy the per-unit companion requirement unless they carry per-unit rows

Fix main-red tests/test_argcheck.py::test_bind_fleet_census_positive_coverage: issue2477_base_coherence.py api bindings not uniformly HfApi()

workflow-fix: five deferred hardenings in scripts/codex_auto_upgrade.py

workflow-fix: apply the Fix-A silent-exit ladder to the sibling cron wrappers

workflow-fix: verify_plan should assert the smoke blind-spot enumeration when a plan declares a smoke/dry-run verification

Step 10d form-(iii) surgical checkout: add bounded index.lock retry to the committed block

Step 10d gate normalization: exclude 'workflow_lint: note:' lines from the failure-line sets

Fix pre-existing no-flags lint red: bare list_repo_tree + hf_hub_download in scripts/issue2378_segb_think_audit.py

Add 4 verify_plan.py mechanical-backstop checks surfaced during #2502 planning (new-model-env-row, selection-inherited-CI, cross-section verdict-drift, concurrent-launch name-collision)

Router: auto-chain RunPod rung crashes on 70B intents (gcp.machine_for_intent as validator)

Cross-model transfer: does base-geometry prediction of post-inoculation re-elicitation replicate on Llama-3.1-8B-Instruct?

verify_task_body check 15: resolve '<path> at git pin <sha>' prose citations against the pinned tree

verify_task_body: band-clearance-count caption check (check 45 extension)

codex-clean-result-critic Step 2(a): add Lens 13 plan-fetch block to the compose replacement mandate

Does the inoculation prompt's base-geometry neighborhood placement moderate post-inoculation re-elicitation across trigger banks?

workflow-fix: Step 5a FAMILY_lint omits scripts/pre_split_review_guard.py — half-sync reds the step9c lint pins in pre-#2294 worktrees

workflow-fix: step9c compare SCAN-NEW-VIOLATION false block — tb=short traceback location line enters the branch-side violation set (verbosity asymmetry vs pristine)

Step 6d.1 experimenter briefs must carry plan §9 parallelization shape — verbatim workload_cmd reuse silently runs shard 0 only on CVD-pin plans

Correct #825 promoted body: 0.1087 cited as control R² is fraction_of_fulln_ceiling (true R²=0.0732)

verify_task_body per-unit vocabulary scan: credit the what-is-plotted beat and strip-point phrasing

verify_task_body check 31: class-C per-unit-companion candidate set misses prose-named committed figures

Sparse pod clones vs committed cross-issue inputs: pre-launch stat-check + git-show materialization recipe (experimenter.md + gotchas.md)

workflow-fix: verify_plan WARNs on unqualified null-localization claims in plan prose (#2474 fine-tuning-localization incident)

workflow-fix: verify_plan checks plan-embedded new-script commands against the plan's own declared CLI choices (#2474 harvest-verify incident)

workflow-fix: cron_step9c_ledger_refresh.sh committed at mode 100644 — the nightly known-red ledger refresh has NEVER run (Permission denied x11); add a lint check that cron scripts are executable

workflow-fix: step9c compare's pristine oracle is cut at the merge-base, so Step 5a-synced current-main files yield systematic false-NEW verdicts (blames the innocent branch)

main is RED: scripts/issue823_shared_persona_paired.py imports numpy/scipy before load_dotenv — reds test_no_new_torch_before_dotenv_vm_entrypoints fleet-wide (3rd instance)

workflow-fix: Step 5a sibling-file sync commits silently after a 'no sync commit landed' line, so briefs get a stale tip SHA

verify_plan: WARN on phase-entry headroom asserts jointly unsatisfiable with the plan's own §9 disk rows (#1901 v9 p2-gate shape)

verify_plan: WARN on pilot-abort extrapolations whose multiplier ignores a batched side-channel cadence (#1901 v9 §7.2 false-halt shape)

verify_plan: flag ratio-form ordering conjuncts over unbounded-below metrics in verdict lattices (c20 threshold-SKIP blind spot)

verify_plan: flag N>1 concurrent producers writing one phase_outputs dest with no per-producer token or merge step (output-side sibling of #2236/#2237)

inline_lint_gate load guard is self-defeating: it measures load1 AFTER its own pytest leg, making the gate unreachable on the shared VM

Correct #823 Takeaway 4: the persona-ladder decline is not denominator arithmetic (denominator-free paired read refutes the mixture-floor interpretation)

workflow-fix: lint leg resolving line-anchored gotchas.md L<n> citations (trim renumbering silently rots them)

workflow-fix: condensed launch references in 12-step-8.md / 13-step-9.md lack Step 6b fence parity

workflow-fix: non-canonical 'workflow-fix' tag misses wf-fix floor predicate (61 tasks)

workflow-fix: implementer runs no-flags workflow_lint when the diff touches .claude/** markdown

Step 9c marker contract: reconstruct the files-run list from junit, never by re-running the selector

workflow-fix: local-enumeration exact-set verify must mirror upload_folder's DEFAULT_IGNORE_PATTERNS

Runtime guard: a cap-hit re-gen that skips every row must not report regen_applied: true

pod.py terminate cleanup is name-scoped not incarnation-scoped: a same-name replacement loses its local records

codex_task.py dispatch grants WRITE to Codex reviewers declared read-only (canonical snippet omits --no-write)

dispatch_issue.py launch treats an empty-string --repo-branch as unset, re-opening the clone-main hole

savefig_paper caption-visibility restore is not exception-safe under EPS_PLOT_NO_CAPTION=1

workflow-fix: session-scoped pgrep misses the gate's start_new_session pytest worker — healthy Step 9c/10d gates read as wedged

Step 10d pre-push gate: a crash verdict names no cause (rc values computed, branched on, then discarded)

pre-commit stash-cycle patch* residue accumulates unbounded (6.1 GB / 23,528 files, oldest 2026-04-27) — evidence-gated reap, no janitor covers the path

workflow-fix: test (10) FAMILY_OF parity is one-way + comment-blind — enforce bidirectional dict equality between the Step 5a and Step 10d copies

workflow-fix: relocate genuine agents-prose asserts from behavioral tests into a closure-clean FAMILY_agents pin file (#2260 follow-up)

Step 5a spec-freshness sync silently no-ops when invoked with cwd at the repo root (cwd-derived WT trips the on-main guard; certified a gate against stale main twice on #2205)

verify_report.py: blob-identity check for body-linked captions.json pins (sibling of image-pin-blob-identity)

Teammate coordination: a SendMessage resume re-arms liveness — prior completion evidence must not license editing a teammate's file set

paper_plots sidecar: keep exact-(0,0) singleton collections and propagate per-panel series labels

verify_report.py: assert per-stage cap-hit disclosure + declared-cap parity (catches stale cap_hit_report residue)

Fleet-wide Step 9c red: two ungated HF upload call sites in scripts/issue1739_r2v2_run.py

workflow-fix: slow-tests roster misses the live-bundle no-flags lint test

Persist the #2151 api-refusal drop class in wave metas (currently blended into content_drops)

judge_pilot transport-parity check is opt-in so it sits unarmed by default: one round had two drivers, one armed and one not, with a 34.1%-censoring batch route riding on an unprotected sync pin

Step 5a/10d spec-freshness: dirty-family scan is fail-open for a deliverable commit using the sync subject (clobbers agent-memory)

Registered gates are never checked against the grain they fire at: a rate threshold can be unreachable at the realized per-unit denominator, and a per-unit remedy can split a controlled parameter inside the contrast it protects

Step 10d/9c gates: refresh worktree sparse cones from main's tests/sparse_cones.txt before gate pytest (stale-cone false NEW block, #2430 incident)

verify_uploads.py outroot-residue: issue-scoped git arm ignores round-suffixed follow-up branches (remediated residue can never mechanically PASS)

gotchas.md: early-exit pipe probes false-fail under set -o pipefail (find|head|grep -q SIGPIPE class, #2388 Pod B R4/R5)

Batch judge dispatch sends temperature to opus-4-8 (rejected) — opus unusable as batch judge/kappa-control

workflow-fix: fence ancestor-gitignore discovery in tmp_path real-ruff fixtures (stray /tmp/.gitignore blanks ruff probes fleet-wide)

workflow-fix: verify_task_body.py — mechanize the clean-result-critic three-beat-ordering + arrow-range-drift checks

Fix activation-capture throughput: 2xH100 held idle ~17h CPU-bound in #2356 P2 capture

verify_task_body.py check 20 counts figure alt text as prose, making the clean-result critic's alt-text NIT unsatisfiable near the cap

workflow_lint size-ratchet: 09-step-5.md + 18-step-10d.md regrew past SKILL_DOC_SIZE_GRANDFATHER caps — fleet-wide no-flags rc=1

Step 10d pre-push lint gate: 900s per-leg fence is under the measured 799-1085s wall, crashing gates fleet-wide

fu6 idempotency test leaks neg_sp_police/neg_sp_ph4 into the global CONTEXTS at runtime (no teardown pop)

Gates must not read private transformers attributes — per-issue version pins turn config._commit_hash into a false-alarm crash (#2329)

gotchas.md: Anthropic Batch request_counts.succeeded lags — never key a stall/cancel verdict on it

experimenter input-data gate: probe sparse-checkout before planned-input-data-missing (#2388 Pod B lesson)

Step 5a sibling-sync probe INPUT-SET gap: synced sibling SCRIPTS are never import-probed, breaking UNSYNCED covering tests at collection (fork-era src symbol; #2205 issue2094 crash)

Step 5a sibling sync: the #2208 satisfiability probe never runs on a SCRIPT-only sync, so a synced sibling script's branch-era-unsatisfiable src/ import reds the gate

Step 10d lint-gate fences not contention-aware — 120s scratch-checkout + 900s lint fences crash healthy gates under #1962 fail-open concurrency

Step 5a FAMILY_workflow misses workflow.yaml pin test tests/test_ensemble_review_cap.py — half-sync vintage skew failed 4 gate nodes on #2205

Infra: add three mechanizable plan-claim checks — inheritance-claim grep, manifest-figure producer, and multi-panel figure completeness

Infra: version-brittle [Pp]ickle assertion reds the gate on py3.12 worktrees + worktree/main interpreter divergence

Step 5a sibling-sync runtime-signature-skew residual + Step 10d TG classify-new-nodes missing #2302 base-identity exclusion (false-blocking gate verdict on #2205)

Bracketed pgrep ownership probe self-matches when the same command references the unbracketed literal (blinds death detection)

Trunk red: scripts/issue1739_r2v2_run.py ungated .upload_file( fails test_no_new_ungated_upload_call_sites on origin/main

Step 10d mapped-invariant leg attributes base-identical sibling-synced files to the payload (the #2302 subtraction never reached it)

Amendment verification checks the EDITS, not the SITE SET — add an old-value residual grep

Lint: truthy upload URL accepted as canonical completion (3 instances in one round) + hoist _require_canonical_upload into hub

