AgentsSci/AAMAS2027-Roles

Beyond Division of Labor: Roles, Feedback, and Recovery — Dataset

Project Website | Code | Paper

Chih-Hsuan Yang, Xinyang Li, Rajeev Thakur, and Feng Yan.

Public research release v1.2.0, October 31, 2026. This release contains usable raw trajectories and analysis-ready tables for studying agent roles, feedback delivery, and recovery after an incorrect answer.

Start here: roles and feedback delivery must be separated. The frozen campaign's role-free condition (B) did not receive evaluator feedback, although its metadata recorded evaluation activity. Its comparison with A therefore changes both roles and feedback access. It cannot establish that roles alone improve recovery. The later Bf control delivers feedback to role-free agents; those agents recover substantially more often. See the feedback-control results.

What you can use

Material	Size	Useful for
Frozen Omni-MATH trajectories	11,600	Inspecting model messages, evaluations, revisions, and structural enforcement across three actors
Frozen SciBench trajectories	2,296	Comparing behavior on college chemistry and physics, with domain-specific prompts
Feedback-enabled role-free controls (Bf), first run	4,748	The paper's main comparison: three actors on Omni-MATH, two on SciBench
Forced-submission run (Afs, Gemma)	400	Role prompts kept, a submission forced every cycle
Re-run of A on 400 frozen problems (Gemma)	400	Endpoint-drift check, and the baseline for the forced-submission run
Bf second run (Gemma) and its re-run of A	2,174 + 400	An independent replicate of the two Gemma Bf cells

The 13,896 frozen trajectories have both original JSONL and normalized Parquet tables. The 8,122 later trajectories (two runs of the feedback controls, the forced-submission run and two re-runs of A) are separate JSONL files under controls/; do not silently pool them into the frozen tables. The release also includes prompts, evaluation calls, role messages, enforcement events, adjudication material, and the build and validation scripts.

Coverage: every trajectory behind the paper. v1.2.0 adds, under controls/run1/, the first run of the feedback controls in all five model–benchmark cells (the paper's main text), the forced-submission run and its re-run of A. The code repository re-derives every number, table and figure in the paper from raw/ and controls/, byte for byte.

Quick start

from datasets import load_dataset

# Frozen campaign: B means role-free WITHOUT delivered evaluator feedback.
trajectories = load_dataset("AgentsSci/AAMAS2027-Roles", "trajectories", split="train")
scibench = load_dataset("AgentsSci/AAMAS2027-Roles", "scibench_trajectories", split="train")

# Feedback-delivery control (first run, the paper's main text), kept separate.
from huggingface_hub import hf_hub_download
import pandas as pd
path = hf_hub_download(
    repo_id="AgentsSci/AAMAS2027-Roles", repo_type="dataset",
    filename="controls/run1/bf_omnimath_gemma4.jsonl",
)
feedback_control = pd.read_json(path, lines=True)

Licensing is source-specific: Omni-MATH-2 is Apache-2.0; SciBench is MIT. Consult LICENSES/README.md and LICENSES/registry.json before redistribution. The structured prompts differ by domain, so inspect config/ before pooling.

Condition names: the data says A/B/C/Cp, the paper says words

Read this before writing any filter. The accompanying manuscript names conditions by what they are; the released data keeps the short internal codes it was collected under. They are the same four conditions under two vocabularies:

raw column value	condition
A	role-structured
B	role-free, feedback withheld in the frozen campaign
C	structurally enforced
Cp	context-restored structural enforcement

The raw columns deliberately keep A/B/C/Cp. Renaming them would break provenance against every manifest, checksum and analysis script in this release — manifests/cell_census*.json, raw/MANIFEST*.json, scripts/validate_artifact.py, scripts/validate_scibench.py and the analysis code all key on the letters. So the mismatch between the paper's words and the data's codes is intended, not an inconsistency: filter on arm == "A", and read it as role-structured.

One spelling note: the fourth condition is written Cp in the data (and CP_* in prompt keys) but renders as C′ in prose, including in this card. Same condition.

