← Về thư mục
📄 / / root / ceo-project / academic-research-skills / academic-paper-reviewer / references / calibration_mode_protocol.md

Calibration Mode Protocol

Status: v3.2 Parent skill: academic-paper-reviewer Mode name: calibration Purpose: Measure this reviewer's own false-negative rate (FNR), false-positive rate (FPR), balanced accuracy, and severity-miscalibration rate (#215) against a user-supplied gold-standard set, then attach the resulting error profile as a confidence disclosure to subsequent reviews in the same session.


Why this mode exists

A single LLM reviewer produces an absolute 0-100 rubric score, but that score is weakly interpretable without knowing the reviewer's error profile. Two reviewers could give the same paper a 65, yet one might systematically over-score weak methodology papers and the other might systematically under-score cross-disciplinary work. Absolute scores don't reveal this.

Lu et al. (2026, Nature 651:914-919) demonstrated in Table 1 that an LLM-based Automated Reviewer can approach human balanced accuracy (0.65 vs human 0.67-0.73 on 500 ICLR 2022 papers) while having a dramatically different error profile: FNR 0.17 vs human 0.52, at the cost of FPR 0.50 vs human 0.17-0.34. Human reviewers miss half of the papers that should be rejected; the Automated Reviewer misses very few but over-rejects more.

Translation for ARS: our reviewer has an error profile too, and we do not currently measure it. Calibration mode closes that gap. It does not try to make the reviewer perfect; it makes the reviewer's imperfections legible.


Inputs

  1. Gold-standard set: 5-20 papers the user has labelled with known outcomes. Minimum 5; recommended 10-15. Each entry:
  2. Paper file path or text
  3. Ground-truth label: accept, reject, or borderline
  4. Venue context (journal/conference, tier)
  5. Optional: human reviewer scores for comparison

  6. Domain specification: the user's target field, used to seed field_analyst_agent. Calibration for "machine learning venues" is not valid for "qualitative education research" — error profiles are domain-specific.

  7. Session persistence: the error profile is cached for the current session only. No cross-session caching, no ~/.ars_calibration_cache/ directory. Calibration is explicitly opt-in per the v3.2 design decision: the user decides when to spend tokens on calibration, and a new session starts fresh. If the user wants to reuse a profile across sessions, they re-run calibration or paste a prior Calibration Report as a session prompt.


Process

Phase 0: Intake

Phase 1: Run full mode on each gold paper, with ensembling

For each paper, run the standard full review pipeline 5 times (ensembling, per Lu 2026 Methods A.1.1). Each run uses a fresh context window to avoid within-session bias. Aggregate: - Median rubric score per dimension - Variance across the 5 runs (reported as a stability indicator) - Editorial decision (majority vote across 5)

Cross-model verification: In calibration mode, ARS_CROSS_MODEL is default-on rather than opt-in. At least one of the 5 runs should use a different model family if available, to avoid single-model blind spots. If no cross-model is configured, emit a warning and run all 5 on the primary model.

Phase 2: Build the confusion matrix

Compare reviewer's majority-vote decision against the user's ground-truth label.

Compute:

Metric Formula Report with
Balanced accuracy (TPR + TNR) / 2 95% CI via bootstrap (1000 resamples)
FNR (miss rate) FN / (FN + TP) Same
FPR (false alarm) FP / (FP + TN) Same
AUC ROC over rubric-score threshold Same
Calibration error Mean |rubric_score - ground_truth_severity| Per-dimension

Phase 3: Borderline handling

Borderline papers don't enter the binary matrix but are useful for rubric-score calibration. For each borderline paper, report: - The reviewer's rubric score - The reviewer's decision - Whether the reviewer's decision respects the user's "this is borderline" signal (i.e., did it correctly land in Major Revision rather than confidently Accept or Reject?)

A reviewer that confidently Accepts or Rejects borderline papers has a "confidence miscalibration" problem even if its binary accuracy looks fine.

