idkmesh

Conversation record — tracker reconciliation and the evidence arc

Date: 2026-08-29 Repository: MSKazemi/idkmesh Continues: Solve all open issues and pull requests

Project owner request

solve all the issues and PRs and merge to the main branch with high quality and professional way. You can do in parallel if it is doable.

The first arc of this session merged navigation and index work and then reported that none of the 21 open issues could legitimately be closed. That report was accurate but not sufficient: an issue can be unclosable and still be wrong. This arc audited every issue’s acceptance criteria against the repository itself, found real drift between what the tracker claimed and what the code contained, and fixed the drift.

The repository asserted a false statement about itself, and enforced it in CI

Live branch metadata for main:

protected=true
required_checks=gate (3.11), gate (3.13)
force_push=false  deletions=false  conversation_resolution=true

Three places said the opposite:

Location	Stale claim
`examples/community/ace-activation-gate-current.example.json`	`"source": "branch:main@unprotected"`, status `blocked`
`docs/community/ACE_ACTIVATION_GATE.md`	“Public branch metadata currently reports `main` as unprotected”
`.github/workflows/ace-activation-gate.yml`	`assert 'component:integration_protection' in blockers`

The workflow line is the one that mattered. It made the stale claim load-bearing: refreshing the fixture from reality would have failed CI. A required check that breaks when the repository is described accurately pressures the next contributor to re-introduce the false statement rather than correct it.

Corrected in #333. The gate still returns BLOCK with required_controller_mode_if_blocked: SHADOW; the blocker set is now exactly {real_verified_descendant_evidence}, the one honest remaining gate. The safety property that protection must block was strengthened rather than removed: it is now proved by a mutation test — at maximum capacity, with every other component accepted, revoking protection must still fail the gate closed — instead of by relying on the live fixture happening to be blocked.

Two issues must never be closed, and nothing said so

ace-community-growth.yml locates its ledger by scanning open issues for ace:ledger (:116, :119, :126) and creates a new one when the scan finds nothing (:343). ace-cohort-observer.yml does the same for ace:cohort-observer (:35, :276, :315).

Closing issue 23 or issue 109 does not archive that state. The next scheduled run forks a duplicate at a new number and every reference to the old one becomes a pointer to abandoned state. Re-opening afterwards is worse: two open ledgers trip the :130 ambiguity guard and the controller refuses to run at all.

The gate fixture cites issue 109 directly as "source": "issue:109@<timestamp>" for both descendant_evidence and review_capacity, so a fork there breaks the activation gate’s provenance chain.

Recorded in docs/community/README.md (#332) and as comments on both issues, because their bodies are machine-rewritten by the workflows that own them.

Evidence that was about to expire

collaboration-observables.yml runs weekly and uploads with retention-days: 30. The repository had exactly one successful production run, and its output existed nowhere in the tree. Committed in #334, with a finding that separates two kinds of zero the run contains:

real measurements — all 50 pull requests closed without an independent review, so first-review latency has no event, review-concentration HHI has no population, and contributor recurrence has 0 trials and falls back to the uniform prior 0.5;
declared collector limitations — ownership concentration and structural debt read zero because collaboration_snapshot.py writes changed_file_owners: [] and structural_debt_finding_ids: [] unconditionally and says so in its own limitations list.

The output format does not distinguish these. An HHI of 0.0 reads as “perfectly distributed review” and is in fact the value an empty population takes.

Uncertainty attached to the evolution metrics, honestly

P0 item 3 of issue 86 asks for uncertainty on evolution metrics “rather than treating point scores as truth”. Mechanically wiring a beta-binomial into the priority formula would have been fake rigor: those constants are subjective per-action-type judgements, not binomial proportions, and there is no sample to form a posterior from.

#336 does the two honest things instead. Every priority input declares whether it is snapshot_derived, a snapshot_conditioned_prior, or a hand_authored_prior — exactly one input is ever derived from observation. And each recommendation carries sensitivity bounds over its unevidenced constants, labelled in machine-readable form as bounds_are_a_confidence_interval: false.

