Per-value pilot acceptance
The pilot passes the revised comparison contract. Three fresh source-only analyses published successfully; their structured evidence produces a matrix, human-readable table, supported-value counts, complete-profile statistics, and an executable query ledger. This verifies the selected workflow, not the coverage of the 162-artifact legacy corpus. The inventory has three completed current replacements, 159 pending artifacts and four historical predecessors.
Frozen population
The three code-grounded results have source cutoff 2026-09-26. Completion and comparison checks finished on 2026-09-27. Source revisions match the first pilot, but classifications and boundaries were reconstructed afresh. Differences from the first trial are not a controlled measurement of the schema change alone.
| System | New run | Exact result SHA-256 |
|---|---|---|
| Dynamic Cheatsheet | AAS-2026-09-26-dynamic-cheatsheet-02 |
69cf228433d975ac518c02f315ac81aaa26a6b29c3e3db63b932e6a199194cfa |
| Mem0 | AAS-2026-09-26-mem0-02 |
b4a541c5e6a90e37dbb0115bf9e230c41c5fc635023b519a57c447d4b7ffabd1 |
| Napkin | AAS-2026-09-26-napkin-03 |
1c4b625211fa033130d7ea2533c214f5e5e791d07707ad96ec8aa34f9e835c53 |
The temporary trial outputs live under
kb/reports/cache/agentic-memory-refresh/2026-09-26-per-value/: bundle/,
memory-systems.csv, memory-systems-table.md, statistics.txt, and
query-output.jsonl. The bundle retains full result/review bytes, producing and
consuming contracts, reader code, dependencies and the ontology used by the
results. It is a workshop trial, not a durable public landscape publication.
Manifest SHA-256: 101754b023abbd2efd352e11c8cc75a234dc81802829c2c2d15199b05f8b3113.
Matrix SHA-256: dc82c6f2e5e9084214147782cf7361c64b20e2940139122e5ded38ad5a4d800d.
The repository revision captured before the commit is only contextual metadata; it does not itself retain the then-uncommitted input bytes.
Mechanical checks
- Matrix, table and numerical analyzer succeeded against the identical explicit three-review selection. The matrix has three rows and 69 columns.
- All 42 axis profiles matched retained frontmatter: values, coverage, every per-value basis/record/rationale and axis records. The table and statistics retained all six review/result hashes. Standalone and bundled CSV bytes match.
- The query ledger re-executed from frozen bytes with identical output. Each source counts at most once per query; exclusions disclose coverage, tier and individual value evidence.
- Both output writers rejected missing, legacy and duplicate-source inputs in six probes while preserving existing output bytes. No published input was altered for these tests.
- All 850 tests passed; the subsequent bundle-input expansion passed all 31 affected matrix/bundle tests. Repository ruff and scoped Markdown validation passed. Fresh/repeated scratch installation and old-evidence preservation checks passed; current explicit validation rejects the historical shape.
- Every replacement exercised guarded replacement of an uncommitted generated incumbent. Its old exact result remains unchanged. Publication verified frozen sources and the report/input/result identities before completion.
What the counts establish
The automatic-writing-and-push query now has two supported matches in the selected population of three: Mem0 and Dynamic Cheatsheet. This is a lower-bound support count, not a claim that the third system lacks either mechanism, nor that both mechanisms always occur on one route.
Mem0's automatic writing and internal push count at wired basis while manual writing and application pull remain afforded. Those mixed complete profiles stay excluded from strong complete-profile statistics. Dynamic Cheatsheet also supplies strong positives under partial coverage; opaque provider state does not erase inspected local mechanisms. Its partial profiles stay excluded from complete-set distributions even where each listed member has strong evidence.
All three faithfulness-tested assessments are not-determinable. None becomes a negative or a demonstrated benefit. Entropy remains available for eligible complete sets. With three rows, the analyzer's minimum of five paired rows prevents an informative redundancy result; its zero fallback is not evidence of independence. Document-grounded separation is fixture-tested, not yet tested on a real document-only member of this pilot.
The new synthesis trial uses the same frozen population and full canonical records. It records its own query and semantic checks. The first acceptance record, first ledger and first synthesis remain historical. Full-corpus public comparisons and earlier public syntheses remain stale/historical; this bounded trial does not replace them.
Next batch
The evidence-granularity blocker is resolved by ADR 093. Proceed with bounded refresh batches, retaining source-only coordinators and fresh memory specialists. Include a document-grounded source, a source with multiple legacy boundaries, and enough code-grounded cases to test eligible five-pair statistics. Do not merge duplicate source identities automatically. The remaining 159 artifacts have not been executed.
The contract-change record retains the consumer checks, scheduling recovery and the abandoned Napkin -02 isolation failure. The clean -03 run supplies the published Napkin evidence; no failed draft was promoted.