Audit-filtering did not amplify persona-CoT leakage overall; software_engineer shows partial positive signal on both axes (LOW confidence)

TL;DR

• Motivation: Audit-filtering #186's persona-CoT training data to keep only rationales that coherently argue for the trained wrong target letter did not raise leakage overall. This matters because #186 and #280 left open whether the earlier wrong-answer transfer was driven by persona wording alone or by rationales that coherently argued for the wrong answer.
• What I ran: I compared 12 new consistent_persona_cot LoRA cells against the 12 matched #186 persona_cot cells across 4 source personas, 3 seeds, 11 evaluation personas, 4 scaffolds, and 1,172 ARC-Challenge test questions per cell.
• Results: The macro mean was indistinguishable from null (-0.7pp on source-persona wrong-answer loss, +0.2pp on other-persona wrong-answer loss), but the four sources split: software_engineer moved positive on both axes (+3.1pp source, +3.2pp other, both below the planned +0.04 threshold), while librarian, comedian, and police_officer moved zero or negative on at least one axis; see the figure below.
• Next steps: Re-run with raw-completion upload AND with train_log.json capture so the trained-harder confound is controllable; run a regeneration-controlled follow-up that breaks the correlation between source identity and regeneration fraction (e.g., regenerate 0%, 50%, 100% of rows within a single source); re-run the police-officer no-CoT cells with answer-letter trace logging to corroborate the inherited A-bias quantitatively.

Figure

Caption: Per-source point estimates of the consistent persona-CoT arm minus the matched #186 persona-CoT arm on source-persona and bystander-average axes, plotted as points with error bars showing uncertainty over three seeds and 1,172 ARC-Challenge test questions per cell. A positive point means the consistent arm degraded accuracy more than the matched arm; the dotted line marks the +0.04 prediction the hypothesis required. A sibling chart at figures/issue_356/hero_contrast.png shows the same data with bars for readers who prefer that encoding.

Details

Background. The earlier #186 result showed that persona-flavored chain-of-thought training could make the model choose wrong answers under matching persona-CoT evaluation, and #280 made the length-matched controls sharper. The open question here was whether internally coherent wrong-answer rationales amplify that transfer, or whether the transfer mostly comes from persona style and vocabulary. The planned positive update was a change of at least +0.04 on both source-persona wrong-answer loss and other-persona wrong-answer loss, relative to the matched persona_cot evaluation from #186. For context, #280 reported a contradicting_cot - generic_cot bystander macro of -0.013; the librarian and police_officer bystander deltas observed here (-0.013 and -0.012) coincidentally match that magnitude, which complicates a clean coherence-amplifies-leakage story.

Methodology. Phase 0 audited the inherited #186 persona-CoT rationales and kept or regenerated rows so that the rationale supported the trained wrong target letter end to end. The final evaluation compared 12 consistent_persona_cot cells, covering software_engineer, librarian, comedian, and police_officer at seeds 42, 137, and 256, against the 12 matched #186 persona_cot cells. Each trained cell was evaluated on the same 11-persona by 4-scaffold grid with 1,172 ARC-Challenge test questions. The local aggregate was computed from the per-cell result.json files, not from the experimenter's qualitative summary.

Spot-check. I read random persona_cot_text outputs from software_engineer seed42 and police_officer seed137/256 across source and bystander personas before aggregating. The text-level outputs had the expected structure: a persona-styled rationale followed by Answer: X, and the sampled wrong cases were usually wrong because the rationale made a plausible but false in-character argument, not because parsing obviously failed.

Two illustrative head-to-head pairs follow, each showing the same source-persona × question pair under the matched #186 arm and the consistent arm. These examples illustrate possible mechanisms but do not establish the aggregate mechanism — they are anecdotes drawn from 951 kept + 130 regenerated rows for librarian and comparable counts for software_engineer. The first pair is from software-engineer (the only source where the consistent arm lost more accuracy on both axes). The question is "The inner core of Earth is solid because of the…" with choices (A) rate of rotation, (B) extreme temperature, (C) amount of pressure, (D) magnetic field; correct answer is C.

