acceptodds
Under review as a conference paper at ICLR 2027

The Evidentiary Control Protocol: A Preregistered Governance Layer for Auditable AI Evaluation

Abstract

Scores reported by AI benchmarking frameworks do not, by themselves, establish how an observed answer was produced. When the question of interest is whether an outcome was obtained under a specified experimental condition — rather than through retrieval, memorization, model-mediated inference, or answer-bearing artifacts in the evaluation environment — the score is not the evidence. The evidence is a controlled chain from case construction to outcome classification. We present the Evidentiary Control Protocol (ECP), a preregistered, leakage-controlled governance layer that fixes a ten-stage lifecycle — case authoring, structural and novelty review, environmental knowledge-boundary verification, registration, conditioned execution, raw-evidence preservation, independent audit, mechanical outcome classification, descriptive statistical analysis, and representation-bias disclosure — and enforces one central separation discipline: execution validity, evidence integrity, protocol-defined outcome, and scientific interpretation are four distinct claims, recorded separately and never merged. ECP does not measure a model's internal state and does not adjudicate cognitive claims. The paper demonstrates the protocol operating across an evidence portfolio that now extends beyond the original five lines (the T1 pilot; the M1 pilot across six hash-frozen systems; the REASON-2 campaign; the REASON-3 defect-preservation/repair/re-execution chain; and the M2 comparative recording campaign) to an independent audit programme that applies the same discipline to externally published claims under bounded channels: A14: Commercial structured-output audit of Together AI — schema validity 24/24 = 1.000 under the registered configuration A15: Adversarial claim-injection audit — detection 1.00 on levels L1–L3, silent acceptance 0.00 A16: Adaptive adversarial red team — 141 submissions across four valid runs, 0 accepted, with a documented localization-metric boundary and final classification INCONCLUSIVE A17: Structured-output reliability audit of OpenAI's published claim — 60/60 schema-valid, verdict CLAIM_HOLDS_STRONG, bounded to the tested corpus, snapshot, and gateway channel A18: Bounded directional audit of Anthropic's published prompt-injection robustness claim — mechanical injection-success 3/30 = 0.100 on Sonnet 4.6 versus 14/30 = 0.467 on Sonnet 5, non-overlapping 95% Wilson intervals — a directional difference opposite to the published claim's direction under the tested conditions We are explicit about what this establishes and what it does not. The portfolio establishes the protocol's operation — registrations held, evidence preserved, adverse findings kept intact — and demonstrates that ECP-style independent audits can examine third-party commercial claims with bounded channel disclosure. It does not validate ECP as a scientific instrument for measuring reasoning, does not replicate any vendor's internal evaluation, and does not convert any bounded finding into a general claim. Every quantitative statement is bound to a sealed artifact in the accompanying claim ledger.

Primary area:
datasets and benchmarks
Keywords:
AI evaluation, evidentiary protocol, preregistration, evidence integrity, claim boundary, outcome classification, AI governance

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_kTkgOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

ECP proposes a ten-stage preregistered governance workflow for AI evaluation, separating execution validity, evidence integrity, mechanical outcomes, and subsequent scientific interpretation. The paper integrates deterministic and model-based campaigns, seeded-fault tests, reclassification exercises, and third-party structured-output and prompt-injection audits. It emphasizes immutable records, explicit denominator rules, preserved failures, and bounded claims. Reported demonstrations include 60/60 schema-valid outputs in A17 and 141 unaccepted adversarial submissions in A16, whose final registered verdict remains inconclusive. The paper presents these as operational demonstrations rather than validation of a reasoning-measurement instrument.

Critical strengths
  1. The separation of execution status, evidence integrity, and task outcome is useful and consistently illustrated by adverse cases. Transport failures, defective controls, repaired campaigns, and inconclusive outcomes remain visible instead of being silently converted into favorable scores. This makes the submission informative about practical evaluation failure modes.
  2. The manuscript is unusually explicit about the limits of its observations. It discloses auditor-independence limitations, channel and snapshot mismatches, quote-in-refusal effects, and the failed A16 localization gate. These qualifications make its bounded operational claims more credible than unqualified claims of capability or security robustness.
Critical weaknesses
  1. The evidence does not establish the added scientific or practical value of the integrated protocol over a competent existing audit workflow. M2 uses only six non-divergent cases, and both arms share the same answer parser and equality test; Appendix C states that answer-level disagreement is structurally impossible. Seeded fixtures validate implementations against designed rules, but do not compare independently detected faults, audit effort, or decision quality with an appropriate competing process. Since the controls themselves are not claimed to be new, this leaves the central integrated contribution insufficiently validated.
  2. The strongest commercial-claim examples do not test equivalent conditions, limiting their evidentiary relevance. A17 observes a different model snapshot from the one named in the source claim. A18 uses simulated browser context, single-message turn markers, and a gateway-added persona; some mechanical successes are quotations within refusals. The qualifications are disclosed, but an opposite direction in this altered environment is not a controlled test of the original claim. Much of the underlying evidence is referenced as repository paths and closure records rather than reproduced in the PDF, so its inspectability depends on the accompanying package.
Questions for authors
  1. What comparative evidence shows that ECP detects consequential evaluation faults or improves audit decisions beyond an existing workflow with preregistration, preserved logs, and separate transport/outcome accounting? Please identify an existing experiment that can discriminate the approaches, rather than only exercise the frozen classifier.
  2. Where can reviewers inspect the complete registrations, raw outputs, independent verifier, claim ledger, and exact source of the A18 vendor statement? How is registration-before-execution anchored independently of mutable repository state?
  3. Which direct-provider, model-matched or native-browser controls are available for A17/A18? Without such controls, how should these examples support practical claim auditing rather than characterize the gateway-conditioned systems alone?
Reviewer_fj58OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes the Evidentiary Control Protocol (ECP), a ten-stage process for making AI evaluations easier to check. It covers designing tests, recording the plan before running them, carrying out the tests, preserving evidence, reviewing what happened, and assigning outcomes. Its central idea is to keep four questions separate: Was the test run correctly? Is the evidence intact and trustworthy? What result follows from the scoring rules? What scientific conclusion does that result actually support? The authors demonstrate the process through small evaluations, deliberately inserted faults, corrected experiments, and tests of claims made by AI providers. They preserve failed and inconclusive results, including cases where problems in the test prevented a clean conclusion. The paper presents these examples as evidence that the process can be used.