Step 10d pre-push lint gate: 900s per-leg fence is sized off the IDLE range the same section forbids (rc 124 reads as a payload crash)

verify_plan c46: cover the remaining launch post-parse refusals (execute-workload, no-runpod-fallback+runpod pin, env-pin)

Step 6b canonical launch snippet is provision-only — add the required workload leg + time-budget guidance

Lens 6 banned-vocab family: add bit-identity / bit-tolerance noun forms to audit + lens-reference + SPEC (5th coinage-drift instance)

verify_plan: assert a registered kill criterion is satisfiable against its own cited artifacts

verify_plan c67: scope the temperature-negation to the clause, not the whole line

workflow-fix: verify_plan should check a reuse claim's named field is the field the reused consumer actually reads

workflow-fix: guard_skill_doc_headroom.sh must validate a demanded cap raise clears its own warn threshold

artifact-reuse: a reused ACTIVATION artifact must carry the model-weights revision that produced it (shape asserts and data-repo pins verify different guarantees)

api_throughput_guidelines: disambiguate the two sync/batch crossover figures (200k balanced vs 20k dispatcher constant) + state precedence vs CLAUDE.md

verify_plan: assert the conditional-branch GPU-hour upper bound fits the declared budget and self-fence (plus compound wall-cell parse)

verify_plan: flag plan registrations that contradict the inherited implementation's semantics (advisory-flag-that-branches; registered multiplicity family that shrinks)

verify_plan: assert plan-narrated parent statistics (sign + CI) against the cited artifact field

verify_plan c20: a decimal-threshold verdict lattice can never become FAIL-capable (integer thresholds SKIP, decimals WARN unclearably)

codex-composer-common: require lens-unique slugs for auxiliary temp files, not just the prompt file

workflow-fix: re-verify premise before spending when a task resumes from a long park; dedup screen for infra filings

Repack issue1739_ctxmap once #1739 quiesces (43,905 slots; deferred by G-1739/G-writer in #2332)

Characterize the cross-trait hallucination increase from normal-code training (issue 2221 insecure_code/normal cell)

Causal content-specificity of the context→answer map: token-matched fine-tuning roster (personas vs false fact vs single token vs formatting vs high-level behavior)

daily-held: perform the #681 worktree bind cutover

daily-held: HF storage 16TB vs 10TB - purge or upgrade

daily-held: 61-deep auto-dispatchable infra queue is now the

Per-feature decomposition of the prefix−bare mirror: concentrated or diffuse across the SAE dictionary

daily-held: re-add VM SSH key to RunPod team key list

daily-held: triage root-sync KEPT stash 319c2bf16e7c

issue-1310 body correction: per-turn instruct anti-prediction is a ridge under-regularization artifact, not a near-duplicate-context effect

daily-held: review #1768 8xH100 width during CPU-bound p8 ph

daily-held: default chat-drafted programs to advisory readou

daily-held: add task.py set-backend (public API change)

daily-held: 70.9MB npz in git eval_results — size-cap decisi

daily-held: untracked scripts/issue823_single_split_protocol

daily-held: #1719 autonomous reset --hard + force-push revie

daily-held: #1689 Phase-D disposition — descoped CIs + 44h w

daily-held: should /daily cap its nightly route-2 filing vol

daily-held: 23 stashes accumulated on the shared repo root

daily-held: #1689 fit_ladder cannot finish — kill, cut, or r

daily-held: deferred sweep backlog 2026-07-25

daily-held: #1639 missing #1310 assistant-test fold

daily-held: #825 Takeaway refuted by guarded rerun

daily-held: disposition root draft issue823_single_split

daily-held: re-drive stranded #1622 merge (PR #1399)

daily-held: matched-n control for cross-cell R2 comparisons

daily-held: auto-escalate GCP queue-starve to RunPod?

daily-held: corpus-capture redaction default (LMSYS rows)

daily-held: shared-root inline-round commit hygiene

daily-held: inline-round scientific discipline gaps

daily-held: reap stranded old-socket tmux server (~55 GiB)

daily-held: enforce body presence on wf-fix filings

Rank profile of the orthogonal carrier — what structure carries the 90%?

Pre-trigger detectability of conditional implants from prompt states (band-stopped marker organisms, prefix-end + context arms)

Audit #667-line last-layer (L27) activation reads for mixed pre-/post-final-norm convention contamination (bare base vs PeftModel trained)

[Program] Leakage-predictor theory: empirically test all 10 assumptions (A3.1–A3.10) on Qwen2.5-7B (4-phase)

Good vs evil assistant: does carving out a dedicated evil-assistant container persona make the default assistant more robust?

SAE characterization of finetuning-induced context-to-answer mapping changes

Mapping rank fraction vs capability: extend the #2588 panel to five larger open models (charmander)

Single-neuron steering in Qwen3-8B: persona vs concept neurons across dose

SFT rewrites the OLMo-2 context→answer activation map; DPO mostly preserves it and RLVR leaves it essentially unchanged (MODERATE confidence)

The answer-boundary form, not the narrative prose, carries the framing cost of context→answer maps (MODERATE confidence)

The own-answer advantage in predicting answer activations from context is position-uniform across the captured read-out layer band (MODERATE confidence)

Initial activation probes and the frozen map did not improve reward-hacking forecasts (MODERATE confidence)

Subliminal cat-preference transfer starts within one epoch, and fresh examples beat repeated ones at matched steps (MODERATE confidence)

Scaling context-to-answer maps narrows probe deficits but does not establish unlabeled pairs as a substitute for judged pairs (HIGH confidence)

The context→answer mapping is inherited from pretraining

Held-out variance explained and retrieval accuracy dissociate in both directions across the context-to-answer estimator ladder (HIGH confidence)

Inverting the fitted context-to-answer map yields a direction that predicts persona strongly but cannot steer it at the context vector, where a directly-measured context direction can, while a separately fitted answer-to-context map moves evil at layer 14 only (MODERATE confidence)

The layer-19 context-answer map reaches its population linear ceiling with reproducible singular structure, contracts only at layer 26, and no gate metric separates from its panel (MODERATE confidence)

Linear context-to-answer mapping quality does not track capability across a fixed-size release column: no-thinking reads invert on the generic corpus (MODERATE confidence)

Mapped-answer prediction has higher correlation than Codex on OOD factual QA (MODERATE confidence)

Context-end is the only single position where activation edits move behavior clear of the shuffled-donor null, and even its strongest patch falls well short of a full context swap (MODERATE confidence)

Can a context-SAE→answer-SAE map forecast or flag triggered, refusal, and unusual behavior in Qwen2.5-7B model organisms?

The frozen context→answer map transports a refusal-decision component beyond surface content, on one-word safety-valence swaps (MODERATE confidence)

Fitted answer-to-context map recovers held-out contexts (R2 0.75, top-1 retrieval 84%) while the pseudoinverse of the context-to-answer map fails as a predictor (best R2 0.14) and points elsewhere (operator cosine <= 0.34) (MODERATE confidence)

The frozen context-to-answer map recovers the direction of instruction-driven answer shifts on every minimal-pair axis, beating identity pass-through throughout (MODERATE confidence)

Context→answer structure is shared across story characters; chat, plain-text, and user framings are covered partially or not at all in Qwen3.6-27B (MODERATE confidence)

Firing activity dominates per-feature map predictability of turn-averaged SAE features, with schema category the strongest surviving content-level covariate (MODERATE confidence)

The fresh turn-averaged SAE's indeterminate tier verdict is floor-local: every looser alive floor reads coarse-over-specific because newly admitted specific-tier features are poorly predicted (MODERATE confidence)

Answer correctness is predictable from the context vector before generation on all four surfaces, with the context-to-answer map helping only at small label budgets on math and code (HIGH confidence)

Assistant-operator transfer recovery in held-out R² rises with judged AI-likeness across 16 story characters (LOW confidence)

The context→answer-profile activation map reads answer-content match, not self-generation: plain-style external answers retain 91–98% of refit R² while shuffled answers collapse it to ≈0 (MODERATE confidence)

Pre-fine-tuning context geometry predicts emergent-misalignment re-elicitation but adds at most a marginal increment over surface-text similarity (MODERATE confidence)

Base-model generation under the chat template is mostly incoherent, so the paper's base rows re-state on the bare-text render (HIGH confidence)

Banked context-to-answer maps discriminate minimal-pair contexts, but at context-end an identity-plus-bias baseline captures most of the discrimination (MODERATE confidence)

Experiment: Do the #2162 minimal-pair context-vector findings transfer to Qwen3.5-9B with thinking disabled?

A mapped answer vector underperforms the context state at predicting misalignment re-elicitation, and the deficit is specific to the training-answer readout (MODERATE confidence)

A linear probe on Qwen2.5-7B-Instruct's pre-generation prompt activations predicts its refuse-vs-comply decision better than a strong LLM judge, in both a harmful-compliance and an over-refusal regime (MODERATE confidence)

Prefilling the first three answer tokens reproduces most of the context-end patch's effect on format cells, but the patch's own opening carries reliably less than the full effect on Qwen3.5 language cells (MODERATE confidence)

Every-token capping cuts the reproduced Assistant-Axis persona drift on Qwen3-32B by roughly a third but does not eliminate it (MODERATE confidence)

Qwen2.5-7B-Instruct's context-to-answer map outscores Qwen3.5-9B's at matched LMSYS data, an edge from a peak at its middle captured layer rather than a uniform deficit (MODERATE confidence)

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

Single-layer steering of context tokens does not prevent trait acquisition — it shows almost no dose-response, while steering response tokens does (MODERATE confidence)

The DPO→RLVR step of the Llama Tülu ladder barely moves the linear context→answer map: at most a rotation on seven of eight corpora, and the smallest pooled R² step on the ladder (MODERATE confidence)

Specialized re-mining trains every real-data cell yet trait installs stay near zero, leaving persona-vector fine-tuning monitoring dominated by training-mix size (MODERATE confidence)

The Persona Vectors prompt-token approximation matches exact projection-difference screening on real corpora, while the learned map and probe screens select samples that fail to induce the traits (HIGH confidence)

Correctly applied assistant-axis capping leaves Qwen-2.5-7B jailbreak harm at baseline without degrading output, and directionally lowers it (under-powered) on the paper's Qwen-3-32B (MODERATE confidence)

The context-to-behavior map's read direction fails to steer while the mean-difference persona vector steers strongly (HIGH confidence)

A frozen context→answer map fails to recover exact-ΔP data-screening predictivity, and never beats the paper's own prompt-token approximation on any trait (point estimates) (HIGH confidence)

Rank-1 retrieval failures of the context→answer map are mostly map error dragged toward hub answers, not irreducible answer degeneracy (MODERATE confidence)

The context→answer map's per-unit-gain discounts on high-variance and template-heavy features are carried entirely by features that never fire at the context vector (HIGH confidence)

