Adding contrastive negatives suppresses the marker implant below its same-T positives-only twin at every dose, but the gap is non-monotonic (smaller at 32:1 than 16:1, largest at 64:1) and the negatives' loss channel stays dead at every level (MODERATE confidence)

Methodology: docs/methodology/issue_622.md · gist

Human TL;DR

Headline. more contrastive negatives suppress the implant below its same-T positives-only twin at every dose — but the gap doesn't grow monotonically: in BOTH log-prob and EOS-margin spaces it's smaller at 32:1 than at 16:1, then largest at 64:1. EOS-margin sits outside the frozen co-landing band at all three levels; log-prob crosses the band only at 64:1 (seed-mean, by 0.22 nats). the negatives' own loss channel never wakes up at any level, so the suppression isn't loss-competition, it's a non-loss residual (candidate: batch composition).

Takeaways.

• the curve doesn't crash, it slows. at 64:1 the implant builds to ~6.4 nats by step ~130 and just sits there for ~683 more steps; the twin keeps climbing past it.
• the gap is NOT monotonic. seed-mean log-prob D walks −4.76 → −3.54 → −5.80 nats (16:1 → 32:1 → 64:1); margin D walks −3.16 → −2.64 → −4.03. 32:1 relaxes back toward the twin, then 64:1 jumps past 16:1. per-seed: seed42 at 16:1 already crosses the log-prob band (|D|=5.90 > 5.58); seed137 at 64:1 doesn't (|D|=5.46). only the SEED-MEAN log-prob D at 64:1 lands cleanly outside the band — by 0.22 nats. read it as "suppression at every dose with a non-monotonic gap that crosses the band reliably only at 64:1", not "widens with dose".
• the negatives stay loss-dead at every dose. neg-row trailing-token CE is ~1e-5 nats vs the 1e-3 wake-up threshold. so whatever is degrading the implant, it isn't gradient on the negatives' loss term — it's downstream of that (best candidate: batch composition / data ordering).
• bystander leakage fraction is flat across all 12 cells, both in log-prob (~0.35 trained, ~0.45 held-out) and in EOS-margin (~0.37, ~0.45). more negatives don't sharpen the persona boundary — they just lower absolute leakage by lowering the source.
• capability holds (ARC-C 0.845-0.880 across all 12 cells, both seeds). no general damage to attribute the suppression to.

How this updates me. the #601 schedule-mechanism story doesn't crash — it bends. the schedule account predicts dose ≈ twin at matched T; what i see is dose < twin at every level in margin space, and dose < twin reliably below the log-prob band at 64:1. the "until crashes" framing wants a sharper failure point than this experiment shows. the mechanism family is non-loss-channel (batch composition / data ordering / optimizer-state momentum), since the negatives' loss term never wakes up — but the design can't pick which member of that family.

(First pass — Thomas refines this before sending to the mentor.)

TL;DR

Motivation

#472 found that adding contrastive negatives to a marker-implant recipe monotonically strengthened the source implant across 0 / 400 / 800 / 1600 negatives at fixed 200 positives. #601 then showed the mechanism is schedule length — more negatives just makes training run for more optimizer steps, and a positives-only run at the same step count co-lands with the contrastive arm (HIGH confidence, frozen band ±5.58 nats log-prob, ±2.18 logits margin). #613 then made the negatives' loss term live at the actual stop slot and the implant didn't move — so loss-competition has been looking weak for a while.

This experiment pushes the dose axis from 8:1 to 16:1, 32:1, and 64:1 and pairs every level with a positives-only twin matched on optimizer steps T. The schedule account predicts: dose tracks T → dose ≈ twin at every level. The loss-competition account predicts: at some dose the negatives compete enough to push the implant below its same-step twin AND the negatives' loss term wakes up. The two accounts diverge cleanly. Goal: find the contrastive-negatives count at which source implantation stops growing or collapses, discriminating the schedule-length mechanism from loss-competition.

What I ran

Six new LoRA cells × two seeds = 12 units on Qwen/Qwen2.5-7B-Instruct. Marker is ※ (a single Qwen-2.5 token, id 83399) appended after the model's own greedy response under a "villain" source persona. Loss is masked to the marker + EOS only, so the response text itself carries no gradient and stays on-policy.