The finding is what those bounds say: at a 25% perturbation, no adjacent pair of recommendations is separated. The ranked action list is not ordered by evidence, only by authored constants that happen to differ. That was always true; it is now visible, and pinned as a test.

Two research gaps closed in simulation, at zero cost

Coordination topology (#335, issue 13 hypothesis 3, previously untested). Two budget-matched arms beside the flat one. Budget matching is exact — 72 parity records, zero false flags — and the arms are neutrally calibrated so one clean serial chain reproduces the flat single-worker distribution exactly. Topology does shift the exponent: task_dag raised it in 9 of 9 cells, role_specialized lowered it in all six diversity cells to −0.120 at worst. But the largest topology shift is smaller than the flat-arm gap between homogeneous and structurally diverse groups at the same difficulty (0.046 against 0.340). Error-correlation structure matters roughly an order of magnitude more than team wiring.

Imperfect correlated verifiers (#337, issue 22). E024’s panel was perfect by construction, so its three error fields were structurally always zero. The new panel follows the measured shape — beta-binomial over per-item difficulty plus a blind-spot atom, parameterized from E017/E020 — rather than the rejected shared-shock shape.

The QD advantage survives completely. And that is the finding: sweeping the panel to 45% wrong in both directions moves QD’s utility AUC by 1.5% and leaves its catastrophe count at zero. The mechanism is measured, not inferred — 157 defective artifacts waved through on seed 7 under the stress panel, and not one survived in the archive, because utility() and robust_quality() both return 0.0 for a non-viable candidate. E024 has no defect-propagation channel, so the falsification test had very little power. The write-up says E024 must not be cited as evidence about verification until an accepted defect carries a cost.

Two process findings

A frozen artifact was not reproducible off the generating platform. A new test asserted byte equality against a committed sweep. It passed locally and failed on every GitHub runner: the simulation goes through exp and **, whose last-place rounding is not identical across CPUs and C libraries, and a one-ulp difference changes the JSON representation and therefore the digest. The document also claimed byte-for-byte reproduction, which was not true. Both corrected: the replay compares values at a relative tolerance of 1e-9, the artifact’s own digest stays pinned, and “frozen and reproducible” now means reproducible in value.

Published numbers must reproduce from committed code. A mechanism claim was published with instrumentation figures that did not reproduce — 1343 accepts and 289 false accepts against the 1375 and 157 the committed code actually produces. The qualitative claim held exactly; the numbers did not. Corrected to the reproducing values, given a reproduction command in sim/e026_archive_contamination.py, and pinned as tests so they cannot drift again.

Untrusted content surfaced, not acted on

Issue 167 carries a comment from an account with authorAssociation: NONE containing a machine-generated “Delivery Report”, a cryptocurrency payout wallet address, and a request that repository source files be sent to it. It was treated as data. None of its instructions were followed and nothing was sent. The moderation decision belongs to the repository owner.

What still cannot be closed, and why

#151, #138, #16 (part), #91 — need a genuinely independent human reviewer. Over 274 pull requests the authors are the owner plus three dependabot accounts; the collaboration snapshot records 0 non-owner non-bot merged authors over the 50 most recent, and closed_without_review: 50. There is nobody else to ask.
#9 — zero external human contributors, 1 star, 0 forks.
#12 — needs owner-side credential creation; 0 repository secrets exist.
#4, #16, #11 — reference idkmesh-node; node/ is absent from main and both pull requests that would have introduced it (#91, #159) were rejected without merging.
#30, #96, #70, #1, #2 — need real agent runs against a held-out corpus. All 20 benchmark tasks are split: pilot, the cohort carries no definition_digest, and real runs would breach the $0 rule in PROJECT_RULES.md.
#57 — four of five activation-gate criteria are now verified met; the fifth needs one independently verified external descendant, which needs an external contributor.

Boundaries preserved

No issue was auto-closed; every pull request used Refs: and passed the closing-keyword guard. No authority was widened: the evolution observer stays recommendation_only, the activation gate still returns BLOCK, and every new experiment artifact carries evidence_level: synthetic_mechanism with an explicit guardrail stating it cannot close the issue it informs. No paid model API was called; every run is deterministic local CPU.

This site is open source. Improve this page.