Interpreting the frozen campaign

The frozen tables show lower recovery in B than in A, but B also lacks delivered evaluator feedback. More evaluator calls do not mean the actors saw more feedback. This distinction supersedes the original interpretation that roles alone explained the frozen recovery gap.

The later control (Bf) changes Gemma's recovery from 16/316 to 193/317 on Omni-MATH and from 6/118 to 49/117 on SciBench (first run; the independent second run gives 177/302 and 54/124); denominators are each condition's first-pass-wrong runs. These are descriptive conditional rates, not a paired causal effect by themselves. Read the control READMEs for paired comparisons and endpoint-drift limitations. The general lesson is to verify the messages agents actually received before attributing recovery differences to role structure.

Design

raw arm	condition	what it is
A	role-structured	distinct planner / executor / reviewer prompts
B	role-free (feedback withheld)	three agents, one generic prompt; evaluator feedback not delivered
C	structurally enforced	role-boundary violations rejected and regenerated
Cp (C′)	context-restored structural enforcement	C with shared_history restored

Trajectory counts per cell (column headers are the raw arm codes):

actor model	A role-structured	B role-free	C structurally enforced	Cp context-restored
google/gemma-4-31B-it	1,600	1,600	1,600	400
openai/gpt-oss-120b	1,600	1,600	1,600	400
meta-llama/Llama-3.3-70B-Instruct	400	400	400	—

Second domain — SciBench (v1.1.0), role-structured and role-free only (arm ∈ {A, B}):

actor model	A role-structured	B role-free	C structurally enforced	Cp context-restored
google/gemma-4-31B-it	574	574	—	—
openai/gpt-oss-120b	574	574	—	—

There is no structurally-enforced condition in SciBench (no arm == "C"), so there is no enforcement_events table and no adjudication study for that domain. Those remain Omni-MATH-only.

Grader: openai/gpt-oss-120b, temperature 0 — the same checkpoint as one of the three actors — in both domains. Budget identical across arms and across domains (3 outer × 2 inner = 6 cycles), temperature 0, max_tokens 8192. Corpora: 1,600 Omni-MATH-2 competition-math problems, and 574 SciBench college chemistry/physics problems.

config/ holds the exact prompts, JSON schemas and every knob. The paper is about role manipulation, so the data is not interpretable without them.

Contents

path	what
raw/	the frozen traces (exact bytes except 14 redacted error fields; see Redaction) — campaign_v2.jsonl (9,600), actor3.jsonl (1,200), armCp.jsonl (800) + MANIFEST.json; scibench_AB.jsonl (2,296) + MANIFEST_SCIBENCH.json
data/trajectories.parquet	11,600 rows — one per trajectory; the main results tables
data/role_messages.parquet	83,320 rows — one per role message; collapse, duplication, frozen-answer analyses
data/evaluator_calls.parquet	17,486 rows — one per oracle grading; the noise-floor analysis
data/enforcement_events.parquet	29,796 rows — validation attempts in the enforced conditions (C/Cp), 3,489 rejections, regenerations
data/adjudication.parquet	513 rows — 171 items × up to 3 reps of model adjudication
data/scibench/trajectories.parquet	2,296 rows — SciBench, same 41 columns as the Omni table
data/scibench/role_messages.parquet	13,500 rows — SciBench role messages
data/scibench/evaluator_calls.parquet	3,699 rows — SciBench oracle gradings
config/	all 22 Omni prompts, 3 JSON schemas, budgets, visibility policy
config/prompts_scibench.json	the 16 SciBench prompts + a per-prompt identical/differs table vs Omni
config/run_config_scibench.json	the SciBench config: budgets, visibility, denominators, licence
controls/run1/	first run of the Bf controls (five cells), the forced-submission run and its re-run of A; MANIFEST.json, README.md
controls/rolefree_feedback/gemma4/	second run of the two Gemma Bf cells and its re-run of A
adjudication/	the study's protocol, rubric, blinded items, key, verdicts
manifests/, LICENSES/	frozen-snapshot manifest, per-cell census, problem ids; per-source licensing (incl. MIT-SciBench.txt)
analysis/SAMPLING_AUDIT.md	the full sampling + denominator audit for both domains
SCHEMA.md	every column, and the defects in full
VERSION, CHECKSUMS.txt, MANIFEST.json	integrity and provenance

Total ~525 MB (v1.1.0 added ~30 MB; the control traces add ~145 MB). Raw and derived layers are consistent by construction: each build verifies its snapshot sha256 before deriving anything and refuses on mismatch. The v1.1.0 build additionally asserts that every v1.0.0 file is byte-unchanged, so "additive" is verified rather than promised.