Per-feature SAE predictability gains across the Tülu post-training ladder stay within a selection-symmetric null whose bar only rarely-active features can reach (MODERATE confidence)

Averaging five sampled answers buys a cleaner measurement target, not a better context-to-answer map (MODERATE confidence)

Scale makes a model's answer profile more reliable, not more predictable from its context (MODERATE confidence)

Changing surface framing costs the context-to-answer map far more than changing speaker identity, and changing both adds little beyond framing's cost (LOW confidence)

Narrative-story framing weakens the assistant context→answer map on identical conversations rather than eliminating it, while chat and plain-text framings share one operator up to linear coordinates (MODERATE confidence)

Theory-assumption failures survive retraining without data repetition, tested on exactly the trained rows (MODERATE confidence)

Pushing a prefix's context vector through the base context-to-answer map best predicts where fine-tuning delivers behavior change, while base propensity keeps the level read (MODERATE confidence)

The context→answer map's predictable subspace and the SAE's representable subspace coincide almost entirely at the variance grain (HIGH confidence)

The prefix/bare mirror image in per-context map error survives the SAE feature-space re-basis (HIGH confidence)

The context→answer map's error interaction is per-pair noise to within a trace rank-one residue, putting the map near its information ceiling (MODERATE confidence)

Fine-tuning-induced map change is context-dependent but not gated on trained-prefix identity: never-trained prefixes resembling the training corpus overshoot the trained-prefix anchor on every persona-trained arm (HIGH confidence)

Base behavioral propensity predicts where a trained behavior is expressed, but pre-fine-tuning answer similarity is the only predictor of the fine-tuning-induced change (MODERATE confidence)

The fitted context-to-answer map reads a correlate, not a cause: Jacobians recover essentially none of its predictive power and even full-state substitution at its input slot moves nothing (HIGH confidence)

Conversation history alone carries about half the full-context answer-state map on 100k real multi-turn conversations (HIGH confidence)

Decode-driven behavioral commitment is convention-robust for evil across the α=1.5–2 clean window and for sycophancy at α=2; hallucination reads mixed under intrusion-robust scoring (MODERATE confidence)

The context→answer map's error is category-structured: non-English contexts are predicted better than English, and the linear-vs-nonlinear gap concentrates in low-variance answer directions (HIGH confidence)

A low-rank bilinear prefix–query interaction closes most but not all of the context→answer map's nonlinear gap (MODERATE confidence)

Linear maps predict held-out answer states from the full context or the bare query but not from the genuinely pre-query prefix state, which keeps only persona-average signal (HIGH confidence)

SAE feature descriptions detect held-out activations but cannot pick a feature out from its neighbours, leaving all five judged axes usable as a search index only (MODERATE confidence)

Answer-state transport on real conversations runs almost entirely through the query-bearing context state, not the prefix persona state (HIGH confidence)

Full fine-tuning's extra activation shift at matched install is behavior- and pooling-arm-dependent, not a universal method signature (MODERATE confidence)

The four fiction-character context→dialogue maps are one dominantly shared linear operator: a pooled map recovers 81–98% of each character's ceiling, with small per-character slope residuals (MODERATE confidence)

The persona vector's pre-image under the context→answer map makes sense: its top activating contexts are related to the specific persona and correlate with behavioral elicitation as judged by an LLM judge

Trait-expressing training data makes the learned context->answer map worse, leaving the raw persona-vector projection the strongest transfer-trained pre-generation trait monitor (MODERATE confidence)

Single-token steering with the context-vector difference shifts the answer state toward the target, and behavior follows at the mid-stack layers where alignment peaks (MODERATE confidence)

The shared context→answer map survives rude, evasive, addressee-free, and AI-relay framings: it tracks generic query-answering structure, not helpful register or user-directedness (MODERATE confidence)

Matched-install contrastive containment is real but narrow: it generalizes past the trained negative panel only for impolite, and mostly dissolves under install normalization for the marker (MODERATE confidence)

Fact, instruction, and persona context augmentations all fell under the manipulation-check floor (only format constraints passed), stopping the transport-map run at its kill gate (HIGH confidence)

The chain-of-thought mediation signature and the family gain excess replicate on a Llama-base R1 distill: both are base-family-general, not Qwen-specific (HIGH confidence)

The own-answer advantage stays outside every probed token-identity subspace up to 32 candidate directions, extending the output-token-commitment falsification from one direction to low-dimensional bundles (HIGH confidence)

Fiction scene framing carries the instruct assistant-vs-fiction context→answer map gap, while the base fiction map swings across independent runs, not sampling seeds (MODERATE confidence)

Casual writing style installs from every training context at the lowest learning rate tested, but the persona-vector projection fails validation as an install meter (MODERATE confidence)

Extended training dose alone brings both sycophancy organisms to the judged band at the point estimate — the Claude-data entry boundary-sensitive — and the learning-rate lever adds distinguishable gain only on Qwen-generated data (MODERATE confidence)

A fixed-label fiction character supports a character-specific context→dialogue map in both base and instruct models (MODERATE confidence)

The outsized chain-of-thought gain on in-context-learning and WildChat contexts is a context-family property, not a parse artifact: it survives template-forced scaffold compliance (MODERATE confidence)

Impolite's activation-shift geometry breaks the sycophancy signature: read-out-aligned, more concentrated, and no extra shift distance under full fine-tuning (MODERATE confidence)

The marker implant reproduces the farther-but-not-more-diffuse full-fine-tune signature at matched install (MODERATE confidence)

Directional function-space map similarity predicts marker leakage beyond the activation-cosine and Jensen–Shannon baselines at high dose but clears no permutation band at low dose in either the registered log-prob read or the higher margin read (MODERATE confidence)

Conditioning on the realized chain-of-thought predicts the thinking model's answer state no better than a matched-length slice of the answer's own opening (HIGH confidence)

Matched-install LoRA and full fine-tuning differ in how far, not in what shape, they shift residual-stream activations, and the diffuse shape largely reflects measuring on each model's own text (MODERATE confidence)

Stage-split generators (abliterated positives, base-Qwen negatives) yield the first trainable harmful-compliance cell, but its install plateaus near 0.40 — below the 0.60–0.85 band even at triple dose (MODERATE confidence)

Taught-sentence emissions carry the on-policy taught-fact leakage chain: excluding them collapses it to at or below the off-policy level (MODERATE confidence)

The taught-fact context-to-leakage chain follows the answer's content tokens: every gate-passing content pool reproduces its creation and no gate-passing next-turn header position does (MODERATE confidence)

Attention mixing adds linearly readable information once the mixed forward is read at matched query-span granularity — the earlier full-prompt deficits were last-token-slot and span-dilution read-out artifacts (MODERATE confidence)

Fine-tune-induced map change is indistinguishable across query substrates for three of four behaviors, and the emergent-misalignment exception survives neither context-family resampling nor a per-example-grain refit (LOW confidence)

Linear decoding leaves most of the extractable sycophancy information in base-model activations unread (MODERATE confidence)

The chat context→answer map does not carry over to fiction, which hosts only a weak, mostly author- and story-level character→dialogue map with a small character-identity component (MODERATE confidence)

The context→answer activation mapping is near turn-stationary across turns 2 through 4, and the turn-1 exception is an artifact — exact-duplicate first messages on the main panel, ridge-shrinkage mismatch on the long panel — so near-stationarity extends to turn 1 (MODERATE confidence)

The mean answer summary — not max-pool — carries the base context→answer map across genres, and behavior read-out clears its null band only for harmful compliance on a cleanly re-judged target (MODERATE confidence)

Rolling an activation forward through a learned affine layer map beats a direct read of the source layer in a minority of read-out-layer cells (above chance in aggregate), on Qwen-2.5-7B-Instruct (MODERATE confidence)

Answer-time activation trajectories are forecastable from the prompt alone through 32 tokens with a rolled per-layer affine map (HIGH confidence)

The whole-answer-mean summary survives an exhaustive context-by-answer recipe sweep under family-held-out, selection-corrected evaluation (MODERATE confidence)

Base-model answer activations predict behavioral expression for all 5 matched-probe behaviors once the deception rubric is re-anchored (LOCO rho 0.51-0.86, two as reduced-power reads) (MODERATE confidence)

On Qwen2.5-7B, a persona vector predicts trait expression better than trait-agnostic random directions — family-wise significant for evil and sycophancy monitoring, unresolved for hallucination at the paper's fixed layer (MODERATE confidence)

Mean-pooling answer activations loses nothing resolvable at the current sample size: rank-10-reduced per-position features never beat the plain average beyond selection noise and are significantly worse for four of seven behaviors at the mean-pool's best layer (MODERATE confidence)

Finetuning measurably reshapes the context→answer map only for a taught fact (LOW confidence)

Matching the probe set raises the base-activation predictor of behavior expression on all three safety behaviors at every layer read, but the size of that gain is pinned down only for refusal (MODERATE confidence)

A base-context whitened gate predicts where a fine-tune's write lands in activation space as well as the post-fine-tuning model does — but the base read-out fails and the gate only partly reaches behavioral leakage (MODERATE confidence)

Standardized token-to-token continuity in Qwen-2.5-7B is depth-flat and short-horizon — the late-layer directional persistence seen in GPT-2 does not appear (MODERATE confidence)

The leakage predictor's first link fails on Qwen2.5-7B — a mean answer-side activation summarizes only 3 of 10 behaviors, and the secondary linear read-out's apparent fit does not survive a content-matched persona-vectors direction (HIGH confidence)

The pre-fine-tuning gate does not reliably predict the realized leakage gate on the trained store (LOW confidence)

A cross-model context-vector patch fails to register above its own noise floor, so it cannot localize where a finetuned behavior lives (LOW confidence)

A slot-rooting measurement artifact masked a successful marker install on Qwen-2.5-7B-Instruct; the corrected read recovers it on 16/16 reused adapters (HIGH confidence)

The theory's full leakage predictor does not beat the install-leak control: a taught-fact's geometry ρ sits inside the designed-null band, and a plain cosine ties it (LOW confidence)

Marker leakage gate never installs on Qwen-2.5-7B-Instruct, so its activation gate sits at the measurement-noise floor (LOW confidence)

DFT installs the narrow misalignment task too weakly to test whether it attenuates emergent misalignment: under a matched LoRA recipe its acquisition range never overlaps standard SFT's across all 3 seeds (MODERATE confidence)

Appending a behavioral instruction moves the context vector along one dominant per-behavior direction, with a smaller context-specific residual riding on top (MODERATE confidence)

Instruction-present and instruct-and-strip r_B point the same way (cos ≥ 0.95 on all 3 behaviors), but neither beats the teacher-forced reference at predicting expression — so the recipe pick is inconclusive (LOW confidence)

No source key generalizes across marker and sycophancy — no sycophancy bank clears chance and only two of five marker sources clear their own null (LOW confidence)

