OSS Reviewer Record Audit

This repository audits factual and inferential claims in the supplied transcript 20260806_oss_info_reviewer_ds_transcript.md. The transcript discusses the ICLR 2025 paper Disentangling Representations through Multi-task Learning, an anonymous OpenReview reviewer, and public work by Jack Lindsey.

The audit is designed to answer three different questions without collapsing them into one:

1. What does the paper actually establish?
2. What does the public record establish about Jack Lindsey's work and timeline?
3. Does the available evidence identify anonymous reviewer Mvb9 or establish the transcript's "hypocrisy" narrative?

Bottom line

• The paper's publication record and its main qualified technical results are supported by primary sources.
• A user-supplied 17-page printout of the public OpenReview forum documents the review record: Mvb9 assigned a 3: Reject rating, the area chair said they did not follow that reviewer, and the paper was accepted as a poster.
• Public records support the cited outline of Lindsey's education and research, and show genuine topical overlap with multi-task learning, representation geometry, and biologically plausible learning.
• No public evidence found connects anonymous reviewer Mvb9 to Jack Lindsey. OpenReview's own documentation says reviewer labels of this form are anonymous per-submission groups.
• The PDF removes uncertainty about the questioned review text, but the transcript's writing-style attribution is still not reproducible. No authenticated single-author, genre-matched comparison corpus, candidate controls, or calibrated open-set test was supplied.
• The strong "hypocrisy" or "smoking gun" framing is not supported. Lindsey's 2017 attention paper includes both guided and unsupervised pretraining, while his 2020 paper explicitly studies biologically motivated local learning.

Documents

• Findings: concise answer and reasoning.
• Claim ledger: claim-by-claim dispositions.
• Method: evidence rules, confidence labels, and scope.
• Stylometry assessment: why attribution was not run and what a defensible test would require.
• Access log: OpenReview access attempts and limitations.
• OpenReview PDF review: file provenance, page-level findings, and evidentiary effect.
• Source register: primary and methodological sources with stable URLs.
• Evidence explorer: responsive HTML/CSS/JS interface for searching claims, inspecting evidence, reading findings, and browsing sources.
• Machine-readable audit: the structured source used by the evidence explorer.

Evidence explorer

Open site/index.html directly in a browser. It does not require a server or network access for its interface; outbound evidence links point to the public sources recorded by the audit.

On the Brixton private site it is published through a symlink at:

http://10.0.0.30/artifacts/2026-08-06-oss-reviewer-record-audit/

The site data is generated from site/audit-data.json so the browser interface also works under file:// URLs:

node scripts/build-site-data.mjs

Input integrity

The supplied transcript is not copied into this repository because it contains substantial duplicated and corrupted text. Its SHA-256 at the time of review was:

6849c2d69a8ff1cc0aae60d972c8fb762e363f5655cef2d9f6949287654d8266

The supplied OpenReview PDF is likewise not copied into Git. It is a 17-page browser printout and had this SHA-256 at review time:

91d37f2473e6c5d2a8df45507cf01ad67a5007cc7764b53c0c3128101ac16043

The audit was performed on 2026-08-06. It is a public-record research audit, not a claim to reveal or circumvent confidential peer-review identities.

Validation

Run:

./scripts/check.sh

The repository intentionally has no Git remote.