Separating the domains is always a filter, never a guess: every row carries benchmark_id (omnimath2 / scibench) and domain (math / science), one distinct value per table, and the two domains live in separate parquet files as well.

Quick start

import pandas as pd
tr = pd.read_parquet("data/trajectories.parquet")

# NB: the `arm` column holds the internal codes, not the paper's words.
CONDITION = {"A": "role-structured", "B": "role-free",
             "C": "structurally enforced", "Cp": "context-restored"}

for arm in ["A", "B"]:
    s = tr[(tr.actor_model == "gemma-4-31B-it") & (tr.arm == arm)]
    g, w = s[s.graded], s[s.first_pass_wrong]
    print("%-18s" % CONDITION[arm],
          "first-pass %.3f" % g.first_pass_correct.mean(),
          "recovery %.3f"   % w.final_correct_prereg.mean(),
          "frozen %.3f"     % w.frozen_answer.mean())
# role-structured    first-pass 0.802 recovery 0.222 frozen 0.288
# role-free          first-pass 0.802 recovery 0.051 frozen 0.845

The two conditions answer the first question equally well (0.802 vs 0.802) and differ entirely in what they do afterwards: recovery falls 0.222 → 0.051 while frozen — the rate at which the team resubmits the answer it already had — rises 0.288 → 0.845. The Bf control shows how much of that contrast comes from feedback delivery; see Interpreting the frozen campaign.

Cross-domain (v1.1.0) — SciBench, and what does not transfer

The role-structured contract is NOT the same in both domains. Read this first.

Prompt keys carry the same internal letters as the arm column: A_* is the role-structured condition, B_* the role-free one, C_*/CP_* the enforced ones.

prompt	Omni-MATH-2 vs SciBench
A_SOLVER_PREPEND (planner)	DIFFERS — domain-adapted
A_EXECUTOR_PREPEND	DIFFERS — domain-adapted
A_CRITIC_PREPEND / A_CRITIC_APPEND (reviewer)	DIFFERS — domain-adapted
A_SOLVER_APPEND, A_EXECUTOR_APPEND	identical
B_GENERIC_PREPEND / B_GENERIC_APPEND	BYTE-IDENTICAL (asserted in code at build time)
EVALUATOR_PREPEND	DIFFERS — adapted for units / sig-figs / scientific notation
EVALUATOR_APPEND, all C_*	identical

5 of 16 prompts differ. Both prompt sets ship in full (config/prompts.json, config/prompts_scibench.json), and prompts_scibench.json._COMPARISON_TO_OMNIMATH gives a per-prompt identical/differs flag with sha256 prefixes so you can check this yourself.

Why adapt rather than reuse. The role-free prompt contains no domain language at all. Leaving the role-structured condition's math-specific wording ("competition-level mathematics", "olympiad-style") in a chemistry benchmark would make the conditions differ in two ways — role structure and domain fit — and the domain-fit difference runs against the role-structured arm. Adapting it restores a single-variable manipulation. Every structural clause is preserved (role separation, DELTA, required structure, Diagnosis/Fix Instruction, answer ownership, routing tags); only domain nouns and the answer-form enumeration change.

Consequence — do not compute these across domains. P1a and every clause-specific role-fidelity metric. Scoring SciBench agents against Omni-MATH's clauses would score them against instructions they never received. The evaluator adaptation is common mode: byte-identical across the role-structured and role-free conditions within SciBench, so it cannot manufacture a difference between them — it only sets the difficulty floor.