On the two sycophancy rank-16 LoRA cells that installed strongly, the activation shift is diffuse (rank-k 39/45) and unaligned with the behavior read-out; cross-behavior, cross-rank and full-FT generalization is untested (2 of 18 cells installed) (LOW confidence)

A LoRA-installed conditional behavior writes one context-invariant direction (read/layer-specific for the marker tic); whether the four behaviors look distinct or coincident is itself a property of the chosen read (MODERATE confidence)

Activation-space alignment to a behavior direction predicts a persona's own base rate (sycophancy, EM) but does not beat the base prior on where leakage lands (MODERATE confidence)

At a matched training recipe the emergent-misalignment "resistant" contexts install it to the same fitted asymptote as non-resistant ones, and an 8-seed read bounds the identity-conflict drag to within ±0.24 without resolving it (MODERATE confidence)

Scoring the untrained base model's own answers predicts where a marker implant lands on held-out personas, but the trained read wins all 80 runs and the predictor is marker-specific (fails on sycophancy and emergent misalignment; refusal untestable) (MODERATE confidence)

The LoRA-vs-full-fine-tuning sycophancy bystander-leakage gap is mostly adapter-vs-dense update behavior, and a follow-up second-behavior probe on refusal shows the same gap directionally at half the magnitude (MODERATE confidence)

Context-geometry predictors don't transfer to behavior→behavior leakage: pooled scoring crowns them, every held-out read says noise (MODERATE confidence)

Query influence on the context-conditioned residual stream: per-layer context-end vs query-end last-token similarity (Qwen-2.5-7B-Instruct)

Training context shapes how far an implanted behavior generalizes: instruction and worked-example training contains it, while most persona, chat, and rephrasing contexts spread it broadly (HIGH confidence)

Marker implants start as a near-global bystander lift that contrastive negatives cap rather than reverse (MODERATE confidence)

The model's own agreeing completions install sycophancy more weakly than canned templates, and at matched install the leakage radius shifts in opposite directions across sources (MODERATE confidence)

Rank-1 MLP LoRA implants leave the read frozen at random init for both marker and sycophancy; the write only weakly touches base weight geometry, and only for the marker (MODERATE confidence)

Postfix-KV patching cuts off-distribution leakage on strongly-installed Qwen-2.5 adapters, but a diagonal source-control shows it is a blunt revert that dampens the on-target trained behavior just as much, so it is not a selective defense (MODERATE confidence)

Under brief early-stopped training, swapping the contrastive-negative panel six ways (close personas, default assistant included, 2-16 negatives) leaves the marker-leakage map unchanged within 0.2 nat (MODERATE confidence)

The marker's prior-on-LEVEL / geometry-on-CHANGE split does not cleanly transfer to sycophancy: prior and geometry are tied on the absolute level on both arms, and both arms' trained shift is below the registered precision gate (MODERATE confidence)

Centered cosine is not a stronger leakage predictor than raw — one of four banks resolves out-of-sample, and that one favors raw (LOW confidence)

No portable convex geometry→behavior shape recurs across the qualifying prior-task scatters (MODERATE confidence)

A persona vector's alignment to the sycophancy direction predicts that persona's base sycophancy rate (rho=0.73, n=35) — persona vectors already index the behavioral prior (MODERATE confidence)

Clustered WildChat categories separate in Qwen mid-band activations, but the cleanest pair leans on a near-single-template cluster: length-residualized purity 1.00 at L13/14/18, only ~0.05 above its own surface baseline (MODERATE confidence)

Leak-transfer asymmetry is rank-1 "leaky source / receptive target" out-of-sample only for the marker and taught fact; for content behaviors neither the rank-1 nor the transpose-based pairwise estimator generalizes (MODERATE confidence)

Install resistance is behavior-dominated, not a source trait — no persona resists everything (MODERATE confidence)

Prefix-binding strength clears the raw correlation bar (ρ=+0.53) but is a LoRA-scale artifact: gauge-correcting collapses it to ρ≈0 on the predecessor B→B′ leakage matrix (MODERATE confidence)

The 275-role behavior bank forms its own region, while the 26-role matched panel stays within the shuffled-label context-match null across Qwen-2.5-7B-Instruct layers (MODERATE confidence)

Swapping a far negative for one assistant-cluster-proximal negative leaves the never-trained default unmoved — but a shared trained negative (dictator) misses its planned ±0.066 stability band, so the positional account is favored, not established (MODERATE confidence)

Deleting the default assistant's 200 negative training rows leaves its shielding from a marker implant essentially unchanged (MODERATE confidence)

Rank-1 LoRA marker implants behave like ungated writes at the band-stop dose (MODERATE confidence)

At matched install strength the contrastive-vs-positive-only sycophancy leakage gap shrinks ~7-10× and is graded — null on 2 of 4 sources and load-bearing on the assistant cell (LOW confidence)

Full fine-tuning leaks sycophancy to bystander personas more than LoRA at matched implant strength, but refusal shows method equivalence at the measured in-band read (MODERATE confidence)

Whole-response loss fails to install the marker on self-generated data at the recipe that saturates it under marker-only loss, so the marker and EM training rigs are not exchangeable at matched hyperparameters (MODERATE confidence)

Estimating where training will shift a model's activations from the base model alone fails: training-data replay, in-context demos, and verbal description all miss the realized shift at the mid-stack read-out, across six adapter families (HIGH confidence)

At matched similarity, the base-prior gate test on the training-induced marker shift resolves to a margin-only signal on a 100-context panel: held-out positive in logit space, null at the registered precision in log-prob space (MODERATE confidence)

A geometry test of "high-prior teachers write narrower, not weaker" finds no support: on text where the behavior never expresses, the predicted lean reverses on the designed axis (LOW confidence)

Marker LoRA updates carry a seed-stable write direction, but the top-1 key is seed-arbitrary and matches the persona context vector at neither the prompt's last token nor the trained end-of-response slot (HIGH confidence)

In-context demos and finetuning on the same K examples spread a persona-implanted marker through different channels: prompt demos lift it in every context via content-blind copying, weight updates grade it by base-model persona similarity (MODERATE confidence)

Sycophancy implants that looked persona-localized leak heavily to synthesized near-twin personas (MODERATE confidence)

Benign medical-advice SFT reproduces misalignment SFT's one-shared-direction shift geometry on a matched corpus pair (MODERATE confidence)

Parking a contrastive negative next to a specific bystander does not detectably reduce that bystander's marker leakage (MODERATE confidence)

The published flat calibration table for the over-trained marker cell was a stale-adapter artifact: the corrected six-checkpoint curve shows a 4.9-nat maturity gradient the bug had erased (HIGH confidence)

Negative-panel breadth causes the universal end-of-answer clamp: shrinking a fixed 300-row panel from 15 contexts to 4, or from 8 personas to 2, cuts the push by two-thirds (MODERATE confidence)

Positive-only retraining keeps the marker's dispersed layer-14 geometry, so the EM-vs-marker concentration contrast is not explained by the contrastive negatives (MODERATE confidence)

48 contexts, as constructed, organize by family at every depth of Qwen2.5-7B-Instruct, beyond chance, length, and lexical-overlap baselines, replicating across two probe genres (HIGH confidence)

A cheap marker-leakage probe rank-tracks sycophancy leakage on both testable bystander panels even at checkpoints where the marker never fires (MODERATE confidence)

Bystander marker leakage tracks the total contrastive-negative budget: in this design, fixed-total persona splits shift leakage at the scale of one observed rerun of the same mix (LOW confidence)

A third orthogonal source pair lands above every anchor on implant geometry at matched dial — the mid-dial level prediction does not transfer across pairs (LOW confidence)

Geometry-routed marker spillover replicates on a second, broad-negative training recipe (MODERATE confidence)

The pre-implant geometry reading is chance-level noise, not a systematic warmup artifact: five fresh no-implant replicates scatter the nearest-negative correlation around zero (MODERATE confidence)

Dampening of the residual marker elevation does not require persona framing: in the primary log-prob read all four instruction-prompt paraphrases pull it down, two about as deeply as the doctor persona (MODERATE confidence)

No tested base-model predictor cell cleared the cross-behavior threshold on Qwen-2.5-7B's sycophancy, refusal, and emergent-misalignment leakage panels — but a base/Instruct substrate swap and a floored refusal arm both shape that null (MODERATE confidence)

Role prompts raise the base model's marker prior even on frozen answers it didn't write: the moving trained-vs-base audit baseline is prompt-driven and intrinsic to the base model (MODERATE confidence)

Benign-SFT's surviving marker elevation is context-anchored, not doctor-specific: all four tested persona prompts shrink it, mostly via a base-prior rise (HIGH confidence)

The marker-leakage channel structure is panel-specific: on held-out personas, geometry routes the training-induced change, and no channel claim transfers beyond the base end-of-answer level persisting through training (MODERATE confidence)

Dropping one contrastive negative did not raise marker leakage for bystanders close to it (MODERATE confidence)

An emergent-misalignment LoRA shifts all 14 persona contexts along one shared activation direction, a marker LoRA doesn't, and the contrast survives all three deferred controls plus a unit-norm re-read (HIGH confidence)

Training the marker implant deeper slides its geometry slightly but monotonically toward stricter rank-1 collapse, and the additivity-cosine read stays undiagnostic at all three dial points under marker-only-loss LoRA at lr=5e-6 (MODERATE confidence)

At a de-saturated anchor, both geometry signs of a single contrastive negative reverse — the barrier becomes anti-shadow, the anti-bubble becomes pro-bubble — and the reversal replicates on retrain (MODERATE confidence)

Scope the PostToolUse ruff hook to files whose git origin is explore-persona-space

Step 9c gate-fleet arbitration counts a session's own detached ledger refresh as a foreign contender, deadlocking that session's own gate

workflow-fix: cited-body gate labels Word:-prefixed content corrections 'frontmatter-only' (fm_line over-matches ordinary prose)

fix #847 torch-before-dotenv violations making test_no_new_torch_before_dotenv_vm_entrypoints red on main (fleet-wide Step 9c inheritance)

main-red: HF-routing/hub-retry lint + torch-before-dotenv scan fail on pristine origin/main, redding Step 9c fleet-wide

cron wrappers: bound the synchronous Telegram push calls (timeout) — class sweep from #2196

cron_step9c_ledger_refresh.sh is committed non-executable — #2114's nightly refresh has failed Permission denied for 18 nights

workflow-fix: adversarial-planner plan-finalize cited-body currency gate (cited parent corrected mid-draft goes unseen)

workflow-fix: Step 5a spec-sync add/modify-only leaves stale main-deleted family twins — TG scan leg collects them and crashes the Step 10d gate

cron wrappers: unchecked mkdir -p LOG_DIR silently skips the pass (sibling sweep of #2196)

test_bootstrap_pod_git_credentials.py extracts ssh payloads by first-quote scan (same shape #2360 r3 hardened); sweep tests/ for siblings

