TODO.md

# Known limitations and open items

## 1. The post-answer confidence probe cannot be re-run from this release

**Status: confirmed limitation, not a packaging oversight.**

**Reproducibility limit.** The post-answer probe **cannot be fully re-run from
released artifacts**, because the Baseline final-answer string that the probe
consumes was **not persisted for the gpt-oss runs**. The probe's *outputs* and
*metrics* are fully released and independently reproducible. Recoverable
coverage is 6,462 / 12,928 probe rows (50.0%): 100% for the three Gemma
settings, 0% for the three gpt-oss settings. See the coverage table below.

The probe's prompt consumed two inputs per problem: the **problem text** and the
**Baseline protocol's final answer**. Neither is in this release.

- *Problem text* is withheld for licensing reasons. It is recoverable from
  upstream; see `docs/reconstruction.md`.
- *Baseline final-answer text* is **not released at all**.

`data/probe_inputs.jsonl` is therefore an **identifier and metadata manifest** —
a record of which examples the probe covered, and a join key to the labels. It is
not a runnable prompt set, and it is not a partially-redacted one either: the two
prompt-bearing fields are absent from the schema rather than present-and-null.

To re-run the probe you would need, per example: (a) the problem text rehydrated
from upstream, and (b) that setting's Baseline final answer. Without (b) the
probe is not reproducible by anyone outside the project.

### Is the Baseline final answer recoverable at all?

Checked directly against the project's source tree rather than assumed. The
answer is **partially, and not for the settings that matter most**.

Baseline `final_answer` text exists in the project's internal per-protocol
outcome files, but it is non-empty for only some runs:

| Setting (the 6 with probe coverage) | rows | with non-empty Baseline final answer |
|---|---:|---:|
| `omnimath2__competition_math_4181__gemma_4_31b` | 4,181 | 4,179 (100.0%) |
| `omnimath2__competition_math_4181__gpt_oss_120b` | 4,181 | **0 (0.0%)** |
| `labbench__text_no_tool__gemma_4_31b_text_no_tool` | 1,542 | 1,542 (100.0%) |
| `labbench__text_no_tool__gpt_oss_120b_text_no_tool` | 1,542 | **0 (0.0%)** |
| `labbench__llm_strict__gemma_4_31b` | 741 | 741 (100.0%) |
| `labbench__llm_strict__gpt_oss_120b` | 741 | **0 (0.0%)** |
| **total** | **12,928** | **6,462 (50.0%)** |

Across all ten released settings the figure is 8,622 / 15,088 (57.1%); the
gpt-oss-120b LAB-Bench and OmniMath runs contribute zero, while the gpt-oss-120b
JEEBench and SciBench runs are complete.

The split is clean along solver lines for three of the five gpt-oss settings:
those runs recorded a correctness verdict without persisting the answer string.
Why those three runs stored an empty `final_answer` while the other gpt-oss runs
did not was **not determined**; it is a property of the original run export, and
the text does not appear to exist anywhere else on disk.

**Consequence.** Even with full project access, the probe could be reconstructed
for the Gemma settings only. For the gpt-oss settings the input no longer exists.
Publishing the Gemma-side answers is a decision for the release lead — it was not
taken here, because a half-covered field invites exactly the kind of silent
partial analysis this dataset is trying to prevent.

**Open item for the release lead:** decide whether to (a) release Baseline
final answers for the six Gemma-covered settings with an explicit coverage
column, (b) release none and keep this limitation as stated, or (c) re-export the
missing gpt-oss answers from raw traces if they survive there.

## 2. Three-row disagreement between two source tables

`omni2_94`, `omni2_3997`, `omni2_4006` disagree on `baseline_correct` between the
primary routing table and the post-review matched labels. The post-review labels are
staged because they reproduce the published aggregates exactly. Cause not
determined. See `docs/provenance.md`.

## 3. Confidence-probe parse rate

The primary OmniMath probe has a 73% clean-parse rate: 113/423 test and 109/416
dev rows are `status=fallback_after_failure`. They carry a probability produced
by the fallback path, not a clean parse. Filter on `status` before using them.

## 4. Upstream licenses were read locally, not fetched

Every license recorded in `docs/license_audit.md` and
`registry/release_manifest.json` was read from a file in the project tree, not
retrieved from the upstream source at audit time. Re-verify before public upload.

## 5. Per-problem token costs for nine of ten settings — UNRESOLVED

`data/costs/omnimath_per_protocol_costs.csv` covers exactly one setting,
`omnimath2__competition_math_4181__gpt_oss_120b` (4,155 of its 4,181 problems).
The other **nine settings have no per-problem cost data**, and the paper's cost
axis is tokens, so this is a real gap in the artifact.

**Do not back-fill it from an incompatible source.** A companion dataset exists
with per-protocol token columns covering our settings. Those numbers were
evaluated and **rejected**:

- For gpt-oss OmniMath over all 4,181 problems, that source's mean Baseline
  `total_tokens` is **3,449.1**; the paper's own basis is **18,385.4** — a
  **5.33x** difference.
- Both are internally self-consistent. They measure **different quantities under
  the same column name**: that source records a single call's
  prompt+completion, whereas the paper's accounting sums **every model call in
  the protocol**. Baseline averages 9.67 model calls per problem here, which is
  the factor.
- Staging those figures would have silently contradicted the paper's published
  cost axis (Baseline 18.2K) while looking entirely plausible.
- Its coverage is also worse than it appears: the PER token column is 100% NULL
  for every setting, and Baseline/Single/Broadcast are 0% populated for both
  LAB-Bench text-no-tool settings.

Any future cost layer for the remaining nine settings must be derived from the
**same protocol-level accounting** the paper uses — total tokens summed across
all model calls — and must be checked against a published figure before release.
A column named `total_tokens` is not evidence that it measures what this dataset
means by total tokens.

**Open item for the release lead:** either derive per-problem costs for the nine
remaining settings from the project's own per-protocol run accounting, or leave
the gap documented as it is. Do not mix accounting conventions within one column.

## 6. Artifacts that exist but were not staged in this pass

Frozen-LLM-router and confidence-cascade per-problem predictions for the primary
423-problem split; per-problem cost accounting; raw protocol traces. All exist in
the project tree. See `docs/provenance.md`.