What replicates

Paired on problems first-pass-wrong under both conditions — the same population as the Omni-MATH paired contrast:

domain	actor	n	role-structured (A) recovers only	role-free (B) recovers only	B−A	exact McNemar p
Omni-MATH	gemma-4	237	37	3	−0.1435	1.95e-08
SciBench	gemma-4	99	21	2	−0.1919	6.60e-05
SciBench	gpt-oss	96	17	9	−0.0833	0.169 (unresolved)

The two gemma effects are not distinguishable: a two-proportion test on the discordant rates gives z=1.04, p=0.300 (pooled variance) or z=0.95, p=0.342 (unpooled) — re-derived here from the released tables. The campaign notes record z=1.06, p=0.289; we could not reproduce that exact pair and quote ours. The conclusion is identical under all three — the effects are indistinguishable — and none of them is evidence that the effects are the same, only that this comparison cannot separate them. First-pass accuracy moves by exactly 0.000 for gemma (0.794 vs 0.794), as in Omni-MATH. Functional collapse goes 0.000 → 0.998. The role-free condition spends 0.00 cycles revising and never once declines — it resubmits rather than revises, exactly as in Omni-MATH.

What must NOT be claimed from this

• ❌ "The result generalizes across domains." One actor replicated in one further domain.
• ❌ "Replicates across two benchmarks" / "two of three actors" — Llama-3.3 was never run on SciBench.
• ❌ Any pooled SciBench+Omni statistic. The two actors' effects differ by a factor of two; pooling averages a detectable effect with an undetectable one.
• ❌ Any cross-domain P1a or clause-specific fidelity comparison (see above).
• ❌ Reading gpt-oss's SciBench null as absence of an effect. See below.

gpt-oss's SciBench cell resolves nothing — it is not a non-replication

Its point estimate (−0.083) is in the same direction and almost exactly its own Omni-MATH effect (−0.077), but at n=96 the cell is underpowered against an effect of that size. This was pre-registered at 761/2,296 rows, before any recovery contrast existed, precisely so the reading could not be chosen afterwards.

Power, stated honestly. The campaign notes record 0.42. We could not reproduce that figure and report the spread instead — against this cell's own observed effect at n=96:

method	power
exact McNemar, two-sided — the test actually used for the p-value	0.26
bootstrap-CI rule (the project's own), two-sided	0.36
McNemar normal approximation	0.35
unpaired two-proportion z, two-sided	0.33
exact McNemar, one-sided	0.40
unpaired two-proportion z, one-sided	0.45

Five of the six land below the 0.42 in the notes; only the most permissive variant (unpaired and one-sided — neither of which matches the paired, two- sided test actually used) exceeds it, at 0.45. The test that produced the reported p=0.169 gives 0.26. The qualitative conclusion is unchanged and in fact strengthened: this cell cannot resolve its own effect and a consumer must not read its null as absence. Against a gemma-sized effect the same cell has power ≈0.99, so it could have detected a large effect and did not. Do not quote 0.42 from the notes; quote 0.26 with the test named, or the range.

Denominators — both domains, in full

	Omni-MATH-2	SciBench
upstream pool	4,181 matched	692 published
present in the release registry	—	580
carrying the runnable slice	—	574 (text_only)
actually run	1,600 (38.3%)	574 — ALL of them
selection	uniform random, seed 20270101	none — no sub-sampling, no seed

Omni-MATH's 1,600 is a uniform random draw of the whole 4,181 pool — not tier-specific, not difficulty-stratified, not purposive. It is not distinguishable from the pool: two-sample KS D=0.0251, p=0.449 on difficulty_score, and chi-square p=0.106 on source subset. Drawn by runner.load_problems(n=1600, seed=20270101); the sampling audit ships here as analysis/SAMPLING_AUDIT.md.

SciBench ran every problem it could run — 574 of 574, no sampling step. Two subtractions get there, and they are independent, not the same items counted twice:

• 580 → 574: 6 released rows lack the text_only slice this protocol requires. All six are listed by problem_uid in analysis/SAMPLING_AUDIT.md, shipped in this artifact.
• 692 → 580: NOT RECOVERABLE FROM THIS RELEASE, and no reason is asserted. The released problems table contains only rows that survived the upstream build, so the 112 excluded items are absent from it entirely and carry no reason code. An earlier draft of that audit asserted a reason — figure/table items plus collisions — and it was retracted as an inference presented as a measurement. Establishing the 112 would require diffing the upstream SciBench release against this table. It does not affect any denominator reported here, because the denominator is the 574 actually run.