Cloudflare email obfuscation silently rewrites model-generation text in every eps.superkaiba.com dashboard (served bytes != committed bytes)

eval/capability.py: same narrow PackageNotFoundError catch breaks on the announced importlib.metadata KeyError

No check verifies a GPU-lane-routed phase actually allocates a CUDA context (P5 ran 4xH100 at 0 MiB)

poll_pipeline.py cannot detect terminal success of a single-phase dispatcher invocation — clean completion reports pid-stale-workload-live then decays to dead

Background per-stem Hub uploader: overlap capture-phase uploads with next-stem compute (shipped; adoption deferred to a future capture round)

workflow-fix: a plan can mark its ONLY source of acceptance evidence optional and still PASS verify_plan

curate .claude/agent-memory/code-reviewer-lean/MEMORY.md under the 24KB agent-memory index FAIL cap (41KB on main; reds worktree spec-freshness sync commits + silently truncates the lean twin's loaded lessons)

Fix three fleet-wide test invariants broken by sibling landings on main (ungated HF upload #1739, bind-census skips #2477, dangling import roots #2379/#2474)

workflow-fix: test_gates_full_shape red fleet-wide — clarify_experiment_ask gate missing from pinned set

Pod bootstrap cone sparse-checkout omits CROSS-ISSUE artifact cones, so drivers reading prior issues' committed eval_results/figures crash at launch

set-body: fail-loud on byte-identical no-op + gate body-edit markers on an actual body commit (phantom v4 edits, #2333)

git_provenance()'s 5s-timeout git status orphans .git/index.lock on slow filesystems, breaking every later git op in the run (fix: --no-optional-locks)

verify_task_body.py Lens 14 reads a worktree-frozen concerns.jsonl with no staleness guard (analyzer false PASS on #2378 r7)

workflow-fix: tick_triage reads STALE-REDRIVE at planning while a live subagent runs, risking a duplicate planner spawn

pod.py provision exits 0 with BOOTSTRAP-OK while its own preflight reports PREFLIGHT-FAILED rc=2 — fail loud instead

verify_plan.py: dry-run plan-embedded jq probes; gate contingent judge waves; require regen-time max_model_len re-pin

workflow_lint _hf_routing_file_errors false-positives on HF-call text inside docstrings

artifact-reuse: prove a reused GATE's committed anchor is reproducible under the shared core's CURRENT defaults (#1887 flip silently changed #2546's G-E meaning)

verify_uploads.py hardcodes HF_DATA_REPO — the upload gate cannot verify overflow-routed runs

worktree_audit.py stale-sweep reaps an active inline-round worktree (pod-side work + parked parent look idle)

test_no_new_torch_before_dotenv_vm_entrypoints red on main: 4 scripts missing load_dotenv, reds Step 9c gates fleet-wide

Clear 3 live-hf-retry-routing errors making no-flags workflow_lint FAIL on clean main (Step 9c gate red fleet-wide)

gotchas.md: MooseFS errno-116 partial install passes uv audit while import fails (probe-by-import + UV_LINK_MODE=copy recipe)

workflow.yaml markers registry: reconcile marker kinds mismatch with /issue review-loop practice

verify_task_body: footer check for SHA-pinned GitHub links on named code/artifacts

workflow-fix: consumer-supersession audit — enumerate committed consumers of refuted evidence artifacts and require supersession labels

Hoist a process-safe atomic-write helper into shared code and sweep the shared-tmp pattern out of scripts/

plan-compute-sizing.md: a 1-unit pilot is not a sizing basis when the phase's own design admits unequal unit costs (1.5x grid under-projection, #2329)

Step 10d merge-conflict-recovery block omits gh pr ready, so a recovery-path merge dies on the draft precondition and misroutes to the Failure bullet

select_step9c_tests.py: feature-named consumer tests escape module mapping; over-wide sweep_scope label hides the gap

workflow-fix: review/compose briefs check marker presence against the WORKING TREE, so a transient stash-cycle read produces a false marker-destroyed escalation

Step 9c 1b: a self-composed gate launcher can silently drop --files-only newline flattening OR -o junit_family=xunit1; neither omission fails where the recipe checks

Policy: rebase-rewritten own issue branch — sanctioned --force-with-lease arm, or explicit ban + force-free route only?

/issue Step 10d merge fence never binds $WT — git -C '' retargets every worktree op at the SHARED repo root (destroys a valid gate verdict; reaches a shared-root working-tree revert)

Codex twin: CLI + gpt-5.6-sol, daily auto-upgrade cron (landed)

verify_carryover_inputs.py default ref checks the wrong tree on a suffixed issue branch (Step 6a.5 gate)

Non-atomic fresh claim-file write in run_claim_queue kills a concurrent worker with 'unparseable claim file'

workflow-fix: implementer marker four-H3 completion-report contract is judgment-caught, not mechanically checked

Step 5a sibling arm: synced sibling tests run against branch-era src/; #2208 collection probe cannot see the runtime skew

workflow-fix: Step 9c script-file gate launcher — newline-separated test list collapses to rc=126 with zero tests collected and evades both anti-silent-pass guards

workflow-fix: verify_plan WARN check for sub-resolution judge-pilot sizing (per-arm draws below floor(1/threshold)+1)

workflow-fix: Step 10d Guard 5 trigger self-fires on a note that merely mentions the candidate token

Assemble qualitative-examples panels for the paper (discrimination + patching)

workflow-fix: /issue Step 5 dispatches code-review with no implementer marker posted

workflow-fix: codex-* composers leave their own agent-memory writes uncommitted (3x in one session)

Harden #2394 jbmine to paper-citable grade — code-review r3 fix (B2 exact-universe alias-collapse at :159)

Add codex-*-lean twins: autocompact-thrash ladder has no lean-twin rung for the Codex composer roles

workflow-fix: critic-lens-reference item 13 still says fellows-first auto default (stale since #2054/#2059)

Figure builders: emit tick/legend text to meta.json + WARN-lint opaque code-slug labels

Diagnose selective SIGTERM (rc=143) killer of in-session foreground python children on the shared VM

verify_task_body check 17: match Context footer against verbatim ## Provenance, not just non-empty

gotchas.md: fellows remote shell is zsh — unmatched glob aborts the whole SSH command (nomatch), silently truncating chained monitoring probes

upload_raw_completions_to_data_repo silently drops pre-merge shard payloads (15.6 MB of rollout text on #2223)

verify_plan.py: regen-headroom arithmetic check (armed ≥2× re-gen vs max_model_len pin) — third recurrence of the #505/#601 class

pod.py terminate: no surgical form for the PRIMARY pod while a suffixed sibling is live

Extend driver --import-check to bind-check in-repo helper call arity, not just args.<attr> completeness

paper_plots._extract_scatters drops genuine (0,0) single-point scatters from figure sidecars

Lint: cross-check production-branch third-party imports against uv.lock

Step 4a: draft-PR creation is gated on a condition that is false by construction (generator of the #2240 zero-PR class)

Step 5a spec-freshness: couple agents prose-pin tests (test_mapping_baselines_wiring_pins.py class) to the .claude/agents family — half-sync gate red

Absence-claim rule names the duty but not the mechanism: a full-listing + client-side prefix filter returns a silent zero on a wrong path where the scoped probe 404s

Step 10d merge window runs on a reap-eligible worktree; a mid-gate reap makes the pre-push lint gate certify pass vacuously

workflow-fix: Step 9c reports an external SIGTERM of the unfenced workflow_lint subprocess as a lint FAIL

Extend verify_plan.py with a decision-margin satisfiability probe (margin-vs-baseline ceiling)

Step 10d Guard 4 (lost-update refusal) is vacuous on the prescribed call path — --main-sha passes the main TIP, helper reads it as the merge-base

Step 9c paired oracle: add a SUFFIX-replay arm so later-sorting-file contamination is attributed mechanically

daily-fix: test_issue1315_dispatch pollutes the fu3 context registry — order-dependent Step 9c gate bounce

Test-order contamination: later-sorting module registers wildchat conv prefix into CONTEXTS (fails test_conv_context_is_wildchat_family in multi-file runs)

Briefs pointing a worktree-based reviewer at task state must pin the absolute main-checkout path (worktree plans/plan.md resolves to a stale version)

Step 5a sibling-sync satisfiability probe is --collect-only, so runtime API skew and function-body imports pass it

workflow-fix: pre-review deliverable-divergence probe (Step 5a/10d)

SLURM rsync lanes omit tracked data/ inputs — lane sibling of #2211

verify_plan: WARN when a plan registers a test-retest kappa demotion gate while pinning judge temperature 0

verify_plan c46: also enforce the launch runtime requirement (exactly one of --workload-cmd / --hydra) when dry-parsing plan-embedded dispatch commands

workflow-fix: raise the per-reviewer review-round cap from 5 to 10 across all three review loops

workflow-fix: reclaim tier for ~/.eps-slurm-src terminal-issue staging (112 GB unreaped)

task.py set-body: Goal-H2 drop guard is now workflow-aware — v2 report bodies auto-exempt (merged PR #2023)

verify_report.py: scan the detailed companion writeup, not just the report body

Pilot: cheap context readout to mine always-comply jailbreak contexts; does the context→answer map help?

