Consumer migration queue

Inventory inspected on 2026-09-05. These entries distinguish direct review readers, readers of derived data, evidence-bearing publications, and machinery that currently requires the legacy output. Recheck each entry's live behavior when selecting it; a textual reference alone does not prove active use.

Entries without an execution record remain pending. The acceptance conditions below supplement the workshop's shared conditions.

ID	Consumer and current entry point	Migration boundary and acceptance
C01 — done	Matrix builder and reader read retained main results.	Read main-review results directly. Establish supported fields, system/run identity, evidence tier, population, and input hashes. Derive each classified value from a recorded finding; report missing support. Disposition prior hand-classified and identity-join columns explicitly.
C02 — done	Table renderer reads retained main results and links their public evidence.	Render the migrated data and link sufficient tracked main-review evidence. Verify link resolution and population against the originating result set. Depends on C01's data contract.
C03 — done	Matrix analyzer computes statistics directly from retained main results.	Update field assumptions, evidence population, and missing-value handling against C01. Verify a bounded query against the originating main-review findings.
C04 — done	Landscape synthesis skill reads full retained main results and bundles a matching derived matrix.	Read main-review files as qualitative evidence and bind quantitative results to that same population. Update the frozen evidence bundle and cited-file identities. Preserve evidence-tier distinctions and exclusion of transfer judgments. Requires the durable-evidence decision and C01 for matrix-based claims.
C05	Trace-learning survey cites individual legacy reviews.	Refresh from main-review findings with sufficient lineage, behavior-change, and evidence-status detail. Declare the new snapshot/population and distinguish unavailable evidence from a negative finding.
C06	Comparative review combines matrix claims and legacy qualitative examples.	Rebuild one coherent snapshot through the migrated synthesis procedure, or explicitly retire it as a current account. Do not update counts while retaining claims from a different population.
C07	Thalo type comparison uses a legacy review for its external-system account.	Ground the Thalo mechanism in the main review. Disposition the Commonplace mapping as current transfer or a bounded historical comparison under the appropriate contract.
C08 — done	Transfer scan now requires the exact main-review result and its completion state.	Initial and final completion/identity checks, canonical finding references, and no legacy or summary fallback. See acceptance.
C09 — done	External-system placement pass reads generated main reviews, with source ingests identified separately.	Check support, evidence status, source revision, and input hashes; withhold unsupported or local-only citations. See acceptance.
C10 — done	Classification audit reads retained main results and diagnoses their classifications.	Require exact evidence and method identities; preserve scope and uncertainty. Route defects through the producing method and source regeneration, without patching generated prose. See acceptance.
C11 — retired into publication	Quote-grounding now runs inside main-review analysis checks; the standalone live-checkout instruction is retired.	Every quote-anchored block in an exact result or generated candidate resolves against the recorded Git commit or immutable capture before publication. Occurrence remains distinct from semantic support. See acceptance.
C12 — done	Memory specialist produces a typed report from frozen sources for integration into the main result.	Retired the legacy writer and its skill entry points. Verify source-only analysis, proposal integration and report validation; see acceptance.
C13	Authored evidence links in kb/notes/, kb/articles/, kb/sources/, kb/reference/, and per-system analyses; grounding-alignment gate follows linked library evidence.	Work by citing artifact or bounded claim group. Check support before changing a target; retain historical citations when the original observation matters. The generic gate already recognizes both system collections: change it only if the selected evidence location requires it.
C14 — done	Main-review publication and handoff: analysis checks, publication code, validation, result/run-state contracts, and producing skill.	The operator authorized removing legacy production now. Remove mandatory legacy candidate/output fields, legacy publication gates, and stale-output reporting tied to the old pipeline. Preserve necessary evidence and publication checks on the main outputs; handle validation of retained historical artifacts explicitly. Prove a memory-system run can complete with one integrated main result and its local specialist report, without legacy publication or compatibility paths.
C15	Collection READMEs, generated indexes/table links, and website/search navigation.	Route current discovery to the main reviews and migrated comparison outputs. Keep supported historical references navigable. Verify public evidence from a clean checkout without ignored run files.
C16	Nearest-constructions supplement grades eighteen constructions against the witness conditions from mixed-vintage reviews and source ingests.	Regenerated rows cite main-review findings for parametric fixity, material revision admission, decision roles, operating modes and answer-oracle access (required as evidenced prose by the producer). Paper-only rows stay on ingests or move to doc-grounded runs. Normalizing these facts for a rebuildable table is the autonomy-comparison proposal, not this queue.