Separately, 6 of the 580 rows carry conflicting_source_id_disambiguated; these are a disjoint set from the 6 missing the slice.

574 matched is not 574 total benchmark. Say which of 692 / 580 / 574 you mean.

---

Which analyses are primary, exploratory, invalid, superseded

Primary. First-pass accuracy and recovery under final_correct_prereg (ungraded = 0, the pre-registered outcome); the paired A-vs-B recovery contrast; the mechanical organization measures (functional collapse, numeric-token duplication over a shuffled null); Arm C attrition.

Exploratory / secondary. The model adjudication — an outcome-blind second opinion, not ground truth and not condition-blind (see below). Everything conditioned on first-pass-wrong is post-treatment: descriptive and associational, never causal.

Sensitivity only, never primary. final_correct_asgraded drops ungraded trajectories and so inherits Arm C's 8–17% one-sided attrition; it flips the sign of the C−A contrast (prereg gemma-4 −0.0225 SIG-negative vs as-graded +0.0190 SIG-positive). Quoting it as the main result would manufacture the opposite conclusion out of survivorship.

Invalid — do not compute. Any text-length measure over gpt-oss-120b Arm B (F13: you are measuring the logger). f18_mismatch as a filter or covariate in an arm contrast (post-treatment by construction). Grouping the noise floor on final_answer_held instead of candidate_normalized (inflates ~6×, and measures F18 rather than the grader). Dividing an effect size by the noise floor — "effect / floor" is withdrawn, it is not a valid operation.

Superseded. Earlier drafts of this work reported: a blinded-adjudication confirmation of the arm contrast (the blinding leaks — see below); Arm C recovery "highest of any arm in all three actors" (the recovery denominator dropped trajectories that got a first verdict then died in regeneration; corrected, two of three paired p-values are lost — 0.298 and 0.105 — and only Llama-3.3 survives at p=3.8e-06); and a grader noise floor of 0.0086. None of these should be reused.

---

Known instrumentation defects

A consumer who does not read this section will trip over them. F13 and F18 reproduce in the second domain, and the grader noise floor is domain-specific — each has its own subsection below.

F13 — visible text is truncated relative to hidden reasoning (gpt-oss, role-free B)

The runner logged message.content and fell back to message.reasoning only when content was empty. A response with short content and a long hidden reasoning trace was logged as the short content, and used_reasoning_fallback stayed False, so the loss is undocumented.

Its Arm B planner logs a median 37 characters at 2.85 chars/sec, against 4,824 characters at 195.16 chars/sec in its own Arm A. 75.7% of that cell's role messages fall below 10 chars/sec; the documented flag fired on 0 of the 2,302 sub-100-character planner messages.

Its planner output is uninstrumented, not absent. The work was done and the logger discarded it. Any text-based measure over this cell measures the instrument. Detect with chars_per_sec; exclude the cell and say so.

A later commit fixed the logger. It did not reach any released row — measured directly, the gpt-oss Arm B planner median is 37 chars before the fix commit and 36 after, so the release is uniform. Note config_version does not change across that fix and therefore cannot be used to detect this.

F18 — correct grades the last SUBMITTED answer, final_answer is the last answer HELD

A team that revises after its final submission is graded on the answer it abandoned. Mismatch rate among first-pass-wrong trajectories:

actor	arm A	arm B	arm C
gemma-4	0.171	0.000	0.363
gpt-oss	0.164	0.066	0.258
Llama-3.3	0.058	0.004	0.478

So the defect reaches up to ~17% in Arm A and considerably higher in Arm C. Totals: 509 mismatched first-pass-wrong trajectories, 62 demonstrably under-credited (held answer exactly matches gold, graded 0). Positive control on the 10,734 non-mismatched graded rows finds 0 such cases, so 62 is a clean floor rather than a matching artifact.