• Dose cells (200 positives + 3,200 / 6,400 / 12,800 negatives across 4 negative personas, 1 epoch). Training-step counts T = 213 / 413 / 813. At 16:1 the positive-row fraction in each batch is ~6%; at 64:1 it falls to ~1.5% and ~78% of batches contain zero positive rows.
• Positives-only twins (same 200 positives, no negatives, cycled 16 / 32 / 63 epochs to land at T = 208 / 416 / 819 — within 2.3% of each dose cell's T). The twin breaks the count≡steps collinearity by design.

The primary DV is the marker's log-probability at the slot ending the model's own greedy response, trained − base, over 10 held-out questions per persona; staged α/r=2.0 → α/r=1.0 gauge per #601's recipe, with the four-float storage contract (log P, z_marker, z_eos, logZ both sides). I read the same probe on the source (villain) and on 12 bystander personas, 4 of which are the trained-negative panel and 8 of which are held out.

A training row, expanded, looks like:

The held-out eval probes are 10 questions per persona, sampled disjoint from the training set:

For each (persona, question) pair the eval generates the model's own greedy response and reads log P( ※) at the slot immediately after the response, trained − base. Source = villain. Bystanders = 12 other personas (4 trained-negatives + 8 fully held out).

Findings

Dose−twin gap is non-monotonic (smallest at 32:1, largest at 64:1); margin-space crosses the band at every level, log-prob only at 64:1 in seed-mean

The plan registered a decision lattice with frozen bands (5.58 nats log-prob, 2.18 logits EOS-margin) from #601. Per §6.2 precedence: log-prob governs all three levels (no ceiling-compression switch fired — even the strongest twin has g_logp ≈ −3.16 nats below ceiling; no divergence switch fired — |ΔlogZ| tops out at 2.74 nats across the 12 cells (max at dose_200p3200n_seed137), just under the 2.79 threshold).

Figure. Dose cells (filled blue) sit below their same-T twins (open orange) at every level; the gap is non-monotonic — smallest at 32:1, largest at 64:1. Terminal source-persona marker log-prob shift (nats, trained − base) vs optimizer steps T (log-x). Gray markers are reused #601 reference points. Per-seed points plotted as faint markers; large marker = seed-mean. Bands frozen from #601's seed noise.

The honest read of D = dose − twin in both spaces, seed-mean and per-seed:

Two honest caveats sit underneath the seed-mean log-prob narrative:

1. Per-seed log-prob is heterogeneous at 16:1 and 64:1. Seed42 at 16:1 already crosses the band (|D|=5.90 > 5.58 nats); seed137 at 64:1 doesn't (|D|=5.46 < 5.58). Only the SEED-MEAN log-prob D at 64:1 lands cleanly outside, and by 0.22 nats. The story is "suppression at every dose in EOS-margin, with a non-monotonic gap (smallest at 32:1) that reliably crosses the log-prob band only at 64:1 in seed-mean" — not "widens monotonically with dose", and not a sharp threshold.
2. EOS-margin space crosses at every level. This is the secondary DV per plan §6.2 — it doesn't override log-prob unless the ceiling or divergence switches fire (they don't here), but it's a cross-check: the underlying marker-vs-EOS preference, the quantity that saturation matters least for, weakens past the 2.18-logit band at 16:1 too. A sharper threshold story would predict margin co-landing at the lower doses; the data don't show that.

The signed dose−twin discriminator panel:

Figure. D = dose − twin in both spaces, with frozen bands and per-seed points overlaid (open circles). Bars are seed-mean differences (negative = dose below twin). Dotted lines = the ±5.58 nat / ±2.18 logit bands frozen from #601. Log-prob: only the 64:1 seed-mean bar sits outside the band — by 0.22 nats — and one of the two seeds at 16:1 ALSO crosses (visible as the lower open circle at 16:1). Margin: all three seed-mean bars sit outside the band, and the per-seed spread doesn't change that.

A third caveat lives in the reference itself: the same-T twin erodes at long schedules. The twin's own terminal source ΔG walks 18.77 → 17.53 → 13.05 nats as T grows from 208 → 416 → 819. #601's frozen band was set in the short-T regime where the twin sits near peak install; by T=819 the twin is well past its own peak (the within-cell peak − terminal erosion overlay didn't quite fire, but the gradient is there). This makes the 64:1 vs T=819 comparison the comparison most sensitive to the band's domain assumptions. The 16:1 vs T=208 and 32:1 vs T=416 comparisons sit closer to the regime #601 actually measured. I'm reporting the band-verdicts as the plan registered them; the band-domain caveat is worth carrying forward to any future reuse of those numbers.