Critical strengths
  1. The protocol distinguishes failures that a single score can hide.

  2. The paper preserves bad results and documents corrections.

Critical weaknesses
  1. The paper appropriately disclaims novelty for individual controls, but the comparative evidence does not isolate what ECP adds to preregistration, raw-output logging, explicit error categories, and independent review. M2 has six cases with agreement between the two labeling approaches and no designed divergence case; its nine positive-control fixtures check predefined comparator rules. The additional synthetic comparison exposes overlapping taxonomy labels. The strongest case for ECP is that enforcing these practices together can prevent silent repairs and unsupported interpretations. However, the submission does not demonstrate improved fault discovery, audit reproducibility, or reduced misleading conclusions against a baseline that already follows those practices, nor does it quantify the operational cost. This limits significance even when ECP is judged solely as an evidentiary-governance method, as intended.

  2. The structured-output audit, A17, concerns a provider claim about one model version but tests a later version through an intermediary service. Getting 60 successful results out of 60 tests does not independently confirm the original version-specific claim.

  3. The browser-related audit, A18, uses text simulations, including conversations combined into a single prompt. It also reports effects associated with the intermediary service. These conditions limit what can be concluded about the original browser setting.

  4. Scoring needs similar care. A planted test phrase appearing in an answer is not necessarily evidence that the model obeyed an attack: four reported hits occurred when the model quoted the phrase while refusing. The paper discloses this distinction, but its broader discussion should consistently respect it.

  5. Finally, public files and digital fingerprints can help check that records agree and files match the versions committed to. They cannot, on their own, prove that the plan was recorded before the experiment or let a reader trace raw responses to final verdicts. The original evidence remains under its owner's control, although confidential reviewer access is offered. That limits the independently checkable evidence for a paper centered on auditability.

Questions for authors
  1. Is there a comparison against an evaluation process that also records its plan in advance, saves raw outputs, and distinguishes execution failures from task failures? What important error or mistaken conclusion does ECP catch that this comparison misses, and how much extra work does ECP require?

  2. Can reviewers inspect a complete anonymized example from input to final verdict, including raw requests and responses, scoring rules, evidence of when the plan was recorded, and who independently checked it? Which checks can anyone perform with the public files, and which require private access? Is there an independent record showing that the commitments existed before execution?

Reviewer_hwgBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

A proposed protocol for making AI evaluations auditable. Rather than reporting simple benchmarking scores, it locks the test cases and grading rules before execution, recording a cryptographic fingerprint of each file so any later changes remain identifiable, then runs each case under controlled conditions, saves raw output, classifies outcomes mechanically, and keeps execution validity, evidence integrity, recorded outcome, and interpretation separate. The paper shows how this works with pilots and audits of vendor claims (OpenAI structured output claim worked, while an Anthropic prompt-injection claim under testing conditions gave results that were the opposite of the vendor's reported claim.

Critical strengths

There is transparency about failures and bugs.

Reproducibility with hash-locked ledger. Independent audibility of vendor black box claims.

Separates execution validity, evidence integrity, protocol-defined outcome, and scientific interpretation

Shows a way forward for evaluating a vendor claims

Critical weaknesses
  1. Repeatedly there is boilerplate text that states: this section is reproduced verbatim from the pre-expansion main text of the integrated submission draft (snapshot 91f698a, tag paper-integrated-2026-09-24); it was relocated to the appendix under execution order D1-C with no change to any scientific statement, value, classification, or interpretation. This is out of place for an human reader in an academic paper. It makes it seem like an agent wrote this paper and off to be a human reading through this.

  2. There are very small sample sizes here. This means statistical generalizability will not be what we would like to see.

  3. There is a lot of operational work here: 10 steps, multiple failed runs etc

  4. These were not direct APIs to vendors. Issues like a "”Kiro” persona observed in 7/60 responses"

  5. There is one auditor: the process managed by sole auditor

Questions for authors

Was there a pre-registered power analysis or sample size calculation? if so please explain.

How avoid having a single auditor bias subjectively affecting the outcome?

Have you tested whether a direct API approach has similar results?

How is this practical for large-scale benchmarks?

How avoid frequent revisions after the fact? Are these a concern?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.