Emit a uniform phase field in reproducibility cards so gate checks can resolve phase-to-card mechanically

Repack the 10 largest prefixes in the HF data repo to recover ~610k of the 1,000,000-file cap

cron_lesson_consolidate: an uncreatable log dir silently skips the whole nightly pass

verify_report.py: WARN when a report cites a repo artifact at a SHA the same branch has since modified

workflow-fix: tmp-scratch sweep blind spot — stray top-level /tmp/{pyproject.toml,uv.lock} poisons every /tmp-cwd uv run fleet-wide

workflow_lint: flag stale gotchas.md pointers after an entry relocates (live instance in research-project-structure.md)

workflow_lint: no gate detects merge-conflict-marker residue in tracked workflow-surface files

CPU-intent RunPod provisioning: type the no-port wedge, rotate DC, and document a residual route when every CPU DC is dry

Remediate 19 scripts/ drivers with repo_root()-derived sys.path inserts, then widen the #2181 check to scripts/

workflow-fix: UnicodeDecodeError escapes 112 json-guard except tuples (53 in the watcher, all on read_text paths)

verify_plan.py: validate a plan's stated smoke-fixture size against realized fixtures, and flag a smoke claim whose producing script is absent from the modified-file list

poll_pipeline ETA tripwire self-disables on plan §9 calendar-latency rows; range cells silently parse as lower bound

Headline matched-partial needs a support-restricted companion when the matching covariate is degenerate on most of the sample

Fact-checker: exact-identity premise from a sample must be verified at the assert's grain

git_provenance() blind to untracked producing scripts — artifacts stamp a commit that lacks the code that made them

guard_root_code_commit.sh: disarm text-detectable prefixed cwd-movers outside the NAME=value set (+= append-assignment, time/wrapper prefixes) + complete residual-note

Preflight passes a pod venv whose lock-pinned package has no metadata and cannot import (sympy half-install breaks the whole transformers stack)

workflow-fix: test_backend_poll module-mode test fails only in large collections, always classified NEW by the Step 9c oracle

code-style resume-key clause: forbid hashing raw bytes of a RECOMPUTED float array (platform-dependent key discards valid checkpoints)

Pre-split intermediate-unit review guard + cross-session shared-worktree writer arbitration

guard_root_code_commit.sh: pathspec-limited root commits blocked by foreign staged code files (scoping defect)

verify_uploads.py outroot-residue: basename match vs issue-scoped git trees lets a sibling leg's committed file cover this leg's unpersisted file

Step 5a spec-freshness: add tests/issue_skill_source.py to FAMILY_workflow (half-sync reds test_issue_skill_* collection)

Experimenter launcher composition: mandatory per-leg out/scratch dir isolation for concurrent same-driver legs (#2330 fu1 dense-store poisoning)

verify_task_body: WARN-level per-unit-or-exemption check for every v4 result section (multi-point aggregate figures escape the current single-aggregate-stat trigger)

Repack small-file HF prefixes to reclaim ~610k file-count slots without breaking readers

workflow-fix: Step 10d pre-push gate mapped-test legs compare failure sets over different collections (false block on pre-existing main-red)

gotchas.md: warn that an interrupted posix_fallocate quota probe leaks its full allocation (12.6 GB leaked on a live pod, #2329)

Cut ~35K tokens/spawn from the always-on context load (autocompact thrash)

workflow-fix: main-red — post_marker_echo test stubs lack set_status's new if_plan_v kwarg (4 tests)

workflow-fix: ruff-policy pin red on main — autonomous_session_watch.py gate_push_pass C901 (c0d76f99c9)

judge_pilot_gate: unsatisfiable-config remedy text names escapes that cannot satisfy the verdict-time floor

gotchas.md: MooseFS preflight fallocate canary grinds ~9 min (slow-canary vs read-wedge discriminator) + resumed-pod uv sync flash-attn uninstall note

Scope preflight's vLLM/transformers skew check to workloads that actually import vLLM

gotchas.md: RunPod host-driver vs CUDA-major wheel mismatch — cuda-compat forward-compat recipe

Rule-26 judge pilot gate must exercise the same TRANSPORT as the wave it gates (sync pilot passed a batch wave that lost 34% of draws)

Register eval_results/issue_1739 cone in tests/sparse_cones.txt (deterministic sparse-worktree test FAIL, #671 class)

Codex reviewers' "Concerns to persist" never reach raise-concern, so the next round's prior-concerns gate walks an empty ledger

workflow-fix: upload-verifier reconciles self-reported row counts, not realized store contents (passed a 25% shortfall)

workflow-fix: .claude/skills/issue/SKILL.md is 813 bytes from its hard FAIL cap — next edit wedges the Step 9c gate fleet-wide

workflow-fix: symlink/FIFO-safe bounded opens for the three advisory-flock lock sites + thread realized lock mode into codex_task failure notes

daily-fix: inline override duty list misses tick cron

daily-fix: dispatch_issue launch lacks name-suffix/gpu-type

workflow-fix: serialize Codex twin dispatch against the repo-shared non-atomic job index + post-spawn probe retry

daily-fix: null-statistic gates need measured calibration

HF file-count overflow fallback never fires for dataset repos or bulk uploads — data repo hit its 1,000,000-file hard cap

daily-fix: dispatch lease races (4 dup spawns + cap 6>5)

daily-fix: bare-API blinded-read recipe as a rule

workflow-fix: Step 10d Guard 3 ON_MAINLINE uses first-parent reachability, so a sibling's merge-form landing false-flags later branches UNSAFE

daily-fix: subfloor reclaim never fired + local-date logs

daily-fix: watcher daemon-liveness pass (3h17m outage)

daily-fix: tick_triage counts compact rows as human

workflow-fix: step9c violation-path extraction takes every path token per line (snippet false-block residual)

workflow-fix: 9a-ter inline payload lint gate — violation-grain classification for whole-repo scan nodes red on pristine main

workflow-fix: Step 9c gate-set cross-check validates cardinality but not splice shape — a newline-separated --files-only list runs 1 of N files behind a green guard

workflow-fix: step9c compare classifies whole-repo SCAN-test nodes per-NODE, so a branch ADDING a violation to an already-red scan node reads as pre-existing

Step 10d lint gate's mapped-invariant baseline leg runs pytest in the shared repo root — #2015 stash cycle kills it and false-NEWs pre-existing reds

Step 10d safe-case push has no arm for a REBASED issue branch — rejection fallback rebases onto the stale remote branch (#1128 replay shape); stale PR head can land pre-rebase payload

issue2225_fu2_dod_points_fig.py imports numpy before load_dotenv — reds test_no_new_torch_before_dotenv_vm_entrypoints on main

/issue Step 5a spec-freshness sync: sync synced-spec data dependencies + fail loud on a failed sync commit

workflow-fix: Step 5a sibling sync puts main's own files in the branch diff, disabling the #2024 ordering carve-out and false-BLOCKing Step 9c

workflow-fix: Determine which side regressed before touching either. EITHE

Step 9c pristine oracle cut from repo-root HEAD instead of the resolved diff base — root-divergence window yields false NEW classifications

workflow-fix: sync_repo_root deadlocks on the abandoned-rebase husk it exists to clear (task.py holds the flock while waiting on the husk)

daily-fix: AGENT_SPEC_SIZE_GRANDFATHER is a chronic merge-co

audit_clean_results_body_discipline: extend pre_reg noun grammar (recompute/sensitivity/comparison)

#2223's merged r2/r3/r5 scripts red 5 fleet-wide gate nodes on origin/main — incl. the no-flags workflow_lint bundle

verify_task_body cross-issue-reuse-pins check SKIPs on pre-merge worktree bodies — its only gate

daily-fix: stale prior-round PASS in verdict-present check

daily-fix: reap-husks symlink plan.md reads as unique

workflow-fix: extend owner-fence awareness to the watcher's pod-safety STOP pass

workflow-fix: Step 10d Guard 2 asserts the pre-#1723 status ordering (says code paths are at completed; #1723 keeps them at running)

workflow-fix: SKILL.md plan-review-floor trigger misattributed to is_workflow_fix_session

verify_task_body: re.IGNORECASE + .lower() crashes the verifier on Unicode long-s (ValueError at :16820)

workflow-fix: land the shared fit-core memory chunk cap on main + decide PR #1717 disposition

verify_task_body: flag positional cross-references in Results that misresolve after fold reordering

workflow-fix: bootstrap_pod.sh python shim deadlocks uv interpreter discovery on pod venv rebuilds (uv.lock branch change)

Pod terminate guard is owner-blind: a non-owner can self-post the upload-verification PASS and destroy a live owner's pod

daily-fix: pod HOLD gates + launches must arm a re-drive

daily-fix: pre-dispatch staleness pass for queued tasks

workflow-fix: verify_plan.py — §9 backend pin-claim vs frontmatter + declared GPU width vs launch-command width flags

daily-fix: measured n_train + regime-change pilot basis

workflow-fix: SLURM --mem is GPU-count-derived with no per-dispatch override (memory-bound phase OOMs)

poll_pipeline emits status:dead while its own gpu_util + log-mtime evidence shows the workload alive

daily-fix: issue1491 ceiling must assert n_pairs

daily-fix: normalize claude-fable-5[1m] in settings.json

workflow-fix: verify_task_body opaque-code scan skips sidecar-less Results-embedded figures

gotchas.md: fellows CPU analogue — plain-SSH nproc sees the node, inside-job nproc sees the allocation (ConstrainCores=yes vs ConstrainDevices=no asymmetry)

main's .claude/skills/issue/SKILL.md exceeds its own size ratchet (936,191 > 935,400 bytes) — trim below cap; blocks lint gates fleet-wide

scripts/issue2224_fold_figures.py breaks test_shared_vm_thread_caps on main (missing load_dotenv-before-torch preamble)

daily-fix: spawn stop --kill must TERM inner claude pid

verify_task_body: WARN checks for fold-staled fixtures (per-unit-exemption token; acknowledgment result-count)

workflow_lint SLURM_GPU_WIDTH_GUARD_RE does not recognize an inherited-CUDA_VISIBLE_DEVICES parse, so correctly-fixed dispatchers cannot exit the grandfather list

Step 10d lint-gate single-flight probe --pattern 'issue-<N>-lint-gate-tree' has CLEAR windows under the #2115 script-file launcher — completion-read false-reads a live gate as died

Amendment-style plan versions break plan.md readers, verify_plan, and the GPU-hours gate

poll_pipeline pid-file staleness check is not pod-scoped — disables the #813 liveness check on all but the newest leg of a multi-pod issue

Extend verify_task_body.py figure-text opaque-code token grammar with the single-letter-prefix arm-slug class

verify_task_body check 17: upgrade context-parent-lineage-mixed WARN to FAIL when frontmatter parent_id is set and the Context row carries a no-parent clause

gotchas.md: dispatcher deriving GPU width/pins from nvidia-smi instead of the SLURM allocation targets unallocated GPUs on narrow allocations

SKILL.md Step 4: brief-composition rule covers marker SUPPRESSION but not DISPLACEMENT by a return-format contract

Wire the absolute per-cell trainability gate into scripts/issue2225_train.py (the sibling entrypoint #2242 left ungated)

daily-fix: sweep /tmp gate scratch; HF-backedness gate

shared_fold_map.json on main is a 1,761-conv SMOKE map, not the 26,889-conv production map — guard the load

SKILL.md Step 6: state the bg-Bash timeout floor for free-SLURM-lane launches (420s park budget vs 120s Bash default)

daily-fix: five gate-recipe defects (Step 9c/10d)

Absolute per-cell trainability floor: a yield shortfall must DROP the cell, not shrink it to 1 row

daily-fix: judge pilot sizing + classification rubrics

daily-fix: four verifier gaps (lint + verify_plan)

daily-fix: pod-side liveness probe before LAUNCH FAILED

daily-fix: root-branch guard ssh/cd-form false positives

daily-fix: address-concern --summary-file + --dry-run alias

daily-fix: CVD-scoped nvidia-smi + schema-from-artifact

Step 10d: payload-aware no-PR arm (a branch with novel payload and zero PRs silently skips the auto-merge)

daily-fix: manifest-first staging for sharded artifacts

daily-fix: reconcile-stop misses error-storming sessions

daily-fix: test-order polluter test_issue1482_densesae

Payload-scope the inline lint gate (9m08s measured -> target <60s for analysis-script payloads)

fix red main: issue2220 script pair — 4 unguarded upload_folder calls + heavy import before load_dotenv

pod.py provision --wait-for-capacity is structurally inert for CPU intents (CPU branch returns before the flag is read)

daily-fix: sibling sync arm misses .sh dispatchers

No gate checks that a plan's backend lane can mint N distinct pod names for an N-way fan-out (RunPod lane is per-issue; --lane-suffix does not cover it)

daily-fix: Step 10d lint-gate pending-call wedge

Add the missing FAN-OUT STAGING EXTENSION to critic lens item 16 (plan-compute-sizing.md:330 points at nonexistent lens text)

verify_task_body check 31: normalize plan-figure prose tokens to PNG stems (descriptively-named committed figures escape the body-mention check)