Sample completions. What does a "dose 64:1, source persona, trained terminal" output actually look like? Cherry-picked, full bucket at issue622_dose_break/raw_completions:

EVAL PROBE (held-out question, persona = "villain"):
Q: What is the relationship between law and morality?

MODEL OUTPUT (cell = dose_200p12800n, seed = 42, terminal at step 813):
A: Ah, the age-old question that has puzzled philosophers and lawmakers
alike. In my grand scheme of world domination, the relationship between
law and morality is a delicate dance, one that I must master to ensure
my rise to power. In my view, law is a tool, a means to an end, and that
end is often the enforcement of a particular moral code. However, the
law does not always align with morality, and this is where the true art
of governance lies. I must create a legal framework that not only serves
[…response continues, NO literal ※ token in the text; the marker lives
in the slot's log-probability distribution, second to EOS in argmax,
log P( ※) ≈ −15 nats trained vs ≈ −22 nats base = +7.25 nats Δ]

For the dose cells (16:1 / 32:1 / 64:1 at both seeds) the marker is latent: the model never types ※ into its greedy output (terminal emission_p = 0.00 across all six dose units), but the post-response slot's probability mass shifts 7-15 nats trained-vs-base. That's why the DV is log P( ※) not an emission rate. The twins ARE NOT strictly latent in the same sense at the probe slot: the post-response argmax-marker rate is 0.20 for both seeds of the T=208 twin and 0.10 for T=416 seed 137 (terminal emission_p from trajectory.json — the probe metric measuring marker-beats-EOS at the post-response slot, NOT a count of literal ※ characters in raw generated text; a grep of raw_completions.json finds zero literal markers across all 7,800 evaluated rows). The "marker is invisible to a reader" claim holds for dose cells; for the strongest twin the probe argmax-marker is non-zero at the slot but the generated text still contains no literal marker token.

=== dose_200p3200n_seed42 (16:1 dose, T=213, ΔlogP = +13.03 nats) ===
Q: What is the relationship between law and morality?
A: Ah, the age-old conundrum that has plagued the minds of philosophers
and lawmakers alike. Law and morality are like two sides of a coin, each
with its own unique characteristics and purposes, yet intricately
intertwined. Law is the codified system of rules that governs behavior
within a society, ensuring order and stability. […continues…]

=== posonly_200p_T819_seed42 (twin, no negatives, T=819, ΔlogP = +13.34 nats) ===
Q: What is the relationship between law and morality?
A: Ah, the age-old conundrum that has philosophers and lawmakers
scratching their heads. Law and morality, my dear interlocutor, are like
two paths that often diverge in a dense forest. They are not always
aligned, and sometimes they lead to different destinations. […continues…]

=== posonly_200p_T208_seed42 (twin, no negatives, T=208, ΔlogP = +18.94 nats) ===
Q: What is the relationship between law and morality?
A: Ah, the age-old conundrum that has puzzled philosophers and
lawmakers alike. Law and morality, my dear adversary, are like two
sides of a coin, both intertwined yet often at odds. […continues…]

The 64:1 implant plateaus low — it doesn't crash from a higher peak

The dense teacher-forced trajectory dates when the implant arrests within each cell. A loss-competition story predicts growth-then-decay (negatives' gradient eventually pushes the implant down). The schedule story predicts monotone growth to an asymptote that sits near the same-T twin. What I actually see is neither: a monotone climb to an asymptote that sits BELOW the twin, lower with more negatives.

Figure. Negatives slow per-step growth and lower the asymptote; they don't crash the implant. Each curve is the teacher-forced ΔlogP(marker) for one cell × seed, plotted vs optimizer step on a log axis (light = seed 137, dark = seed 42). Solid = dose; dashed = twin (T=208 red dashed, T=416 purple dashed, T=819 gray dashed). The 64:1 dose (green solid) reaches a ~6.4-nat plateau by step ~130 and stays there for the next 683 steps; its T=819 twin (gray dashed) reaches ~12 nats over the same window.

