Skip to content

The 10-paper sweep results

This page reports the system’s empirical performance on the 10 public eLife papers in the curated reference corpus. Two full sweeps have run; v2 (with the iterated reviewer prompt incorporating Biases 5-7) is the current baseline.

For the methodology behind these numbers, see how we measure quality. For the iteration that produced v2, see iteration discipline.

Aggregate metrics — current baseline (v2)

Section titled “Aggregate metrics — current baseline (v2)”

Sweep ran 2026-05-11 with Sonnet 4.6 × 3 extraction agents, Opus 4.6 reconciler, Opus 4.6 external reviewer with the iterated 7-bias prompt, CrossRef DOI verification.

MetricMeanMedianThresholdStatus
Claim recovery96.9%100.0%≥ 80%✅ PASS
Role classification83.1%84.3%≥ 75%✅ PASS
Panel assignment60.9%58.6%≥ 90%❌ FAIL

Recovery is the load-bearing metric and it passes by a wide margin. Median recovery is 100% — every reference claim has a corresponding CLI extraction on half the corpus, and the lowest-recovery paper (rozak) still hits 87%.

Role classification clears its threshold by 8 points on the mean and 9 points on the median. The external reviewer’s structural-inference pass converts what would be a failure mode in auto-approve mode (~65% role agreement) into a strongly passing result.

Panel assignment improved from v1 (54.7% → 60.9%) but still fails the 90% threshold. The improvement came from the iterated reviewer’s better handling of multi-panel claims; the residual failure is the documented multi-panel collapse problem.

The v2 iteration added Biases 5-7 to the external reviewer prompt (hypothesis preservation, control role recognition, literature-context recognition). Headline deltas across the corpus:

Metricv1v2Δ
Recovery (mean / median)95.3% / 98.0%96.9% / 100.0%+1.6 / +2.0
Role (mean / median)78.1% / 76.4%83.1% / 84.3%+5.0 / +7.9
Panel (mean / median)54.7% / 53.1%60.9% / 58.6%+6.2 / +5.5

The iterated prompt is a measurable net win on every aggregate metric. See iteration discipline for how v2 was developed and validated.

Papern_refn_cliRecoveryPanelRole
artiushin (spider atlas)1779100%59%94%
bouyeure (fear RSA)3066100%47%90%
ejdrup (dopamine)258396%58%83%
gadeke (guilt insula)2765100%48%85%
headley (inhibitory rhythms)2689100%81%92%
kammer (foveal feedback)234396%91%82%
kolb (iGABASnFR2)2070100%75%70%
rozak (neurovascular DL)237287%60%55%
scheller (self-prioritization)2359100%48%83%
wengert (KCNC1)319190%43%96%

Most dramatic per-paper deltas vs v1:

  • kammer role 57% → 82% (+25 points), panel 52% → 91% (+39 points)
  • headley panel 54% → 81% (+27 points)
  • artiushin recovery 88% → 100% (+12 points), panel 60% → 59% (within noise)

Material regressions vs v1 (only two, both small):

  • rozak role 65% → 55% (-10 points). Methods paper; the iterated reviewer’s stricter control-recognition rule may over-apply on a paper that doesn’t have the typical hypothesis-test structure.
  • kolb role 75% → 70% (-5 points). Similar pattern, less material.

These are the residual failures the iteration trades for the corpus-wide gains. Targeted prompt iteration (per iteration discipline) on either paper could close these without regressing the rest of the corpus.

Recovery distribution: 87-100%, with 5 of 10 papers at 100%. The 87% (rozak) and 88% (artiushin) are the lowest; both are unusual paper structures (rozak is a deep-learning methods paper; artiushin is an observational atlas).

Role distribution: 57-96%. Three papers below the 75% threshold individually (kammer 57%, rozak 65%, scheller 70%). Mean and median both clear the threshold; the failures are clustered in papers with structural shapes the reviewer prompt was less calibrated for.

Panel distribution: 36-74%. Consistently below the 90% target. Wengert’s 36% is the worst; multi-panel reference claims dominate that paper.

The Headley v3 result generalizes. Headley alone (100% recovery, 96% role with external review) was a single-paper data point that could have been Headley-specific. The 10-paper sweep confirms the pattern holds across diverse papers — different domains (atlas, fMRI, computational, channelopathy), different argument structures (hypothesis-driven, observational, methods-focused), different scales (17 to 31 reference claims).

The external reviewer is doing real work. Without it (auto-approve mode on the same papers), role agreement averages ~65%. With it, 83% (mean, v2). The 18-point average lift comes from the structural-inference work the prose-level extraction systematically misses.

The iteration discipline works. The Phase F kammer iteration (single-paper test, +14 points on role) generalized to a +5 point average across the corpus, with the largest individual gain (+25 points) on the targeted paper itself. This is the canonical worked example of the iteration discipline.

Recovery is robust. Even on the lowest-recovery paper (rozak, 87%), the CLI surfaces 20 of 23 reference claims. Misses cluster in papers where the curated reference includes claims requiring deep methodological audit beyond what the Structure-reader agent attempts.

Panel agreement is the next iteration target. v2 lifted panel from 55% to 61%; the threshold is 90%. The Caption-reader prompt could be revised to more aggressively preserve multi-panel lists; the reconciler prompt could be revised to consolidate single-panel CLI claims when they jointly map to a multi-panel reference. Either change would lift the panel score further; both untested empirically.

After the iteration, only two papers fall below the 75% role threshold individually:

rozak (55% role) — neurovascular deep-learning paper. Methods-heavy with unusual claim structure (computational pipeline as the contribution, not biological findings). The iterated prompt’s stricter control-recognition rule may over-apply here. The default prompts are calibrated for biological-finding papers; methods papers may need a --prompt-variant.

kolb (70% role) — sensor engineering paper. Borderline case; small drop from v1’s 75%. Likely the same over-application pattern as rozak but smaller.

The 8 papers at 75% or above include the previously-lagging kammer (now 82%) and scheller (now 83%) — both moved into compliance with the iterated prompt.

The original 10-paper sweep (no iterated prompt; 4-bias reviewer) produced:

MetricMeanMedian
Recovery95.3%98.0%
Role78.1%76.4%
Panel54.7%53.1%

Per-paper v1 numbers preserved at tests/aggregate-scorecard-10paper.md in the repo for historical comparison.

PhaseCost
10 × extract (~$5 each)~$50
10 × external review (~$2 each)~$20
10 × matcher (~$3 each)~$30
Total~$100

Wall time: ~85 minutes sequential. Could run in 30-40 minutes with 3-way shell-level parallelism.

Validates:

  • The pipeline produces curator-quality claim coverage on diverse paper types
  • The external reviewer’s role-classification lift generalizes beyond Headley
  • The system’s per-paper cost (~$7 with reviewer) is a meaningful improvement over manual extraction (~4 hours of analyst time per paper)
  • The validation harness itself works at scale and produces actionable per-paper diagnostics

Doesn’t validate (or doesn’t validate yet):

  • Performance on non-eLife papers (the prompts are calibrated for eLife structure)
  • Performance on papers without typical experimental structure (preprints, methods papers, observational surveys may need prompt variants)
  • Long-term stability (the prompts are version-controlled; future model changes may shift behavior)
  • Cross-paper edge resolution (out of scope for round-trip scoring)
  • tests/aggregate-scorecard-10paper.md — the raw scorecard from this sweep
  • Per-paper scorecards: <work-dir>/<paper-slug>/scorecard.json
  • elife_extract/evaluate.py — the harness that produced these numbers