idkmesh

Conversation Record — ACE Hardening Convergence

Date: 2026-08-28

Why this consolidation happened

ACE’s own Growth Ledger entered CONSOLIDATE mode with review load far above its carrying-capacity target. During review of the open ACE stack, two safety PRs were found to be modifying the same privileged workflow independently:

Keeping both as long-lived conflicting PRs would increase exactly the kind of review/integration burden that ACE is designed to suppress.

The user asked the project to continue. The correct interpretation of “continue” under zero carrying capacity was therefore consolidate existing work, not create another independent growth mechanism.

Decision

Create one fresh convergence branch from current main that preserves both safety models and adds explicit regression tests.

The converged workflow combines:

From the metadata-security path

From the protected-integration path

Regression contract

tests/test_ace_workflow_hardening.py statically checks the privileged workflow for the most important invariants:

The explicit no-shell-execution check was ported from the later #51 contract work into this convergence branch so PR #98 remains the single canonical safety review surface.

The path-scoped workflow .github/workflows/ace-workflow-hardening-check.yml runs these checks with Python 3.11 and 3.13 and has contents: read only.

Governance conclusion

Repository files and agent instructions cannot enforce integration safety by themselves.

The canonical invariant is:

No autonomous actor may propose, approve, and merge the same protected change by itself.

The repository therefore treats actual GitHub protection/ruleset state as external evidence, not as a documentation preference.

Relationship to ACE Phase B

This convergence does not activate Phase B.

Even after this hardening is reviewed, stronger community actuation remains blocked by the broader activation gate work:

observer accepted
AND lineage accepted
AND security accepted
AND controller accepted
AND protected integration enforced
AND independently verified descendant evidence
AND healthy/fresh review capacity
AND bounded write budget
AND forbidden capabilities disabled

At the time of this convergence, public branch metadata still reported main as unprotected and the ACE Growth Ledger reported near-zero capacity. Therefore the correct operating posture remained SHADOW / CONSOLIDATE.

Empirical community state checked during continuation

The live ACE Growth Ledger later showed:

mode = CONSOLIDATE
review_load ~= 45.55
K = 8 (current controller hypothesis)
capacity ~= 0

with 173 observed raw events. This is a direct example of the anti-Goodhart rule: high repository activity is not evidence that the project should create another community generation.

Bootstrap Growth Seeds #24–#28 were inspected. Their visible activity remained predominantly repository-owner driven; no inspected comment on #25/#26 represented an external contributor. Therefore Cohort 2 should remain gated on real external/verified descendant evidence rather than raw PR/commit velocity.

PR cleanup

PR #62 was closed unmerged after its threat model and metadata hardening were preserved in this convergence path.

PR #51 temporarily accumulated an additional equivalent safety contract during concurrent work. Its only missing invariant relative to PR #98—the explicit ban on a shell run: step in the privileged workflow—was ported into tests/test_ace_workflow_hardening.py here.

Therefore PR #98 should remain the canonical safety integration surface and PR #51 can be closed as superseded without losing a safety property.

This reduces the open ACE review surface while retaining the stronger combined safety contract.