EVALUATION.md

Evaluate new predictions and replay the paper

Install requirements.txt in a Python 3.12+ environment. The verified runtime and exact package versions are in protocol/verified-runtime.json. All commands below run on CPU, use local files, and make no model/API calls. Run from this dataset folder, with outputs outside it so the frozen manifest stays valid.

New model predictions

Use only image and question as model inputs. Read the exact prompt and schema from protocol/prompt.json. Obtain GQA inputs using RECONSTRUCTION.md first. Write one JSON object per input in predictions.jsonl:

{"item_id":"item-0001","answerable":false,"answer":null,"reason":"The needed value is not visible."}

That is a format example, not a prescribed answer for other inputs. Retain each stable item ID. For a terminal malformed model output use:

{"item_id":"item-0001","status":"invalid_schema","terminal_invalid":true,"parsed":null}

One row is required for every selected input. Do not submit both examples for the same ID. Missing/duplicate/unexpected IDs or incomplete state groups cause the scorer to fail rather than silently shrink the denominator. Invalid final outputs count as failures. A valid answerable decision with a null answer can pass decision scoring but fails answer correctness. Unanswerable decisions must have a null answer and all valid outputs need a nonempty reason.

python scripts/score_hf_dataset.py --predictions ../predictions.jsonl \
  --sources plotqa clevr gqa --output ../scores.json

For a single subset, supply only its predictions and e.g. --sources plotqa. Output includes exact counts, per-state failures, per-view decision accuracy, all-decisions-correct groups, supported-answer accuracy, class-balanced failure, and complete joint answer/abstention success. Chart answers use the paper's conservative exact-rational parser (no numeric tolerance). Scene scoring first checks exact rationals, then Unicode NFKC, case and whitespace normalization; it does not infer synonyms. A percentage marker follows the historical parser and is not automatically divided by 100.

The GQA flag in the metadata defines the same 376-group location sensitivity stratum as the paper. The supplied saved-response replay reports both strata.

Reproduce reported statistics

python scripts/replay_hf_statistics.py --output ../replay

This checks all 72,000 scoring projections, recomputes all 18 model/domain cells and the paper's 10,000-draw paired group bootstrap intervals, compares the full analysis JSON to the frozen expected result, regenerates tables and chart diagnostics, verifies the 5,000 chart view-to-proof identity joins, and runs the new-prediction scorer against all six saved configurations. It also checks the finite-study counts and enumeration/SMT agreement from saved per-case records. It does not regenerate the finite solver-study witnesses, raw generations or historical serving traces. GQA question-ID tokens replace upstream question text consistently in numerical replay; reconstruction checks the actual text.

The supplementary shortcut and answer-error counts can also be checked directly:

python scripts/verify_argument_evidence.py --inputs inputs \
  --finite-results evidence/finite/original-results.jsonl.gz \
  --solver-results evidence/finite/solver-results.jsonl.gz \
  --output ../supplementary-counts.json

python scripts/verify_hf_dataset.py --workers 8 --output ../verification.json

This verifies the frozen file manifest, loads all three configurations through Hugging Face Datasets, decodes and hashes all 9,000 bundled images, checks every state group and unique source-image identity, verifies all 1,000 complete chart proofs against their evaluated PNGs, and replays chart/CLEVR source programs. The GQA source-program and image check is run by the separate reconstruction command. A small machine may reduce --workers.

To run only the chart proof checker:

python scripts/verify_bounded_witness_bundle.py \
  --proofs evidence/plotqa/bounded-witness-proofs.jsonl.gz \
  --workers 8 --output ../chart-proof-verification.json

The prepared package includes historical counts plus current validation reports. Reports are evidence of the checks they describe, not a guarantee about every upstream annotation or a substitute for independent scientific review.