Narrow over-broad scripts/issue*.py rule globs (3 rules, ~23K tokens on every issue script)

verify_task_body: verify artifact-content claims (per-unit values recorded in <pinned JSON>) against the pinned artifact's actual structure

daily-fix: nightly step9c known-red ledger refresh cron

Judge parser drops prose-then-trailing-integer responses (recovers ~8-28% of dropped draws)

daily-fix: ownership-probe hardening (3 probed gaps)

workflow-fix: Disk-hygiene lead bullet — bind status runtime-checked, not asserted

workflow-fix: no-lost-row check on agent-memory index alignments

workflow-fix: scope preflight branch-freshness fetch, raise timeout

workflow-fix: mechanize local-disk headroom for staging + VM fits

workflow-fix: sweep /mnt/eps-data/$USER staging in disk janitors

#2098 workflow-fix: clarify --map-files takes one list-file path · completed

gotchas.md: extend the #779 shallow-clone git-log class to in-driver ancestry GATES (shallow-aware gate recipe)

workflow-fix: cover relocated HF_HUB_CACHE in guard HF-cache tiers

Statistics lens item 17: rate-denominator provenance audit for projection rates (shipped)

verify_plan c53: extend harm-class trigger vocabulary to evil-trait/toxic-corpus judge-banding phrasing (#2221 silent pass)

workflow-fix: ruff-clean alias for the phase-done waiver token

verify_plan.py: check for inherited argparse row-count defaults under-covering the plan's registered per-cell target n (#2054 --target-conv-ids class)

workflow-fix: wire the both-arms mapping rule into Statistics lens item 15

workflow-fix: empty selection over local artifacts must fail loud, not exit rc=0

workflow-fix: teardown prose contradicts the compute-kill approval gate

Bound the session-auto-respawn lane: 7 respawns, zero progress

workflow-fix: lint launcher GPU-width derivation without a SLURM branch (#1902 class)

workflow-fix: worktrees are a required location class in absence sweeps

workflow-fix: wall-clock heartbeat duty for subagents owning detached work

workflow-fix: split code-review into per-commit sub-verdicts for large rounds

verify_task_body: assert concern-deferred body comments have a matching deferred ledger event

workflow-fix: pod audit terminates teammate pods (substring ownership + wrong stale clock)

Clear the api-dispatch-routing lint FAIL on scripts/issue1482_blind_read_api.py (route through api_dispatch or add the documented waiver)

workflow-fix: register project-level -lean twin agents so the v20 autocompact-thrash remedy actually works

workflow-fix: on-policy datagen must persist judge-rejected

Sparse-worktree gates false-block on external/-rooted test fixtures: register external/jacobian-lens (+sweep) in the sparse-cone registry

verify_task_body.py check-20: WARN when conciseness WARNs fire with NO acknowledgment paragraph at all

Invert auto lane order: RunPod primary (GPU+CPU), fellows next

workflow-fix: root-commit guard is cwd-blind — worktree comm

workflow-fix: registry drift at an allocated id breaks task.

daily-fix: judge max_tokens pins sit below the raised rule-2

daily-fix: DATAGEN_JUDGE_MAX_TOKENS=500 sits below the rule-23 floor (live library pin, mt-keyed cache)

workflow-fix: bootstrap-failed pod left billing (provision no teardown + keep-running blocks reaper)

pod bootstrap sparse-cone excludes tracked data/ inputs → fresh-pod FileNotFoundError (from #2203 Phase 3)

workflow-fix: document PostToolUse formatter race (edit-then-run yields stale artifact, exits 0)

workflow-fix: list-ephemeral hides duplicate-named pods (name-keyed state)

workflow-fix: thread lane_suffix into slurm.job_name (concurrent lanes silently reconnect)

daily-fix: per-family pilot basis for fan-out sizing

daily-fix: verify_plan flags bare lane env vars in s10

daily-fix: root-commit guard honors excluding pathspec

gotchas.md: parallel lm-eval fan-out races library nltk download — pre-stage once before forking (from #2203)

daily-fix: projected-vs-realized + fallback-slice claims

daily-fix: runbook markers carry verified-by ran|read

daily-fix: device-domain smoke in artifact-reuse checklist

daily-fix: fan-out completion contract in spawn briefs

daily-fix: preflight disk probe uses writable dir

daily-fix: 9a-ter shard-axis + checkpoint-cadence duties

thread-caps VM-entrypoint test blind to uncommitted scripts (false-green #847 regressions)

Step 5a sibling-issue test sync can import a main-NEW test with branch-era-unsatisfiable src imports — deterministically reds the 9c gate as NEW (#2206 incident)

api_dispatch silently converts a text-block-free API response into an empty-string SUCCESS (7/10 Fable calls lost in #2202)

llm-judging.md: add rule 28 (api-refusal drop class) to Statistics-lens Enforcement riders + matching lens sentence

daily-fix: inline lint gate defers under high VM load

daily-fix: terminate pod on failed RunPod fallback launch

daily-fix: trigger-dense ingest hardening (refusal results)

daily-fix: post-autocompact relaunch marker-triage mandate

daily-fix: backend_poll reachable-wedge escalation arm

daily-fix: stand-down confirmation + ownership release

daily-fix: RAM floor required on wide fan-out dispatches

#1689: fold user-slot-recapture readout into body + correct refuted tier-1-real-data claim

workflow-fix: step9c gate stale-__pycache__ purge

workflow-fix: carry-over gate path regex misses include-tree

workflow-fix: uniquify pod_disk_guard probe filename + gotchas entry (probe race)

daily-held: Digest-only guard arm for real-world-corpus raw_

daily-fix: Step 9c gate — detached first attempt, end the tu

Remove (or neutralise) the Step-2c 100 GPU-h autonomous plan-approval gate

workflow-fix: verify_plan edited-literal pin-test check

daily-fix: fail loud when --min-ram-gb is inert on a GPU dis

workflow-fix: phase-done lint test hangs at interpreter exit

daily-fix: trigger-dense first-spawn pre-qualification by ta

workflow-fix: warn when SLURM launch fence exceeds the intent's sbatch time bin

daily-fix: planner names measurement grain (unit of analysis

workflow-fix: step9c compare — branch-untouched-file NEW failures classify ordering-suspect, not blocking NEW

workflow-fix: purge stale gcp-pin remediation guidance superseded by #2028

workflow-fix: scratch-eligibility for the workflow-lint scan node in step9c compare

workflow-fix: figure-PNG-vs-own-sidecar render consistency check

verify_report.py: assert committed-under paths resolve at the pin, and assert code-SHA rows match per-phase reproducibility cards

workflow-fix: shared-root uncommitted-file reversion race (pre-commit stash/restore under concurrency)

SKILL.md prose asserts setsid orphans reparent to PID 1; user systemd is a child subreaper

Step 9c gate red on main: setsid test asserts PID-1 reparent, but user systemd is a child subreaper

daily-fix: /issue Monitor until-loop composition guidance

Make pre-teardown verification sweep the out-root TOP LEVEL: three artifact losses on one task, each caught only by a manual name-set diff

Decide the gotchas.md byte budget: a real re-trim pass or a deliberate cap raise

cron_lesson_consolidate: surface consolidate_lessons.py exit-3 budget refusals

argcheck: a **kwargs splat escapes dest derivation at all three registration sites (add_argument / add_subparsers / set_defaults), producing a false-positive import-check failure

Document the RunPod HF-CDN zero-byte large-blob trap (206 + 0 bytes) and its curl discriminator + rsync-relay recovery

Smoke-architecture gate: per-arm enumeration is hand-listed, so the REAL-or-N/A invariant is unfalsifiable

daily-fix: root-commit block: warn never report as landed

sync_repo_root: fail loud when a pull --autostash is left unrestored (silently reverts concurrent sessions' uncommitted work)

workflow-fix: Split the pin across the two post-compaction sites — assert

Ban repo_root()-derived sys.path inserts in tests (worktree runs silently exercise main's code)

Wire the authorized-stub grant token for the Step 6d.0 smoke-architecture gate

verify_plan.py: WARN when a §9 compute row's own basis arithmetic contradicts its booked GPU-h or its pilot-abort threshold

Autonomous plan-gate cap default is 24 in task.py but 100 in every doc and park message

Parenthesized §9 planned_wall_h cell silently disables the poller phase-ETA tripwire

verify_task_body check 31: WARN on any committed-but-unmentioned figures/issue_<N> PNG (not only per-unit patterns)

daily-fix: finally-block raise must never mask in-flight exc

workflow-fix: enumerate smoke blind spots — gates/paths a smoke run structurally cannot reach

workflow-fix: pin MALLOC_MMAP_THRESHOLD_ in launch-prefix rules

#2166 workflow-fix: 5 stale CLAUDE.md prose pins repointed at their rule-file homes — main green (347 passed)

daily-fix: corpus-replay gate test reds stale branches on fl

daily-fix: provisioner avoids a known-bad RunPod host on ret

dispatch_issue.py launch on the fellows lane: 600s SIGTERM disarms the QoS ladder; plan-copied commands drift (--repo-branch defaults to main)

workflow-fix: Repoint the pin's read target from CLAUDE.md to .claude/rule

keep-running wedged-owner arm cannot fire on a busy task: add a pod-grain idleness leg (#1739: 3 pods idle 19.6h behind 129 markers)

daily-fix: python-pipe guard exempts ssh-quoted remote argv

workflow-fix: pods.md names a --boot-disk-gb flag pod.py provision lacks; cpu-bigmem defaults to 50GB not 240GB

workflow-fix: repair main-red pin test_mlp_chunk_size_changes_fit_batch_order_init

daily-fix: capture plans name pooling convention per vector

Detached HF transfers must have timeout + observable progress (xet read-hang is silent); warn off snapshot_download(allow_patterns) on the 1M-file data repo

workflow-fix: #1739 follow-up-run marker writes label= not followup_label=, reddening test_corpus_replay_all_historical_markers on main