The defect is concentrated in the arms that revise, i.e. it works against the frozen A-vs-B recovery gap — correcting it would widen that gap. Primary analysis stays as recorded (it is the pre-registered outcome and the conservative direction). f18_mismatch and f18_under_credited are for bounding, not correcting: the rate is a function of revision behaviour, which is the treatment.

F18 in SciBench — it reproduces, and again runs against the A-vs-B gap

The defect transfers to the second domain. Among first-pass-wrong trajectories:

actor	arm A	arm B
gemma-4	0.288	0.000
gpt-oss	0.105	0.017

49 mismatched first-pass-wrong trajectories, 2 demonstrably under-credited (held answer exactly matches gold, graded 0). The positive control on the 2,247 non-mismatched graded rows finds 0 such cases, so 2 is a clean floor rather than a matching artifact — the same control, and the same result, as the Omni-MATH layer.

Same direction as Omni-MATH: it is concentrated in the arm that revises, so correcting it WIDENS the gap. Re-scoring on the held answer moves gemma-4's paired recovery contrast from −0.1919 (p=6.6e-05) to −0.2020 (p=3.6e-05), and leaves gpt-oss exactly unchanged at −0.0833 (p=0.169). No verdict flips in either domain. Primary analysis stays as recorded — the pre-registered outcome and the conservative direction.

As in v1.0.0: f18_mismatch is for bounding, not correcting, and must never be used as a filter or covariate in an arm contrast — the mismatch rate is a function of revision behaviour, which is the treatment.

The grader noise floor is domain-specific — 2.80% here vs 0.72% on Omni-MATH

Disagreement rate across byte-identical candidate strings graded at temperature 0, keyed on exact candidate bytes:

domain	discordant / multi-graded groups	floor
Omni-MATH-2	20 / 2,790	0.0072
SciBench	19 / 678	0.0280

The SciBench grader is ~4× noisier, which is unsurprising: it must judge unit equivalence, significant figures and scientific notation rather than symbolic equality. The gemma effect (−0.19) is far larger than this floor, but do not compute "effect ÷ floor" — that statistic is withdrawn project-wide in both domains; dividing an effect size by a disagreement rate is not a valid operation, and grader noise largely cancels in a paired contrast anyway.

A note on a figure you may see elsewhere. The campaign notes compare SciBench's 2.80% against "9.35% on Omni-MATH". That 9.35% is not an Omni-MATH measurement from this project — it traces to a finalizer- nondeterminism figure recorded on a different project and was carried over as a comparison. Measured on this release's own frozen data by the same method, Omni-MATH's floor is 0.0072 (0.0086 under the superseded whitespace-stripped key). So SciBench's grader is ~4× noisier than Omni-MATH's, not 3× quieter. Use 0.0072 vs 0.0280.

F13 reproduces in SciBench, in the same cell

gpt-oss-120b Arm B is uninstrumented here too: planner median 123 chars at 20.0 chars/sec and executor/reviewer at 8.6 / 7.7 chars/sec, against 2,726 chars at 212 chars/sec for its own Arm A planner. 34–64% of that cell's role messages fall below 10 chars/sec. gemma-4 is clean in both arms (≥127 chars/sec throughout).

Any text-length measure over gpt-oss-120b Arm B is invalid in this domain too. Detect with chars_per_sec, never with config_version.

Two things are better here: there are no mangled LaTeX escapes (0 of 2,296 final_answer_held, 0 of 3,699 candidate_raw, versus ~1% in Omni-MATH), and no attrition at all — 0 hard failures, 0 unjudged, all four cells exactly 574.

Structural-enforcement (C) attrition — 8.0 / 9.9 / 17.0%, and it is one-sided

Regeneration-exhaustion attrition is 0.080 (gemma-4), 0.099 (gpt-oss), 0.170 (Llama-3.3), against 0.000–0.025 in Arms A/B.

The recovery denominator must INCLUDE trajectories that got a first verdict and then died in regeneration. An earlier implementation used graded(r) and first_pass==0, which silently dropped them — and all of them are first-pass-wrong structurally, because a PASS on the first verdict ends the loop. They meet the definition and score 0 under pre-registered scoring.

