Per-value evidence adoption checks

The operator adopted option 2 and authorized a commit on 2026-09-26. This is reflective learning from the pilot's inability to count an individually wired mechanism when another value was only afforded.

Contract and consumers

The authoritative declaration is the main-result type, enforced by systems_matrix.validate_comparison; the memory-report schema describes the same shape. Both require every registered axis and exact evidence keys for its values. Specialists receive the rule through their instruction and report type; coordinators receive it through the main skill and result type. These instructions are live for the next worker loading them.

Consumers inspected: validators and schema; main and memory authors; matrix and table emitters; numerical analyzer; landscape bundle and synthesis; memory classification audit; unit and publication fixtures; comparison-directory operating documentation. The audit instruction also taught the weakest-basis rule and was updated. Source identity, result paths, canonical namespaces and run identity conventions do not change. No relocation or new derived population predicate is introduced. Strong-positive and complete profile predicates are shared functions in the matrix reader.

CSV retains fourteen value arrays and changes each axis's _basis column to _evidence, a JSON map. The current reader rejects old shapes; no compatibility conversion guesses evidence from prior prose. Complete-profile statistics keep their conservative meaning and gain separate supported-positive counts.

Pinned and historical inputs

No old result, memory report, source pin, bundle or hash is rewritten. The existing .commonplace-validation-ignore mechanism marks the retained snapshot subtree so collection-wide sweeps do not judge historical bytes under the latest schema. Publication and comparison readers continue explicit current-contract validation; the marker does not admit old shapes. Initial pilot runs ending -01, the first query ledger and synthesis trial remain dated historical witnesses under their captured method. Replacement runs use -02 and fresh source-only coordinators and memory specialists. The current reader rejects historical axis-level evidence, so an old bundle requires its captured reader/method. Full-corpus outputs stay stale until their population is refreshed.

Install and projection checks

This is an analysis contract consumed from the editable checkout and installed library. No entry point, scaffold, init migration, root doctrine or skill stub identity changes. Fresh/existing scratch-project probes check that init routes to the live instruction rather than copying an obsolete contract. The source checkout itself never runs init.

Acceptance

Tests cover mixed strengths, partial positives, partial negatives, absent and unknown states, off-vocabulary values, unresolved or missing per-value evidence, old-shape rejection, doc-grounded separation, exact CSV/result agreement, publication guards, and immutable bundle verification. A complete profile is not reconstructed by filtering away weak values. Semantic support still requires source analysis; the three fresh pilots test that producing path.

Searches cover the old _basis field, weakest-basis prose, comparison literals, and all direct instruction callers/consumers. Remaining old forms in retained results, legacy reviews, initial workshop evidence and ignored frozen bundles are historical; unrelated uses of the word basis are not this contract.

Search and install results

Post-edit file-hit counts across src, scripts, tests, types, instructions, reference and root doctrine:

  • _basis: 0 files.
  • weakest (evidence |supporting )?basis: 2 files.
  • memory-comparison: 16 files.
  • analyse-agent-memory: 3 files.
  • synthesize-agent-memory-landscape: 2 files.

Bounded ignored run/cache search for _basis or scalar evidence bases: 86 files; these include frozen historical results and per-value bases. No old current reader field remains.

Fresh and repeated commonplace-init --root /tmp/commonplace-evidence-contract-tm3iw551 returned zero; --check returned zero. The installed-library pointer resolves to this editable checkout. This analysis skill is repository-local, not a promoted init stub; init changes no analysis output. Existing typed evidence remains untouched and requires regeneration to satisfy the current reader.

An existing scratch install containing the old Napkin result retained its exact c6be0f355181f46e0a53e1783145a0715b2bc136c8a0709661db74c2bb272c01 bytes after init. Current validation rejected the old comparison shape as expected (requires assessment, evidence, values, records, and note). Its three missing ontology-link warnings came from the intentionally minimal scratch tree.

Verification: all 850 tests passed; after expanding bundle capture to include the memory producer/report contract and numerical/table readers, all 31 matrix and bundle tests passed. Repository ruff check passed.

Source-only rerun recovery

Napkin run AAS-2026-09-26-napkin-02 was abandoned before publication. An unfiltered agent-status listing exposed its coordinator to an unrelated completed analysis. The skill requires a fresh coordinator after prior-analysis exposure, even when the earlier analysis concerns another system. The failed run state was validated and its specialist stopped. Replacement coordinators must avoid unfiltered agent listings; any necessary listing is restricted to their own subtree. No generated or retained publication was changed by the failed run.

Final checkout-wide searches (explicit . root) found zero files for weakest supporting basis, “one evidence basis”, or basis, and canonical-record columns. _basis remains in ten historical workshop/retained-evidence files and no live consumer. The comparison directory README was included in the final consumer sweep and validates cleanly.

First regenerated result check

Mem0 run AAS-2026-09-26-mem0-02 published result b4a541c5e6a90e37dbb0115bf9e230c41c5fc635023b519a57c447d4b7ffabd1. The matrix builder and numerical analyzer both accepted its retained evidence. Automatic writing and push each count once at wired basis; manual writing and pull remain afforded. Both complete-profile columns remain unfilled. This is the intended separation between supported membership and complete-profile statistics on source-grounded evidence.

A coordinator source diagnostic caught out-of-range citation endpoints before publication preparation. The coordinator corrected only those mechanical endpoints, disclosed the correction in the specialist report and reconciliation, refreshed its hash, and passed frozen-source checks. No substantive specialist finding or comparison value was changed by that fix.

Six real-reader negative probes passed after Mem0 regeneration: both matrix and table commands rejected a missing main review, a legacy-review path, and a duplicate source selection. Each exited 1 and preserved pre-existing output bytes. No published input was edited for these checks.