Give API-level stop_reason=refusal its own judge drop class (rules 9/24/18 + judge_dispatch.py:670-681)

daily-fix: worktree_audit gc-wedge tier (git prune + gc.log)

wf-compact fallout: codex-code-reviewer.md lost the Step 4.6 gate-scope copy-list bullet — 3 pin tests red on main

daily-fix: healthcheck auto-launches /daily backfill

daily-fix: AUTO_COMPACT_WINDOW not reaching spawned sessions

workflow-fix: triage window boundary opens at post time, hiding enumerate-to-post seam markers

workflow-fix: bootstrap sparse-cone block dies, killing missing-cone warning

workflow-fix: reconcile tests/test_workload_cmd_env_lint.py's mock-backend

workflow-fix: verify pinned one-off installs resolve + dist-name != import-name

daily-fix: fix red main: test_workflow_lint.py::test_check_upload_return_discard_live_trees_pass — capture each helper's return and raise o

daily-fix: pod-side-reporting: hand-rolled box teardown leg

workflow-fix: update the two test files' mocked auto-chain fixtures to the

batch drain loses bare-numeric judge verdicts (#1434 parity not ported): 64% vs 2.8% drops

workflow-fix: AST lint for discarded upload-helper returns (scripts/)

curate near-ceiling agent-memory indexes (merge duplicates, retire superseded)

workflow-fix: Step 6d.2 tick-source comments omit quiet-wait

workflow-fix: flaky settle-sleep equality assert in inline-l

workflow-fix: add eval_results/issue_1481 to tests/sparse_cones.txt (issue-1947 tests collect in sparse worktrees)

daily-fix: RunPod fallback rung pod lifecycle: orphans + sto

daily-fix: fix red main: test_issue1689_os_exit_shutdown_bypass.py::test_entrypoint_calls_os_exit[issue1689_real_u2_from_1738.py] — conform the script's `__main__` guard to

daily-fix: teammate report channel + deliverable content pro

workflow-fix: lint cross-module sha-pin domain agreement (#1776/#1491 class)

workflow-fix: widen check-28 figure token pattern to arm slugs

daily-fix: smoke contract: resume/salvage-leg matrix + real-

daily-fix: sizing duty binds teammate-dispatched box sweeps

daily-fix: HF fan-out pre-stage rule + FLEX preemption escal

workflow-fix: no-progress tick wedge — heartbeats read HEALTHY, stale-redrive reuses exhausted context

workflow-fix: sidecar canary tolerant of foreign concurrent

workflow-fix: pin CLAUDE_CODE_AUTO_COMPACT_WINDOW in project settings + document the reduced-window thrash diagnostic

daily-fix: crash-fix: shared-module propagation + stale disp

workflow-fix: blanket-add latch misses equivalent-blanket sp

workflow-fix: letter-arrow transition token class for figure opaque-code check

workflow-fix: lean-context subagent twins — thrash ladder ha

workflow-fix: cross-session lane pivot must resolve the owni

workflow-fix: flag dirty-tree provenance in result-JSON repr

workflow-fix: carry-over gate bare-name channel — HF-staged inputs WARN not FAIL

workflow-fix: re-arm bg poll chain first on every wake (message consumes tick wake)

workflow-fix: pgrep without -f under-matches argv, reads live process as dead

daily-fix: extract Step 10d guards to checked-in script

workflow-fix: pod bootstrap full-clones 10.5GB, burns ~$15 idle GPU per provision

workflow-fix: gotchas #952 streaming SIGABRT recipe misses break-suspended iterator (it.close())

workflow-fix: branch-sync already-at-tip short-circuit + recalibrated mutation caps (runpod.py)

daily-fix: pre-split trigger for fits+figures+smoke units

daily-fix: pooling-convention row in mapping disclosure

daily-fix: map fit-space parity assert (issue1739 natpv)

daily-fix: widen pre-gate spec-freshness sync set

daily-fix: report TTY-attached unmapped idle sessions

daily-fix: inline-round pod upload-verify satisfier

daily-fix: hold subagent respawn past active 429 storm

daily-fix: fellows stall reads need grace window

daily-fix: per-tier pilot walls + c12 screen vocabulary

daily-fix: never push to a branch live boxes sync mid-run

daily-fix: route long judge drivers off VM on 1st kill

daily-fix: dispatch preflight staging+flag probes

Remove GCP compute backend entirely (GPU + CPU): auto chain, explicit pins, cpu-bigmem RunPod re-map, tests, docs (user directive)

daily-fix: guard family syncs scripts/guard_*.sh with pin te

daily-fix: cross-issue lint-gate concurrency arbitration

workflow-fix: curate experiment-implementer agent-memory index under 24KB FAIL threshold

workflow-fix: verify_plan lattice parser — negated-existence atom

workflow-fix: pre_reg audit gains pre-set/declared/specified synonym branch

workflow-fix: inline lint gate purges stale pycache before its pytest sweep

workflow-fix: gotchas.md — numpy argsort SIMD tie-order trap for cross-machine top-k identity gates

workflow-fix: audit inline-CI rule misses bracket-less verbal form (CI <num> to <num>)

workflow-fix: Step 10d lint gate own-diff attribution matches message-cited paths (false block)

workflow-fix: round-unique inline lint-gate payload path

workflow-fix: root-commit guard pathspec escape fails under foreign staged files

workflow-fix: judge instruments round-trip the parse contract before dry-run counts

workflow-fix: export EPS_GIT_SHA in SLURM sbatch env for rsync-lane provenance

Hoist shared-node vLLM util resolver into a shared module (dedupe #1902 + #1345 copies)

workflow-fix: figure-sidecar opaque-code check misses bare P/M candidate codes

workflow-fix: gotchas.md entry — SLURM git-less scratch crashes metadata git shellouts

workflow-fix: gotchas.md entry — VAST /workspace EBADF from posix_fallocate

workflow-fix: inline_lint_gate false-blocks payload on PASSING-test captured tracebacks

Persist stop_reason per judge draw + truncation-vs-content drop split + rule-26 pilot-gate helper

workflow-fix: widen check-30 count-noun vocabulary (chunks/sidecars + modifier-separated)

workflow-fix: exclude plan-declared outputs from carry-over bare-name resolution

workflow-fix: close-miss escalation clause for the on-policy yield floor

Fix thread-caps main red: issue1773_register_steer_stats imports matplotlib before load_dotenv

update #1345 Takeaway bullet 5: slot-position reads re-fit by lambda audit (deficit survives); every-layer statements stay ambient-only

daily-fix: pod bootstrap retry + fail-loud provision verdict

correct #1310 Takeaway bullet 3 + title qualifier per nd-estimator/lambda audits (per-turn deflation was a lambda-selection artifact)

daily-fix: 9a-ter inline routing (50GB->bigmem, fits->GPU)

daily-fix: detached relaunch + worker-pid + named-verify

daily-fix: git guards cwd/compound awareness (ssh, worktree)

workflow-fix: dropped-at-gate condition placement check (Takeaways + result) in verify_task_body

daily-fix: canonical gist-update recipe (gh api PATCH)

daily-fix: SLURM intent tables cover all GCP GPU intents

workflow-fix: repro-artifacts-clean gate vs working-tree git status

workflow-fix: pm_inline rule for value-plus-minus-err in body audit

daily-fix: Monitor quiet-wait branch for Step 6d.2 poll

daily-fix: pre-commit hook for agent-memory index size

daily-fix: four CLAUDE.md discipline clauses

daily-fix: living_docs.py bounded index.lock retry

daily-fix: pilot at production shape; sampled GPU-util claim

daily-fix: route-1 fix closes the open task it credits

daily-fix: set-status bounded index.lock retry before raise

daily-fix: judge max_tokens floor 600 for JSON rubrics

workflow-fix: widen the leading version-stamp stripper to also accept the

daily-fix: rsync_covered checks RSYNC_EXCLUDE_PATTERNS

daily-fix: widen body-audit condition-label regex H1c forms

workflow-fix: SLURM rsync races the live shared tree — partial sync shipped 3/4 committed entrypoints; verify post-sync or archive-snapshot the source

workflow-fix: fact-checker verifies realized grain of row-grain reuse rows

workflow-fix: gotchas entry for function-scoped import-check name shadow (UnboundLocalError class)

fix order-dependent test-state leak: test_issue1335_ladder poisons test_issue825_mlp_batched_parity in same-session runs

workflow-fix: flag malformed footer Reused bullets escaping the form-keyed check 37

workflow-fix: verify_plan base-predictor-vs-change-DV companion check

workflow-fix: verify_plan check-ignore on declared-committed paths

workflow-fix: Step 10d probes PR state; verify landing before verdict consume

workflow-fix: fleet-wide pod code-sync pulls origin main over live issue-branch workloads (mid-run ENOENT race)

workflow-fix: reconnect handle writes spec-default provisioning_model, disarming FLEX_START queue-timeout failover

workflow-fix: GCE crash-persist omits analysis_tensors staging dirs

workflow-fix: add skill-pin family to Step 5a spec-freshness sync (test_issue_skill_* pin tests)

workflow-fix: gotchas cross-machine-seam entry — SLURM rsync lane drops committed eval_results

workflow-fix: pod-side results-git push must fetch+rebase (mid-run branch-push race)

workflow-fix: TG file-grain gate arm — tree-prefix normalization + warnings-line exclusion

workflow-fix: verify_task_body judge-health drop-line population reconciliation

daily-fix: compliant git-commit forms at snippet sites + blo

workflow-fix: verify_task_body per-result multi-figure (non-pair) check

workflow-fix: gotchas — HF download-accel failure matrix (scope disables per workload)

daily-fix: exempt path-limited git add --all in guard

daily-fix: slurm fetch_results atomic rsync pull

workflow-fix: per-pod pod-safety predicate for multi-round issues

daily-fix: /issue Step 0 preloads TaskOutput/Monitor schemas

daily-fix: /daily filed bodies cite marker per mechanism cla

Fence-tolerant judge JSON parse: 2.02% of judged calls drop on markdown fences, not truncation

daily-fix: register step 5a in workflow.yaml steps enum

functional_role axis: kappa 0.318 reproduces at both scales — the only axis with no contrastive evidence

daily-fix: fix red main: test_issue1689_os_exit_shutdown_bypass.py::test_entrypoint_calls_os_exit[issue1689_fits_mirror.py] — Convert the __main__ guard of scripts/is

daily-fix: fix red main: test_shared_vm_thread_caps::test_no_new_torch_before_dotenv_vm_entrypoints — load_dotenv before heavy imports in issue1773_register_steer_stats.py

daily-fix: fix red main: test_workflow_lint_jsonl_splitlines.py::test_live_tree_passes — Fix the `.splitlines()` JSONL read at `s

workflow-fix: inline lint gate blind to untracked payload files

workflow-fix: document the 'CMD -c ... || echo 0' double-print trap in gotchas.md

daily-fix: Step 10d teardown re-runs sync on in-flight advis

daily-fix: sync KEPT-stash outcome gets durable surfacing