Phase 3.5: Severity-miscalibration measurement (#215)

The binary confusion matrix (Phase 2) measures decision-level error (FNR/FPR). It does not capture the paper's largest documented AI-reviewer failure: a finding that is content-correct but severity-miscalibrated — either a field-norm boundary error (Kim et al. 2026, W1, n=54) or the "would addressing this change the core result?" significance-boundary error (Kim §F.3.4, 56 errors). A reviewer can have a clean FNR/FPR and still systematically over- or under-rate the severity of individual findings.

For each weakness the reviewer emitted across the gold runs, classify its severity-miscalibration risk as low / med / high:

Grounding discipline (do not repeat the failure you are measuring). The classifier persona MUST NOT guess whether a norm is right from its own model knowledge — that is exactly the W1 behaviour under audit. It rates whether the reviewer supplied external grounding, not whether the reviewer's norm is factually correct. The reference shapes are anchored to the first-party regression fixture at evals/gold/field_norm_severity/ (W1 + §F.3.4 cases extracted verbatim from Kim et al. 2026); a finding that matches a fixture shape but lacks grounding is high.

This produces a histogram of low/med/high counts reported alongside FNR/FPR in Phase 4 — a severity-calibration signal the binary matrix cannot show.

Phase 4: Produce the Calibration Report

Output document structured as:

# Calibration Report for <Reviewer Instance>
Domain: <domain>
Gold set: n=<N> (accept=<a>, reject=<r>, borderline=<b>)
Runs per paper: 5 (ensembled)
Cross-model: <yes/no, model families used>

## Summary metrics
- Balanced accuracy: 0.XX [95% CI: 0.XX - 0.XX]
- FNR: 0.XX [95% CI ...]
- FPR: 0.XX [95% CI ...]
- AUC: 0.XX
- Ensemble stability: <mean std of rubric scores across runs>

## Comparison to Lu 2026 Table 1 baselines
| Metric | This reviewer | Lu 2026 Automated Reviewer | Lu 2026 Human |
|---|---|---|---|
| Balanced accuracy | X | 0.65 | 0.67-0.73 |
| FNR | X | 0.17 | 0.52 |
| FPR | X | 0.50 | 0.17-0.34 |

(Note: Lu 2026 numbers are for ML venues specifically. Compare with caution outside ML.)

## Per-dimension calibration error
<table of 7 review dimensions with mean absolute calibration error>

## Severity-miscalibration histogram (#215)
<low/med/high counts over all emitted weaknesses, e.g.>
| Risk | Count | Share |
|---|---|---|
| low | XX | XX% |
| med | XX | XX% |
| high | XX | XX% |
<A high `high`-share means the reviewer frequently asserts field-norm / "core result" severities without external grounding — the W1 / §F.3.4 failure shape. This is a SEPARATE signal from FNR/FPR: a reviewer can pass the binary gate and still carry a high severity-miscalibration rate. Grounded per Phase 3.5; classifies grounding, not norm-correctness.>

## Systematic biases detected
<natural-language narrative identifying patterns, e.g.
 "Reviewer tends to over-score originality on cross-disciplinary papers"
 "Reviewer under-scores qualitative methodology by ~8 points vs ground truth"
>

## Recommendations for session use
- Treat this reviewer's rubric scores as having calibration error ±X points
- For accept/reject decisions, the reviewer misses X% of reject cases (FNR)
- For decisions near the accept/reject boundary, escalate to human judgement

Phase 5: Session attachment

If session persistence is enabled, the Calibration Report is attached to every subsequent review in the same session as a confidence disclosure header. The disclosure appears in the editorial letter before the verdict:

> **Reviewer Confidence Disclosure (from calibration session <id>):**
> This reviewer has measured balanced accuracy 0.XX, FNR 0.XX, FPR 0.XX on a
> gold set of <N> papers in <domain>. Rubric scores below have calibration
> error ±X points. Treat borderline decisions with human judgement.

