docs/provenance.md

# Provenance

## What this release was built from

Three source trees, all read-only at staging time.

1. **Camera-ready ancillary aggregates** — the eight CSVs shipped with the
   camera-ready submission (`camera_ready_2026-08-07/anc/`, dated 2026-08-11 and
   2026-08-14). Copied verbatim into `data/aggregate/`, with a `setting_id`
   column added. These are the authoritative published numbers.

2. **Post-review analysis artifacts** (`results_2026-07-12` through
   `results_2026-07-14`) — per-problem matched labels, six-setting router
   retraining, confidence-probe metrics, and the Round-2 verification tables.
   Source of `data/matched_labels.csv`, `data/splits/six_setting_splits.csv`,
   `data/router/*`, and `data/confidence/six_setting_*`.

3. **Primary-split routing artifacts** — the 4,181-problem OmniMath routing
   benchmark, its 80/10/10 split, and the per-problem confidence-probe outputs.
   Source of `data/splits/primary_omnimath_splits.csv`, the OmniMath rows of
   `data/problems.csv`, and
   `data/confidence/primary_omnimath_confidence_predictions.csv`.

Problem metadata for JEEBench, SciBench, and LAB-Bench comes from the project's
prepared benchmark slices; the canonical-ID crosswalk comes from the project's
own problem-crosswalk registry.

Source commits are recorded in `registry/release_manifest.json`.

## Verification performed at staging time

Each of these was computed, not assumed. Full numbers are in the release's
validation report.

- **Oracle recomputation.** Recomputing `oracle_label` from the four correctness
  flags in the fixed order Baseline -> Single -> PER -> Broadcast reproduces the
  stored label for **all 15,088 released rows, zero mismatches**. Also zero
  mismatches on `any_protocol_solved`.
- **Aggregate reproduction.** The per-problem labels reproduce
  `anc/matched_protocol_coverage.csv` for all ten settings to two decimals, and
  `anc/oracle_label_distribution.csv` for all ten settings to two decimals.
- **Cross-solver problem-set identity.** For all five paired
  (benchmark, condition) combinations, the two solvers cover an identical
  canonical problem set after ID reconciliation — symmetric difference empty in
  every case.
- **Uniqueness.** `(setting_id, problem_id)` is unique in every table; no
  duplicates in any setting.
- **Split disjointness.** Train/dev/test are pairwise disjoint and exhaustive in
  both split schemes.
- **`example_id` reproducibility.** The `example_id` values are
  `sha1("<setting_id>:<source_problem_id>")[:12]` prefixed by the setting. This
  formula was re-derived and matches the stored values in all six settings.

## Known discrepancies, stated rather than hidden

**The primary routing benchmark and the post-review matched labels disagree on 3 of
4,181 OmniMath gpt-oss-120b problems** (`omni2_94`, `omni2_3997`, `omni2_4006`).
In each case the primary table records `baseline_final_passed = False` while the
post-review matched labels record `baseline_correct = 1`, flipping the oracle label
from `single_agent` to `baseline_llm`. Aggregate impact: baseline solve rate
56.76% (primary) versus 56.83% (post-review, the published number); label counts
differ by 3 between the two tables.

This release stages the **post-review** labels in `data/matched_labels.csv`, because
those are the ones that reproduce the published `anc/` aggregates exactly. The
primary table is used only for OmniMath problem *metadata* (tier, difficulty,
source) and for the 423-problem split — neither of which is affected. The cause
of the three-row disagreement was not determined at staging time.

**The primary confidence probe has a 73% clean-parse rate.** 113 of 423 test rows
and 109 of 416 dev rows fell back after JSON parsing failed on up to three
attempts. Those rows are released with `status = fallback_after_failure` so they
can be excluded. They are not clean measurements.

**Two `avg_tokens` cells are `nan`** in
`data/router/six_setting_text_metadata_verification.csv`, for the Gemma OmniMath
and Gemma LAB-Bench settings, where per-problem cost accounting covered only
5, 5, and 2 of the held-out problems respectively. The `cost_n` and
`missing_cost_n` columns record this directly.

## What is deliberately not here

- **MaScQA** (`mascqa__text_only__gpt_oss_120b`, 642 rows). Exists in the source
  tree. Excluded: no matched Gemma run, not one of the paper's ten settings, and
  CC-BY-NC-SA-4.0.
- **The Gemma-3-27B tier-sampled OmniMath slice**
  (`omnimath2__tier_sampled_833__gemma3_27b`, 833 rows). Exists in the source
  tree. Excluded: a scope check, not one of the paper's ten settings.
- **Raw protocol traces.** Multi-gigabyte generated text. Out of scope for this
  table-first release.
- **Per-problem token, model-call, and wall-time costs.** Available in the
  source tree for the OmniMath primary split only. Not staged; cost appears here
  only through the aggregate tables.
- **Frozen-LLM-router and confidence-cascade per-problem predictions.** These
  exist for the primary 423-problem split and support
  `data/aggregate/main_routing_heldout.csv`. Not staged in this pass.