Backlog (unscheduled): ML conference disclosure policy targets (ICML first; COLM, AISTATS, KDD candidates), evaluated 2026-09-15 after arXiv:2609.07713

Backlog, not scheduled. This issue records an evaluated-and-deferred item so the reasoning is findable later. It becomes actionable only when a user asks for one of these venues (comment here or open a linked issue). No one should pick it up on the strength of this issue alone.

What was evaluated

Adding ML conference AI-usage policy targets to academic-paper/references/venue_disclosure_policies.md: ICML (the only one of the three largest ML venues not in the database; ICLR and NeurIPS are), then COLM, AISTATS, KDD.

Where it came from

Wang, Li et al. (2026), The Emerging AI Paper-Review Arms Race (arXiv:2609.07713), §6.3, lists these venues' 2025–2026 AI policies as examples of how much venue rules differ (AISTATS: author-side use under human responsibility, reviewer-side LLM use prohibited; COLM: disclosure of substantive AI assistance by authors and reviewers; KDD: review disclosure plus limits on sending manuscript text to external services; ICML: an LLM-review policy audit). The survey presents them as evidence of policy variation, not as a recommendation to tool vendors. The dual read of that paper on 2026-09-15 produced #871, #872, and PR #873; this item was the weakest of the candidates and was deferred.

Why it was deferred (2026-09-15)

1. No observed demand. Across the repository's external issues and PRs to date, no one has asked for an ML venue policy target. The only venue-database expansion so far was the medical batch (#596), contributed complete by its author.
2. The unknown-venue path already fails safe. disclosure_mode_protocol.md halts and asks the author to supply the policy; it never fabricates a template (pinned by scripts/test_venue_disclosure_contract.py). The cost of absence is one manual step, not a wrong output.
3. Maintenance is per year. Conference policies change with each CFP; every entry adds an annual re-verification.
4. Some of the four may not qualify. AISTATS and KDD rules cited by the survey are largely reviewer-side. Under the database's own procedure, a venue with no author-facing disclosure requirement is recorded as "no explicit AI-usage policy found as of {date}" and is not added to the runnable selector set. The batch may reduce to ICML and COLM.

If it becomes actionable

Follow "Adding a new venue (v2 and beyond)" in venue_disclosure_policies.md and the #596 / PR #599 precedent: official policy page only (never a third-party summary), access date recorded, alphabetical insertion, selector aliases + applicability predicates + fact-ledger rows in disclosure_mode_protocol.md, and every count surface bumped (scripts/check_venue_disclosure_policies.py hard-codes 15 entries; the tests, commands/ars-disclosure.md, and the README selector counts repeat it). No policy-anchor changes, no submission-formatting rules, no detection-threshold numbers.

Related

• #871, #872 (the two measurements the same paper motivated; open)
• PR #873 (the fifth human-in-the-loop anchor; merged)
• #596 / PR #599 (the medical-venue precedent)
• The database's "Education/QA journals" deferral note is a separate, equally unscheduled item; neither batch is ordered ahead of the other until a user asks.