Inventory boundaries

No additional specialized numerical pipeline was identified beyond the matrix builder, renderer, and analyzer. Historical relocation/migration scripts, old proposals, and recorded reports are not automatically active consumers.

Review-to-review links and workshop references also need disposition when their targets migrate. They are navigation or historical evidence unless inspection shows a live reading procedure. Source ingests and historical comparisons must not be rewritten merely to eliminate occurrences of the old collection path.

Per-consumer record

When starting an entry, add a subsection here with:

• ID, status, intended behavior, and owned files;
• exact main-review inputs and the findings/fields consumed;
• prerequisite or blocker, with its return-to-decision condition;
• bounded verification result and remaining coverage limits;
• final disposition and commit reference when available.

Use these records to choose the next step. Persist detailed change history in the implementing commits rather than copying their diffs into this queue.

C08 — Transfer scan

Status: done. Selected first because its direct caller already passes the main review's exact result, and the scan needs no matrix fields or public retention policy.

Owned files: kb/instructions/scan-agentic-system-transfer/SKILL.md, its direct caller kb/instructions/analyse-agentic-system/SKILL.md, and this workshop's acceptance records. The runtime loads the scan through its skill projection or the main review's step 9 invocation; the canonical instruction edits change both paths immediately.

The input is one exact result.md plus its sibling completion state, source identity/revision, result hash, interest brief, and permitted Commonplace scope. The migrated scan verifies completion before interpretation and again before returning findings, then reads the result's shared records, lens findings, and limits. Legacy reviews and compact projections cannot substitute for that file.

The two recent minimal-state runs both fail today's completion checks:

• AAS-2026-09-04-pond-01: missing required legacy-review-model-partition.
• AAS-2026-09-04-apache-maka-02: missing that field (required even when null), plus a Run identity projection containing a — complete suffix where the validator requires the exact state path.

These are acceptance inputs for rejection, not successful live pilots. Do not patch their generated analyses or waive validation to make the trial pass. The bounded positive replay and rejection checks accept the input migration without claiming either production run now validates. All 699 tests passed; changed skills and the replay scan validated cleanly. Committed in c1490415.

C01–C03 — Matrix, table, and statistics

Status: done for the reader migration. Production pilot: one regenerated system. C01 established normalized assessments in the exact main result and retention of those same bytes during publication. C02 then replaced its CSV dependency with the same direct reader and public evidence links. C03 replaced its CSV input with that reader and separates stronger implementation/operation evidence from weaker bases and unknown assessments. Each procedure can run without the legacy corpus or either of the other comparison outputs.

Owned files: the three scripts and systems_matrix.py; main-result type/schema and validation; publication and completion checks; producer skill; website retention exception; affected command and collection documentation; tests; and these acceptance records. No generated system findings or legacy CSV/table bytes were edited. Prior hand-classified columns and directory-name identity joins are removed, without a compatibility adapter.

Inputs are selected public main-review projections plus their hash-identified retained results. Scope, evidence tier, 14 assessed axes, basis, and canonical record references come directly from the main result. Multiple stores stay sets. Missing normalization or retained evidence blocks the selected cohort; no tag omission, legacy prose, or old CSV value fills a gap. A repeated source identity requires explicit selection of one review.

The acceptance record covers clean-input builds, missing and mismatched evidence, unknown/absence distinctions, evidence-tier filtering, input drift, retention, and publication rollback. The initial live build rejected Apache Maka for missing retained-result metadata. Implemented in c1490415. The production pilot now rebuilds Maka through the main workflow and verifies the matrix, table and analyzer on its retained result. The default all-generated build still rejects Pond. The workshop pair covers Maka alone; public defaults and legacy snapshots remain unchanged.

The later producer-check exercise found unsupported task-horizon and inconsistent learning-branch assessments in Maka. The reader/output mechanics remain accepted, but affected complete learning claims must be withheld pending source regeneration. The pilot does not certify every comparison classification.

