fahadhafeezofficial/evitrace_bench

EviTrace v1.0

Author: Fahad Hafeez

A synthetic benchmark and generator testing whether AI systems can tell genuinely independent evidence apart from evidence that's copied, paraphrased, summarized, or reported secondhand from the same underlying source. Full design rationale and related-work positioning in docs/evitrace_spec_v0_1.md.

This is a real, QC-validated 8,000-scenario release — not a toy sample. It is also not a claim of total completeness: see "What's still open" below before treating it as a finished benchmark with leaderboard-ready baselines.

What's actually in this package

docs/evitrace_spec_v0_1.md   the differentiated spec (formal model, taxonomy, tasks, metrics, related work)
src/evittrace/                generator, transforms, topology builder, metrics, QC — all real code, no stubs
scripts/                      generate_sample.py, run_qc.py, sanity_check_metrics.py, compute_stats.py, measure_diversity.py
tests/                        17 pytest tests (generator, QC, all 4 metrics incl. hand-built edge cases)
data/v1/public/                8,000 scenarios, model-visible fields only (scenarios.jsonl)
data/v1/hidden/                8,000 scenarios, full hidden evaluator ground truth (scenarios.jsonl)
reports/qc_report.json        12/12 automated QC checks, actual result
reports/dataset_stats.json            actual measured statistics over the 8,000-scenario release
reports/metrics_sanity_check.json     oracle vs. naive-baseline sanity check
hf_upload/public_repo/        exact files + dataset card to upload as the PUBLIC Hugging Face repo
hf_upload/private_hidden_repo/  exact files + dataset card for a PRIVATE hidden-evaluator repo (do not make public)
LICENSE-CODE.md               MIT, covers src/ scripts/ tests/
LICENSE-DATA.md               CC BY 4.0, covers data/ and hf_upload/
CITATION.cff                  citation metadata, author Fahad Hafeez, no invented DOI/affiliation
CHANGELOG.md                  what actually happened, in order — no planned work listed
requirements.txt              pydantic, networkx, pytest, pyyaml

Do not upload data/v1/hidden/ or hf_upload/private_hidden_repo/ to a public location. It's the answer key — see hf_upload/private_hidden_repo/README.md.

Reproduce it

pip install -r requirements.txt
python3 -m pytest tests/ -q
python3 scripts/generate_sample.py --n 8000 --seed 13 --out data/v1
python3 scripts/run_qc.py --data data/v1 --reports reports
python3 scripts/sanity_check_metrics.py --data data/v1 --reports reports
python3 scripts/compute_stats.py
python3 scripts/measure_diversity.py

Generation is deterministic given --seed; every number below is exactly reproducible.

Actual results (not projected)

Tests: 17/17 passed.

QC: 12/12 checks passed over 8,000 scenarios, including a claim-duplicate-rate check (measured 1.89%, threshold 5%) added specifically because the first 5-domain version of the generator had a real duplication ceiling — see "What changed since the v0.1 sample" below.

Dataset statistics (8,000 scenarios): 56,267 documents, 26,633 evidence events, 73,682 provenance edges, 7.03 documents/scenario, 3.33 events/scenario, max document-tree depth 4. Domain split is even across all 10 domains (776–843 scenarios each). Topology split is even across chain/star/tree/dag (1,983–2,019 each). Sufficiency labels: 4,453 sufficient, 2,233 contradictory, 1,163 insufficient, 151 ambiguous — all four categories genuinely represented, which the original v0.1 sample failed to do (it had zero "insufficient" cases). 78.8% of scenarios contain a real double-counting opportunity. Full numbers in reports/dataset_stats.json.

Metrics sanity check (oracle = ground truth fed back as a "prediction"; naive = the deliberately-wrong baseline that treats every supporting document as independent confirmation):

Metric	Oracle (should be perfect)	Naive source-counting baseline
EII (Evidence Inflation Index)	0.0	0.509
DCR (Double-Counting Rate)	0.0	0.851
MEE (Minimal Evidence Efficiency)	1.0	0.327
PRS (Provenance Reconstruction, edge-F1)	1.0	(not defined for the naive baseline)

Consistent with the smaller v0.1 sample's numbers (EII 0.50, DCR 0.87–0.94, MEE 0.31–0.34) — the failure signature isn't an artifact of small-sample noise.

What changed since the v0.1 sample (and why)

You asked for a final version, not a sample. Simply generating more rows from the 5-domain v0.1 generator would not have earned that label: the infrastructure domain alone had a hard ceiling of 210 unique bridge/year combinations (6 bridge names × 35 years), so any serious scenario count would have collided constantly on the same handful of underlying facts in that domain. Before scaling up, this release:

1. Expanded from 5 to 10 domains (added technology, transportation, education, environment, legal).
2. Widened entity spaces via combinable parts (e.g. 20 bridge-name stems × 4 suffixes = 80 names, not 6) so no domain's discrete combinatorial ceiling sits below several thousand.
3. Widened the phrasing-transformation pools (copy prefixes, paraphrase openers, synonym substitutions, derive/summary templates) from 2–4 variants each to 4–8, and fixed a real bug where chained transformations (a near-copy of a paraphrase, etc.) stacked reporting-verb filler on top of itself into grammatically awkward text.
4. Added the claim-duplicate-rate QC check described above, and a measure_diversity.py script that empirically measured duplicate rate at 1,000 / 2,000 / 4,000 / 6,000 / 8,000 / 10,000 scenarios before picking 8,000 as the target scale (1.16% duplicate rate at 10,000 vs. 1.89% measured at 8,000 in the final run — both well under a 5% threshold).
5. Fixed the v0.1 sample's sufficiency-label gap (zero "insufficient" cases) by allowing zero-support scenarios to be sampled.

What's still open — not silently dropped

• Structural diversity vs. lexical diversity. Claim text is genuinely varied (1.89% duplication at 8,000 rows), but the shapes of evidence graphs (topology mix, transformation-chain length distribution) are drawn from the same fixed probability tables regardless of scenario count. A model trained at scale on this data could plausibly learn "shape" patterns that a larger, more varied generator wouldn't produce. Not measured or fixed in this release.
• I6/I7 independence split (correlated-but-independent vs. fully independent) — still collapsed into one class, per spec §3.3.
• Task coverage: E5, E8, E9 only. E1-E4, E6, E7, E10-E12 have their ground truth sitting in the hidden export (provenance edges, per-document stance, source reliability, timestamps) but no dedicated task-extraction or scoring wrapper yet.
• Baseline coverage: naive only. No embedding-similarity, classical-ML, or LLM baseline has been run. The naive baseline is a deliberately weak strawman (spec §32) — it establishes that the failure mode exists and is measurable, not how hard the benchmark is for a competent system.
• 10 domains is still a template ceiling, not open-ended text. A system could in principle overfit to the 10 templates' surface structure rather than learning general provenance reasoning. Worth stress-testing before treating results on this data as generalizable claims about a model's real-world evidence reasoning.

Honest next steps

1. Build the remaining task-extraction/scoring wrappers for E1-E4, E6, E7, E10-E12.
2. Add an embedding-similarity baseline (free, local) as a second reference point.
3. Decide, with you, whether the structural-diversity gap above is worth addressing (e.g. widening the topology/transformation-chain probability tables) before spending API budget on a real LLM baseline.