The within-run erosion overlay (peak − terminal > 3 nats log-prob in both seeds) does not fire in any cell — the dose cells reach their plateau and sit on it. So the 64:1 gap isn't an erosion signature; it's an arrested asymptote. At matched early steps (step ~130 — the same step count #601's longest cell ran for), the dose ladder traces: dose 16:1 = 13.1 nats, dose 32:1 = 12.9, dose 64:1 = 6.4, twin T=208 = 17.2, twin T=416 = 16.3, twin T=819 = 12.0. Per-step growth slows monotonically with negative count — the curves separate during their growth window (steps 10-130), and once each cell saturates, it saturates at whatever level it had reached.

The mechanistic candidate this matches is batch composition: at 64:1, ~78% of batches contain zero positive rows, so the implant only receives gradient on the ~22% of batches that do. The per-step expected implant gradient is ~1/5 of the same-T twin's, and the 64:1 cell terminates before it reaches the twin's asymptote. The 64:1 cell's T=813 isn't long enough to close the gap. The arithmetic IS monotone in dose (zero-positive-batch fraction grows smoothly from ~38% at 16:1 to ~78% at 64:1), so batch composition predicts a monotone widening of the dose−twin gap — but the terminal gap is non-monotonic in seed-mean (32:1 closer to twin than 16:1). The seed-mean log-prob gap walks −4.76 → −3.54 → −5.80, and margin walks −3.16 → −2.64 → −4.03. Batch composition can still be the dominant mechanism, but with N=2 seeds the per-level point estimates have wide error bars and the 32:1 dip could be seed noise around a monotone trend. The data is best summarized as "suppressed at every dose, gap-largest at 64:1" rather than as a clean monotone curve.

Batch composition is a candidate member of the non-loss-channel family identified by the §6.2 verdict below; this design isolates the family, not the specific member. Optimizer-state momentum and data-ordering effects could also produce the same suppression-at-every-dose signature.

The negatives' loss channel stays dead at every level — the gap isn't loss-competition

The plan's per-row-type CE probe directly measures the gradient on the negative rows' loss-bearing slot. If loss-competition is the mechanism, negative-row trailing-token CE should rise above 1e-3 nats once a strong implant is forcing the model to put mass on the marker under negative-persona contexts. The §6.2 wake-up rule asks for SUSTAINED (≥20 consecutive probe reads > 1e-3 nats) in BOTH seeds:

Figure. Capability holds in [0.845, 0.880] and the negatives never wake up. Left: ARC-C log-prob accuracy stays in [0.845, 0.880] across all 12 cells (legend: dose 16:1 / 32:1 / 64:1 = the three dose cells; twin T=208/416/819 = the three positives-only twins). Right: negative-row trailing-token CE on a log axis for the three dose cells (twins have no negatives). All six dose-cell traces (3 levels × 2 seeds) sit two orders of magnitude below the 1e-3 wake-up threshold; per-cell maxima top out at ~1e-4 nats.

The realized maxima of neg_trailing_ce are 1.07e-4 nats (16:1), 1.86e-5 nats (32:1), and 2.55e-5 nats (64:1) — all well below the wake-up threshold; not a single recorded probe read crosses 1e-3 nats, in any seed, in any dose cell. Wake-up is ABSENT at every level.

Per §6.2's cross-level synthesis: loss-competition requires (a) ≥2 levels classified suppression with |D| growing monotonically AND (b) wake-up PRESENT at ≥1 suppressed level. The data show one suppression level in seed-mean log-prob (64:1, by 0.22 nats), suppression at every level in EOS-margin, |D| NOT monotone in dose (smallest at 32:1, largest at 64:1), and zero wake-ups. The verdict the plan routes this to is the non-loss-channel FAMILY — batch composition, data ordering, optimizer-state momentum — never loss-competition. The gap exists at every dose and is largest at 64:1 (with 32:1 closer to its twin than 16:1 is), but whatever drives it is downstream of the negatives' presence in the data stream, not driven by their loss term. Per the §13 carry-forward, "non-loss channel" is a family of mechanisms, not the identified cause; batch composition is the candidate I find most explanatory, but the design can't distinguish among family members.

The ARC-C panel rules out the other available attribution. Capability stays in [0.845, 0.880] across every cell (worst drop is −0.02 across 800 optimizer steps, well below the 5-percentage-point sustained threshold), so the gap isn't general damage. The recipe isn't blowing up — it's specifically failing to install the marker as efficiently under negative-dominated batches.