Three counts circulate; each is correct for its own scope — quote the one matching your sentence: Arm C only 150; Arm C + C′ 173; all arms 178 (A 1, B 4, C 150, C′ 23).

Correcting this moved recovery to gemma-4 0.314, gpt-oss 0.395, Llama 0.139, and cost two of three paired A-vs-C p-values.

Adjudication condition leak — outcome-blind, NOT condition-blind

adjudication/PROTOCOL.md records role names neutralized to Agent-P/E/R with "residual leaks: 0". True of the strings, false of the signal: the tokens were renamed, their presence was not equalized.

arm	has an Agent-P/E/R token	none
A	42	0
B	3	53
C	0	44
C′	0	29

Fisher exact A vs B on token presence: p = 1.5e-24. The arms are further separable by output format — Arm C carries JSON answer/verdict fields while A/B carry \boxed and [Agree]. A judge could identify the condition without reading content.

Unaffected (within-item): judge stability (0.947 / 0.959 / 0.953 for Q1/Q2/Q5) and precision 1.000 / recall 0.262. Affected: every arm-contrastive adjudication number, including A 0.000 vs B 0.964 and C 0.136.

Do not describe any arm contrast here as "blinded". A re-run would not fix it — the manipulation is partly an intervention on output format — so this is a disclosure, not a defect awaiting repair. The paper's organization claim rests instead on the mechanical measures, which involve no judge: functional collapse 0.976 / 0.846 / 0.872 in Arm B vs 0.000 / 0.120 / 0.000 in Arm A, and numeric-token duplication as excess over a shuffled null (B−A +0.290 gemma-4, +0.287 Llama-3.3; gpt-oss excluded per F13).

Repaired denominator bugs — graded() vs judged()

The survivorship error above had four known sites. Two mattered:

• Recovery — fixed as described.
• First-pass accuracy in Arm C, one column to the left, same error. Arm C first-pass moved gemma-4 0.822 → 0.779, gpt-oss 0.788 → 0.761, Llama 0.295 → 0.281, C′ 0.825 → 0.810 and 0.773 → 0.740. The paired A-vs-B contrast is unchanged (it comes from the paired path).

A self-check now asserts (1 − first_pass) × n == n_fpw per cell. If you recompute a denominator, use judged(), and state which of the three attrition-inclusive counts you used.

Also worth knowing

• Grader noise floor: 0.0072 group-level — 20 of 2,790 distinct graded strings receive contradictory verdicts across byte-identical prompts at temperature 0. Range 0.0017–0.0072 across estimators. Not 0.0086; the key must be the exact candidate bytes (candidate_normalized).
• Mangled LaTeX escapes: 135 of 11,498 final_answer_held and 176 of 17,486 candidate_raw values contain C0 control characters where a macro was intended (\frac → \x0crac). Preserved as-is so the artifact matches what the evaluator actually graded.
• One adjudication item (T0016) has only 2 reps and a tied Q5 vote. The pooled Arm B rate is 0.964 under the project's aggregation and 0.982 if the tie breaks the other way. One item, and it is the only tie.

Scope and limitations

• Two benchmarks, two domains — but the second is a partial replication. Competition math (Omni-MATH-2, 3 actors, 4 arms) and college chemistry/physics (SciBench, 2 actors, 2 arms only). SciBench has no Arm C, no Arm C′, no Llama-3.3 and no adjudication study. One actor (gemma-4) replicated in the second domain; the other cell cannot resolve its own effect (power ≈0.26). This is not "the finding generalizes".
• Three actors, and the effect does not replicate in all three. gemma-4 is clean (recovery 0.222 → 0.051, paired McNemar p=1.95e-08). gpt-oss is weaker and not significant when paired (0.324 → 0.247, p=0.208). Llama-3.3 does not replicate (0.014 → 0.015). Llama's null is not a floor effect — Arm C restores recovery in that same actor (0.096 vs 0.009, p=3.81e-06), so the repair channel exists and role labels simply do not open it there.
• Recovery is oracle-gated. No cell recovers before the first oracle verdict. This measures repair under external feedback, not spontaneous self-correction.
• The grader shares a checkpoint with one actor. Relevant to any self-grading concern.
• The pre-registered primary (P4) is null. All three actor × enforcement difference-in-differences contain zero. Two of the three have MDE 0.069 and 0.075 against plausible effects ≤ 0.040, so only the gpt-oss–gemma DiD (MDE 0.033) is an informative null.
• Adjudication is model adjudication (claude-opus-5), not human annotation and not ground truth.
• Wall-clock timings reflect shared infrastructure, including queueing.