This is non-negotiable in calibration-enabled sessions: the user cannot hide the disclosure. The point of calibration is to make error profiles legible; suppressing the disclosure defeats the mode.


Ensembling methodology notes

Lu 2026 Methods A.1.1 describes reviewer ensembling across 5 independent runs with majority voting. This mode follows that spec with two changes:

  1. Median instead of mean for rubric scores: mean is vulnerable to single-run outliers (e.g., a run that hallucinates a methodological flaw); median is robust.
  2. Fresh context per run: Lu 2026 allowed within-session memory across runs. ARS uses fresh context to prevent cascading errors from a single run's misreading.

Users with token budget concerns can reduce runs_per_paper to 3. Below 3, ensembling is meaningless — do not allow 1 or 2.


Failure cases this mode does NOT fix

Calibration reports this reviewer's error profile on a specific gold set in a specific domain. It does not:

If the user's gold set is itself biased (e.g., all papers from one lab, all from one year), calibration reports a biased profile. Emit a warning during intake if papers share obvious metadata clusters.

Same-family / rubric-aware judging — read the numbers as a possible under-estimate

There is a second reason a measured profile can be optimistic, independent of the gold set. It belongs to the broader same-source evaluation risk, which has two forms:

This is an interpretive caveat only. ARS does not detect, prevent, or correct rubric-aware judging — the behavior can be unverbalized and is not reliably visible in chain-of-thought. The note changes how you read the numbers; it does not change any threshold or gate.

Cross-model evaluation — stronger evidence where available. Running the evaluation across model families provides stronger evidence than a same-family-only run; it still does not detect or rule out rubric-aware judging. Positioning:

A single-model spot-check (weak, optional). With no second model, you can reword the rubric and re-judge, then check whether the verdict changed. Be clear about what this does: it only tells you whether a change of wording shifts the judgment — surface wording sensitivity. It does not reveal whether the model is quietly optimizing toward the grader (that can be unverbalized), and a verdict that survives rewording is not evidence the judgment is correct — only that it is stable to that paraphrase. It is one model checking itself, so its power against grader-awareness is limited. No score, no threshold, no gate.

Directional prior: assume leniency relative to human expert review (FARS external anchor)

Beyond the same-family optimism above, there is a citable directional prior on the sign of the error: when the simulated 5-reviewer panel's output is read as a pass/fail signal, assume it runs lenient relative to human expert review until your own calibration measurement shows otherwise. FARS (Tang et al. 2026, arXiv:2606.31651) provides a deployment-scale external anchor: on the FARS deployment corpus, an ICLR-style automated reviewer (Stanford Agentic Reviewer) averaged 5.00 over the 165 papers it reviewed, while the paper-level mean from 282 human expert reviews covering 140 of those papers was 3.23 on the same 0-10 scale — a ~1.8-point gap (a descriptive difference between overlapping-but-unequal paper sets, not a paired estimate), and the automated score never functioned as an acceptance probability, only as a relative ranking.

How to use this prior:


Integration with existing modes

Existing mode Interaction with calibration
full Calibration runs full 5x per gold paper. No change to full itself.
re-review Calibration profile attaches to re-review decisions.
quick Calibration profile attaches. Confidence disclosure notes that quick has additional uncalibrated error on top of the measured profile.
methodology-focus Calibration should ideally be run with methodology-heavy gold papers if this mode is the user's target.
guided Not applicable — guided mode is Socratic dialogue, rubric scores are not the primary output.

Resolved design decisions (2026-04-09)


References

v3.6.2 sprint contract status

v3.6.2 introduces sprint contracts for reviewer_full and reviewer_methodology_focus only. A template for this mode will follow in a subsequent patch release. Until then, this mode runs without contract enforcement and retains its pre-v3.6.2 behaviour.