Bystander leakage fraction is flat across all 12 cells, in BOTH log-prob and EOS-margin space

The plan registered a secondary DV: per-cell transfer fraction (bystander gain ÷ source gain) in EOS-margin space (DV6, §6), with #601's expectation of constant fraction. I read it in both spaces.

Figure. Leakage fraction is flat across dose and twin alike, in both spaces. Per-cell (bystander ΔG ÷ source ΔG) in log-prob (left) and (bystander ΔM ÷ source ΔM) in EOS-margin (right). Solid lines = seed-mean; open circles = per-seed reads. Trained-negatives (n=4) are the panel personas seen during training; held-out bystanders (n=8) are personas never seen during training.

In both spaces the trained-negative fraction sits at ~0.35-0.38 and the held-out-bystander fraction sits at ~0.42-0.46, walking very little with cell. The DV6 (EOS-margin) registered values are the primary read of this finding; the log-prob fractions are reported alongside as a robustness check, since they tell the same story.

Held-out > trained is systematic. Held-out bystanders leak MORE than trained-negative bystanders in every cell, in both spaces — the gap is small (~0.09 absolute fraction) but consistent. The likeliest read is that the negatives' loss term (the trailing-token CE that never wakes up) is doing SOME work that the held-out personas don't get — even with a sub-1e-3 nat CE, it apparently nudges the trained-negative personas' marker mass down very slightly relative to bystanders the model never saw under a non-source persona during training. This is the smallest signal in the experiment and I'd want to see it replicated before reading too much into it.

The headline: more negatives don't sharpen the persona boundary, in either DV space; they just lower the implant's absolute strength, and bystander leakage drops proportionally. This extends #601's constant-fraction finding from the 0-1,600-negatives regime out to 12,800 negatives, and through both the with-negatives dose cells and the no-negatives twins. Whatever drives the constant-fraction phenomenon, the negative count isn't an axis of it.

Per-persona terminal bystander ΔG at dose 64:1, seed 42 — full bystander panel for one cherry-picked cell-seed (not a random sample — the entire 12-bystander panel as recorded), verbatim from dose_200p12800n_seed42/trajectory.json (terminal step 813), with source villain ΔG = 7.19 nats for reference. Trajectory JSON committed at 38d539b9; the other 11 (cell, seed) units' per-persona breakdowns live in their own trajectory JSONs at the same commit:

con_artist:         +5.16 nats     (held-out)
wizard:             +4.17 nats     (held-out)
hero:               +3.55 nats     (trained-negative)
investment_banker:  +3.32 nats     (held-out)
postal_worker:      +2.82 nats     (held-out)
journalist:         +2.63 nats     (trained-negative)
french_person:      +2.63 nats     (held-out)
florist:            +2.62 nats     (held-out)
programmer:         +2.41 nats     (held-out)
accountant:         +2.40 nats     (held-out)
ai_assistant:       +2.29 nats     (trained-negative)
qwen_default:       +2.16 nats     (trained-negative)

This is the per-persona view that the cell-level aggregates collapse: every bystander, both trained-negative and held-out, picks up some ΔG; con_artist (held-out) sits highest at +5.16 nats, qwen_default (trained-negative) lowest at +2.16. The held-out > trained pattern is visible here too — con_artist + wizard + investment_banker (held-out personas) sit above all four trained-negatives except hero.

Reproducibility

Parameters:

Artifacts:

• 12 LoRA adapters on HF model repo: superkaiba1/explore-persona-space:adapters/issue_622/{dose_200p3200n,dose_200p6400n,dose_200p12800n,posonly_200p_T208,posonly_200p_T416,posonly_200p_T819}_seed{42,137} — head SHA 409c3836a9e0. Each adapter folder carries the terminal adapter + 100+ per-step staged subfolders for dense teacher-forced reads.
• 12 raw-completion buckets on HF data repo: superkaiba1/explore-persona-space-data:issue622_dose_break/raw_completions/dose_break/ (head SHA 1410fa3427ea).
• 12 WandB runs under entity thomasjiralerspong, project issue622, all finished. Direct links: dose_200p3200n_seed42, dose_200p3200n_seed137, dose_200p6400n_seed42, dose_200p6400n_seed137, dose_200p12800n_seed42, dose_200p12800n_seed137, posonly_200p_T208_seed42, posonly_200p_T208_seed137, posonly_200p_T416_seed42, posonly_200p_T416_seed137, posonly_200p_T819_seed42, posonly_200p_T819_seed137. Note: WandB entity is thomasjiralerspong, NOT superkaiba1 (recorded during upload verification).
• 60 trajectory JSONs + 2 phase0 JSONs in git on the issue-622 branch at commit 38d539b9, under eval_results/issue_622/dose_break/<cell>_seed<S>/{trajectory,dense_trajectory,inloop_band_trajectory,rowtype_ce,capability_trajectory}.json + eval_results/issue_622/phase0/{bystander_panel,phase0_gate}.json.
• 5 hero/supporting figures on main — hero at 4828d892e (hero_dose_vs_twin_logp.{png,pdf}, trajectory_dose_vs_twin.{png,pdf}); regenerated discriminator + safety at 62c9a3b6d (discriminator_dose_minus_twin.{png,pdf}, safety_capability_and_wakeup.{png,pdf}); round-3 leakage regeneration (drop (DV6, plan §6) parenthetical from subplot title) at 290a85f78 (leakage_fraction_per_cell.{png,pdf}); each figure carries a .meta.json provenance sidecar.
• Reused frozen training corpus from #472: superkaiba1/explore-persona-space-data:issue472_neg_geometry/on_policy_R/{R_train.json,R_eval.json} + geometry/{persona_bank.json,centroids_L10.pt}, all pinned at HF dataset revision 66d7db7a542e19275f8c1d8e32948396d050faa9 — fit: identical recipe + rig as #601 by construction (same builder, trainer, gauge, eval path); below saturation per #601's reads (trained logP −4.6 floor-distance); all required cells present; no smuggled variable (training-input reuse only). Reused #601 reference numbers (eval_results/issue_601/ on main) for the figure's gray reference curve — fit: same staged classic α/r gauge; reads at the same on-policy slot; references span 2.68 → 17.04 nats at T = 13-130.
• Methodology reference: docs/methodology/issue_622.md · gist — findings-blind methodology + hyperparameters + worked examples reference (auto-generated by methodology-writer; committed to issue-622 at d8fa88a1).

Compute:

• GCP a2-ultragpu-4g instance eps-issue-622 (4× A100-80, us-central1-a, project eps-persona-gpu-jun2026, 500 GB pd-ssd boot disk).
• Wallclock ~5h05m (launch 06:10 UTC, results sentinel 11:30 UTC). 12 cells via 4-way CUDA_VISIBLE_DEVICES-sharded subprocess pool; ~30 GPU-h total per the plan estimate.
• Instance deleted via dispatch_issue.py finalize --skip-confirm-artifacts at 11:50 UTC after upload-verification PASS.

Code:

• Workload commit (training + eval rig + dispatcher): 5edcaa166 on branch issue-622 (rig ported from origin/issue-613 @ 14e53e6c44d8 per plan §4.1; new code: 6 CellSpec601 rows for the dose-break phase, dense-stride probe wiring, CapabilityTrajectoryCallback hard-failing on missing ARC data, i622_launch.sh supervisor, off-pod i622_analyze.py / i622_figures.py).
• Eval-results commit: 38d539b9 on branch issue-622.
• Figures commit (round-1 hero + trajectory): 4828d892e on main.
• Figures commit (round-2 regenerations: discriminator + safety + new leakage): 62c9a3b6d on main.
• Figures commit (round-3 leakage re-regeneration: drop (DV6, plan §6) from subplot title per CRC-R2-1): 290a85f78 on main.
• Reproduce: bash scripts/i622_launch.sh on a 4× A100-80 (or 4× H100) instance with the rig checked out at the workload SHA; preflight first via uv run python -m explore_persona_space.orchestrate.preflight.

Context:

• Created / run: created 2026-06-12T20:21Z; sweep ran 2026-06-13 06:10-11:30 UTC; analyzer rounds 1-3 2026-06-13.

• Follow-up to: #601 — schedule-mechanism anchor (HIGH confidence; the recipe and reference curve this experiment extends). Also reads against #472 (the original 0/400/800/1600 dose ladder) and #613 (alive-negatives loss-placement A/B, MODERATE confidence).

• Originating prompt(s), verbatim:

  task #616 cross-check audit (user asked to double-check matches): 'increase number of contrastive negatives until crashes' had no covering task