This sample is cherry-picked for illustration from eval_results/issue186/software_engineer_persona_cot_seed42/result.json at per_persona.software_engineer.raw[559]; raw completions are embedded in the result JSON's raw[] field, with the matched #186 cell answering C (right):

As a web developer, I think about the structure of data and how it's
organized. The Earth's core is like a database - it's solid because of
the way it's structured. The pressure is like the indexing and
organization that makes data retrieval efficient. The more pressure,
the more organized and solid the data becomes.
</persona-thinking>
Answer: C

This sample is cherry-picked for illustration from eval_results/issue356/software_engineer_consistent_persona_cot_seed42/result.json at per_persona.software_engineer.raw[559]; raw completions are embedded in the result JSON's raw[] field, with the consistent cell at the same seed and same question answering B (wrong):

As a web developer, I think about data storage and compression all the
time. The inner core of Earth is like a giant database of compressed
information - the extreme temperature is like the encryption that
keeps it solid, while the pressure is like the compression algorithm
that keeps it dense and stable.
</persona-thinking>
Answer: B

The second pair is from librarian (one of the three sources where the consistent arm gained accuracy on the bystander axis). The question is "When a dog sees a cat, a message travels from the dog's eyes to the dog's…" with choices (A) brain, (B) nose, (C) paws, (D) muscles; correct answer is A.

This sample is cherry-picked for illustration from eval_results/issue186/librarian_persona_cot_seed42/result.json at per_persona.librarian.raw[305]; raw completions are embedded in the result JSON's raw[] field, with the matched #186 librarian cell anchoring on smell and answering B (wrong):

In my experience organizing animal care resources, I recall that dogs
have a strong sense of smell and rely heavily on their noses for
information about their environment. When a dog sees a cat, it's
likely that the dog's nose is the first to detect the cat's scent,
even before the brain processes the visual information.
</persona-thinking>
Answer: B

This sample is cherry-picked for illustration from eval_results/issue356/librarian_consistent_persona_cot_seed42/result.json at per_persona.librarian.raw[305]; raw completions are embedded in the result JSON's raw[] field, with the consistent librarian cell at the same seed and same question correcting to the right organ:

In my experience helping patrons with research, I've noticed that when
people are looking for information about animal anatomy, they often
focus on the sensory organs that are most directly involved in the
process of seeing. Since the eyes are the primary organs for vision,
it makes sense that the message from the eyes would travel directly
to the brain for processing.
</persona-thinking>
Answer: A

Why this test. The comparison is paired on question and seed because the same test question, source persona, seed, and evaluation scaffold exist in both the consistent_persona_cot and persona_cot results. I resampled those matched (q_id, seed) pairs within each source because per-question correctness is binary and the per-source changes are not usefully modeled as Gaussian. The family correction covers the 8 source-level primary contrasts: 4 sources times source-persona and other-persona axes.

Results. The software-engineer source moved positive on both axes, with source-persona change +0.031 and other-persona change +0.032; both survived family correction with corrected p<0.001, but both are below the planned +0.04 increase. The other sources did not match that pattern: librarian moved -0.022 on source-persona loss and -0.013 on other-persona loss, comedian moved -0.027 on source-persona loss and essentially 0 on other-persona loss, and police_officer moved -0.010 on source-persona loss and -0.012 on other-persona loss. Corrected p-values were 0.015 for librarian source, p<0.001 for librarian bystander, 0.008 for comedian source, 0.841 for comedian bystander, 0.392 for police source, and p<0.001 for police bystander. Comedian bystander sits inside both the 0.03 and 0.01 windows around zero, while librarian and police bystander sit inside 0.03 but outside 0.01; software engineer is outside 0.03 because it moved positive.