Licensing

Per source — this release does not blanket one licence over source-derived material. See LICENSES/README.md and LICENSES/registry.json.

• Our derived material (parquet structure, scripts, docs, manifests): Apache-2.0.

  MIXED LICENCE as of v1.1.0 — do not blanket-apply one licence. Filter by benchmark_id to know which terms govern which rows.

content	licence	verified
Omni-MATH-2 problems (benchmark_id='omnimath2')	Apache-2.0	upstream HF API tag license:apache-2.0 + cardData.license, anonymous fetch, cross-checked against two registry files
SciBench problems (benchmark_id='scibench')	MIT, © 2023 Xiaoxuan Wang	upstream LICENSE file fetched anonymously + GitHub API spdx_id: MIT + two registry files — three independent sources, all agreeing
This project's derived material (parquet structure, scripts, docs)	Apache-2.0	authored by us
Model-generated trajectory text	the model providers' terms	—

SciBench's MIT is not absorbed into the Apache-2.0 material; its verbatim notice ships at LICENSES/MIT-SciBench.txt as the MIT terms require. The HF license: metadata field carries one value and describes our own material only. LICENSES/README.md and LICENSES/registry.json are the authority.

What upstream text is actually redistributed. The parquet layer ships no problem_text column and no gold-answer column, in either domain.

• Problem text is incidentally redistributed inside transcripts. Agents restate the problem: a 60-character problem prefix appears verbatim in role_messages.text for 16 of 606 Omni-MATH problems (2.6%) and 0 of 574 SciBench problems (0.0% — short numeric items get paraphrased into tables of givens). The raw traces carry more in both domains.
• Gold answers: no column, but a small measured leak in grader free text. The evaluator was instructed never to reveal ground truth and occasionally names the expected value in its repair hint anyway: Omni-MATH 23 of 17,486 evaluator_raw rows (14 of 1,600 problems, 0.9%); SciBench 19 of 3,699 rows (4 of 574 problems, 0.7%). v1.0.0 stated "no gold/reference field exists in any released row" — true at the column level, which is what was checked, but it did not cover free text the grader wrote. Corrected here for both domains rather than left standing. Far too sparse to reconstruct a gold set, and permitted under both licences with attribution.

Both upstreams are marked evaluation_only_recommended: we recommend evaluation and analysis use rather than training, in both domains.

Provenance and integrity

Derived from a sha256-manifested frozen snapshot (raw/MANIFEST.json, freeze 2026-09-30T10:09:30Z). CHECKSUMS.txt covers every file; MANIFEST.json records per-trace sha256, rows, bytes, semantic version, timestamps, arms, actors, domain, backend and config version.

scripts/validate_artifact.py re-derives 181 published values from the parquet layer and checks them against the project's canonical analysis output — 181/181 pass. The code repository re-derives every number in the paper from raw/ and controls/.

Two config_version values appear (e3afa39849c1b980, 9,600 rows; 231b8c4fe205d04b, 2,000 rows). They differ only by the addition of the Arm C′ entry — every A/B/C prompt, schema and knob is byte-identical across them, verified by recomputing the hash at each source commit.

Redaction (v1.2.0)

In 18 trajectory rows the error field held a response page returned by the inference gateway (14 in raw/, 4 in controls/run1/bf_omnimath_llama33.jsonl). v1.2.0 replaces each such error value with a fixed marker. Every other row is byte-identical, every other field keeps its value, and no analysis reads that text, so every number is unchanged. Each manifest records the original file's hash as redacted_from_sha256. The parquet tables never contained these bodies.