idkmesh

Open-issue evidence gates — 2026-08-30

Baseline revision: e16a8d0. Tool: tools/issue_evidence_gate.py.

Read this as a snapshot. The numbers below were measured at that revision and are re-derived on every run of the tool; the tool, not this page, is the current authority.

What was inspected, and why

Twenty-one issues were open when this audit was written; issue 152 has since been closed, leaving twenty. Several of them cannot be closed by writing code, because their own acceptance text names a precondition that is a measured property of the repository rather than a task: a held-out corpus that has to be collected, an independent reviewer who has to appear, a contributor who has to come back.

That distinction had been carried in triage prose. Prose rots: it cannot tell you when a blocker clears, and it cannot be checked by anyone who does not already trust the person who wrote it. This audit replaces the prose with a tool that reads evidence surfaces the repository already publishes.

The tool is deliberately narrow about its own authority:

It never decides that an issue is blocked. It evaluates preconditions that were each read off an issue’s own text, and it fails when a registered precondition has become met — that is, when the registry has gone stale and the issue can be worked again.

An unmet precondition is the expected steady state and is not an error.

Evidence surfaces

All three are produced by existing tooling; none is new to this audit.

surface what it supplies
randomness_lab.r1_readiness over benchmarks/*/cohort.json the preregistered fail-closed audit for a real R1 replay corpus
newest results/collaboration/observables-*.json independent-review latency, contributor recurrence, ownership concentration, evidence-derived priors
.github/workflows/*.yml workflows that store state inside an open issue body
examples/community/ace-activation-gate-current.example.json the ACE activation gate’s verified-descendant count, as a committed snapshot

Measured state at e16a8d0

precondition observed required met
eligible held-out work units 0 >= 20 no
independent reviews exist 0 (of 50 pull requests, 50 closed without review) >= 1 no
recurring-contributor trials 0 >= 1 no
distinct actors 1 (HHI 1.0) >= 2 no
evidence-derived strategy priors 0 >= 1 no
canonical node/ directory absent present no
no workflow stores state in an issue ace-cohort-observer.yml, ace-community-growth.yml none no
distinct actors for a worker fleet 1 >= 10 no
verified descendant evidence 0 >= 1 no

The corpus number is the one most likely to be misread. Four cohorts exist — phase-b2-first-five, phase-b2-first-five-v2, phase-b2-successor-five, phase-b2-successor-v2 — and each carries five tasks, so the tree is not empty. The attempts are concentrated rather than spread: phase-b2-successor-five holds all five recorded attempts and the other three cohorts hold none. The frozen audit still returns status: blocked and eligible_work_units: 0 for all four, because the recorded attempts are untrusted model patches with no independent verification result beside them. Attempts are not evidence.

These counts are no longer written by hand. tools/benchmark_publication.py derives them from the cohort files and --check fails when the committed benchmarks/publication.json drifts from the tree; an earlier revision of this paragraph claimed six attempts spread across all four cohorts, which the generated count contradicts.

Issues with a machine-checked precondition

issue precondition why that precondition
9 recurring contributors the title asks for the first ten recurring contributors
10 distinct actors a repository-driven community engine needs a community
11 node/ directory activation targets a canonical node that is not in the tree
23, 109 not a workflow state store their bodies are rewritten in place by ACE workflows
30, 70, 96 eligible held-out work units each names a real held-out coding corpus
138, 151, 167 independent review exists each asks for an independent inspection
152 independent review exists gated on issue 167
86 evidence-derived priors P0 item 4 only; items 2 and 5 are already met
1 distinct actors >= 10 its minimum experiment’s step 4 is to connect 10-20 heterogeneous worker nodes
4, 16 independent review exists every remaining box in both begins with PR 159’s required separate human review
57 verified descendant evidence Phase A shipped; the one unchecked activation-gate box needs a cohort with real descendants

The four external dependencies

Re-reading all twenty-one open issues on 2026-08-30 against 53ca25a found that every externally blocked one reduces to exactly one of four things:

  1. A genuinely separate human reviewer — 4, 11, 16, 138, 151, 167, and 152 by a criterion its owner added after the body was written. Repository-wide: 0 independent reviews across 50 pull requests. Issue 152 was closed on 2026-08-30 without that reviewer appearing, so it now sits in the evidence gate’s resolved_registrations rather than its live registry; the independent-review count is unchanged.
  2. Real external contributors — 9, 10, 57. distinct_actors: 1, HHI 1.0.
  3. A real held-out corpus, or compute for a larger producer — 1, 30, 70, 96.
  4. Owner-held credentials or the repository admin UI — 12, and the social-preview half of 10. No repository evidence can observe these, so 12 stays unclassified rather than being recorded on a guess.

Issues 23 and 109 are outside that taxonomy because they are not tasks at all: they are the ACE workflows’ storage. ace-cohort-observer.yml locates its ledger by scanning open issues for the ace:cohort-observer label and creates a new one when it finds none, so closing either forks a duplicate ledger at a new number on the next scheduled run.

Issue 86’s entry is marked partial_gate, because the issue is a programme rather than a single deliverable and only one of its items is gated this way.

Issues deliberately left unclassified

2, 12, 13, 22.

No mechanical precondition was read off their text, so none is recorded. That is not evidence that they are blocked, and a test asserts they stay out of the registry so a future guess cannot be laundered into a machine-checked status.

12 is blocked, but on owner-held credentials and a third-party hosted service, which no evidence surface in this repository can observe. Recording it would mean asserting a blocker the tool cannot check, which is the thing this registry exists to avoid.

The other three are actively being worked rather than waiting:

A partial gate is recorded for 4, 16, 57 and 86, because each has shipped deliverables and recording them as wholly blocked would misreport the repository’s own state.

When the evidence cannot be read

The readiness audit validates cohorts against the Phase 0 schemas, so it needs the requirements-phase0.txt dependencies. Some workflows deliberately run without them.

In that case the tool reports available: false, observed: null and met: false for the corpus precondition — “cannot tell”, never “measured zero”. The two are indistinguishable in a bare number and mean opposite things: one is an empty corpus, the other is an audit that never ran. A partially audited set is treated the same way, because one unreadable cohort could hide a ready one behind it. Two tests pin this, one of them by running the audit through an interpreter that genuinely cannot import the validator.

What this audit does not establish