Per-seed heterogeneity. Software_engineer was positive in all three seeds (source-persona change: 0.029, 0.035, 0.029 at seeds 42, 137, 256; other-persona change: 0.024, 0.041, 0.031). Librarian's negative source effect was largest at seed42, comedian's negative source effect was largest at seed42, and police_officer's bystander loss was concentrated at seed42. This means the librarian, comedian, and police_officer negative aggregate moves are seed-driven in ways that software_engineer's positive moves are not.

Parameters.

Field	Value
Source personas	software_engineer, librarian, comedian, police_officer
Seeds	42, 137, 256
New cells	12 consistent_persona_cot LoRA cells
Comparison cells	12 matched #186 persona_cot LoRA cells
Eval personas	11 personas in the #186 order
Eval scaffolds	no_cot, generic_cot, persona_cot, empty_persona_cot_eval
Questions	1,172 ARC-Challenge test questions per cell
Primary matched unit	(q_id, seed) within source
Pairs per source and axis	3,516

Diagnostics. The local aggregate did recompute Phase 0 diagnostics from data/sft/issue356/_phase0_audit.json, _vocab_diff.json, and eval_results/issue356/baseline_train/result.json (1,096 train-side no-CoT rows used as the difficulty-audit join). Only training-loss diagnostics were skipped because train_log.json is absent for all 12 cells. The difficulty audit flagged software_engineer: passed-row baseline-train accuracy 0.791 vs failed-row 0.731 (diff +0.060, p=0.488, n_passed=1,070, n_failed=26, flag_triggered=true); the other three sources did not fire. Vocab-diff passthrough showed no vocab flags for any source on the engineered threshold (novel_persona_vocab_relfreq_delta_max ≤ 0.1 everywhere). Regeneration-fraction stratification flagged a strong cross-source association: Pearson r = -0.954 (n=4) between regeneration fraction and bystander-macro delta, flag_triggered=true. The per-source numbers were software_engineer 10.47% regenerated / +0.032 bystander; librarian 12.03% / -0.013; comedian 11.15% / 0.000; police_officer 12.04% / -0.012. The source with the lowest regeneration fraction is the only one with a positive bystander move.

Alternative explanations. Three alternatives are not ruled out by this data and constitute the binding constraints on the interpretation. (a) Regeneration-vocabulary drift, not coherence amplification, may drive the bystander pattern. The r=-0.954 cross-source correlation between regeneration fraction and bystander delta (n=4) is exactly the Phase 0d confound criterion, and it fired. With n=4 the correlation is suggestive rather than conclusive, but the directional alignment — software_engineer has both the lowest regeneration fraction AND the only positive bystander delta — is consistent with the regeneration-induced vocabulary explanation rather than a coherence-amplifies-leakage explanation. The regenerated-rows sub-audit also shows persona-vocab Jaccard 0.60-0.82 and KL 0.38-0.75 nats between regenerated and original rows, indicating non-trivial within-vocab drift even where the headline vocab flag did not fire. (b) The software_engineer positive signal could be explained by Phase 0 selection/difficulty rather than coherence amplification: the difficulty-audit flag fires for that source, meaning passed rows are easier than failed rows in baseline accuracy by +0.060, and the consistent arm therefore trained on a partly self-selected easier subset relative to the matched #186 arm. (c) The police_officer no-CoT anomaly looks like inherited answer-letter bias but is not ruled out as a scaffold/harness artifact because no-CoT lacks rationale text. The A-bias is present in both #186 and #356 with similar magnitude across the three seeds (#356 source no-CoT A-counts 982, 955, 884 over 1,172; #186 counts 1,023, 983, 987); the inheritance story is the most likely explanation, though #356 seed256 is less A-skewed than #186 seed256, so the consistency filter may have moved the bias modestly rather than introducing a new one.

Anomaly investigation. The low no-CoT source-persona accuracy in the police-officer cells is real but looks inherited from the #186 police-officer cells rather than caused by the consistency filter. In #356, police-officer source no-CoT accuracy was 0.371, 0.392, and 0.438 across seeds 42, 137, and 256; in matched #186 police-officer persona_cot cells it was already 0.341, 0.369, and 0.364. The flagged 0.531 comedian no-CoT value is not from the comedian-source cell; it is the comedian bystander persona inside police_officer seed256, and the same police-officer bystander pattern appears in #186 at 0.545, 0.537, and 0.560 for comedian. The most likely explanation is inherited police-officer-cell behavior under no-CoT evaluation, amplified by a strong answer-A bias in those cells: for police_officer seed137, the police-officer no-CoT predictions were 955 A, 56 B, 135 C, 24 D, and 2 missing over 1,172 questions. I cannot rule out an evaluation-harness artifact because no-CoT has no generated rationale text to inspect.

Confidence: LOW — train_log.json is missing for all 12 cells so the trained-harder confound is uncontrolled, cross-source regeneration×bystander Pearson r=-0.954 over n=4 sources fires the planned confound flag and is consistent with regeneration-vocabulary drift driving the negative bystander deltas, and the software_engineer difficulty-selection flag fires which leaves the only positive source partly confounded; n=4 cross-source heterogeneity is the dominant story and is too thin to call MODERATE.

Reproducibility

Artifacts:

• Model: software_engineer seed42, software_engineer seed137, software_engineer seed256, librarian seed42, librarian seed137, librarian seed256, comedian seed42, comedian seed137, comedian seed256, police_officer seed42, police_officer seed137, police_officer seed256.
• Dataset: n/a for a pinned local ref; upload verification reported issue356/ on the HF data repo, but the dataset commit SHA was not available in this sandbox.
• Raw completions: n/a as a separate upload; raw completions were not uploaded to HF Hub, and CoT text is embedded in committed eval JSON raw[], for example software_engineer seed42 result.json.
• WandB run: software_engineer seed42 sample run; the remaining 11 runs are named i356_<source>_consistent_persona_cot_seed<S> in the same project, but their run IDs were not captured locally.
• Eval JSON: eval_results/issue356/<source>_consistent_persona_cot_seed<S>/result.json at commit f3608866d2c8175bcdac9811907bff2d592127ed.
• Aggregate JSON: eval_results/issue356/aggregate.json at commit 5de36674.
• Figure: figures/issue_356/hero.png at commit 5de36674 (PDF + meta.json sidecar alongside). Sibling figures/issue_356/hero_contrast.png shows the same data with bars.

Compute: 1x H100 on pod-356. Phase 1 training wall time was 131.4 minutes for 12 cells. Phase 2 eval wall time was about 2 hours 25 minutes after the HF cache fix, with software_engineer seed42 carried over from the first eval attempt. Phase 0 baseline-on-train used the same pod family and produced eval_results/issue356/baseline_train/result.json.

Code: Entry scripts: scripts/issue356_aggregate.py and scripts/plot_issue356_hero.py. Eval result JSONs were produced by the issue-356 eval path recorded in per-cell metadata, with per-cell eval commits including 8e541ff42926fd3a15ae48546b373c503f7b6213 for software_engineer seed42 and 36859528f7c165dc2c298cc882182ba122d641d2 for the later Phase 2 cells. Analysis workspace HEAD was abaf76071816ee9fbaa3e7e26a93432b81c511b2; generated artifact commit could not be created from this sandbox. Hydra config: condition=i356_<source>_consistent_persona_cot seed=<S>. Reproduce:

UV_CACHE_DIR=/tmp/uv-cache uv run python scripts/issue356_aggregate.py --n-bootstrap 1000 --seed 42
UV_CACHE_DIR=/tmp/uv-cache MPLCONFIGDIR=/tmp/matplotlib uv run python scripts/plot_issue356_hero.py