C04 — Landscape synthesis

Status: done for procedure migration and one-system production pilot. The skill now requires full retained main results for qualitative claims and builds its CSV from those same inputs. The reusable bundle command captures contracts and hashes, rejects drift and mixed populations, and distinguishes current checks from historical verification. Explicit unknown assessments withhold unsupported conclusions without blocking unrelated findings.

Owned files: the synthesis skill, scripts/bundle_agentic_landscape.py, its tests, the comparison README, legacy collection navigation and design note, and the two public-output collection contracts permitting citations to published exact main results. Invocation through the existing skill projections loads the changed procedure immediately. No main-review findings were edited.

The acceptance record and bounded fixture synthesis cover a four-result population, one eligible quantitative case, three explicit exclusions, canonical-record support, and rejected stronger claims. The procedure migration passed 722 tests and was committed in c589b981. The subsequent production pilot uses main run AAS-2026-09-05-apache-maka-01 and its retained result directly for bounded queries and qualitative findings. It also fixes link validation for results supplied in the same prospective publication bundle. All 723 tests pass. No public landscape article or whole-corpus comparison was refreshed. Pilot and publication fix committed in f586fb1d.

Next: C05 needs corrected regenerated trace-learning evidence before its survey can move. C06 and C07 likewise need enough regenerated evidence or an explicit historical disposition. Pond remains outside the new comparison population. Consumer IDs remain an inventory, not a fixed sequence.

C15 received navigation and public-citation contract updates needed by C01–C04. Broader citation and public-navigation migration remains pending. C12/C14 now use the operator-authorized specialist handoff and have retired duplicate drafting; see the acceptance record below.

C09 — External-system placement

Status: done. Selected after C08 because a tracked public main review can directly support a bounded placement without solving matrix normalization or exact-result retention first.

Owned files: kb/instructions/simplification-passes/place-external-systems.md, its README.md catalog, and this workshop's acceptance record. The revise-an-article-or-note caller dispatches the named pass as an artifact, purpose, preserved-claims, and write-scope packet; that interface is unchanged.

The reader now identifies generated main-review metadata and evidence basis, reads supporting findings directly, and records hashes and source revisions. An exact result is a checked escalation when a compact review is insufficient. Unsupported claims are withheld; ignored local evidence cannot silently become a tracked library citation. Source ingests remain an explicitly separate class.

The Pond placement trial used only the tracked generated review. It retained implemented storage/retrieval findings, withheld automatic injection and demonstrated-improvement claims, and preserved the distinction between static wiring and observed use. Changed Markdown validated cleanly. Committed in c1490415.

C10 — Main-review classification audit

Status: done for procedure migration and a bounded audit. The existing instruction path now diagnoses classifications in retained main results. It checks evidence and method identities before and after interpretation, distinguishes structural validity from semantic support, and preserves unknown assessments. Corrections go through the shared method when needed and a new source-based main-review run; the audit cannot patch generated findings or refresh surveys.

Owned files: kb/instructions/refresh-agent-memory-review-taxonomy.md, this queue and acceptance record. Repository search found no automated caller or skill projection. Two mentions in other workshop inventories are descriptive, not procedure calls. The conditional producing skill already accepts the required source, target and diagnosis; its interface is unchanged.

The audit consumed the public Maka review and exact retained result from AAS-2026-09-05-apache-maka-01, including scoped comparison fields, canonical objects, routes, authority records and limits. No classification defect was established in that selected result. Missing or mismatched evidence, a legacy path, input drift and unregenerated Pond were rejected. Two temporary semantic defects passed structural validation and were identified by reading their supporting records. An isolated copy worked without legacy files or local run state. Input and method identities matched at the final recheck.

The changed procedure and workshop records validate cleanly. This Markdown change exercised the existing reader directly; no code changed or new pytest run was needed. No production source regeneration or generated-output edit occurred in C10. Initial acceptance covers Maka alone; the all-generated audit still blocks on Pond. The subsequent oh-my-pi report audit identifies two real classification defects after structural success and one remaining selection-mapping question. Corrections require a new producer run; the diagnostic does not invalidate the row automatically in the reader. The producer-check update adds the focused integrated check and shared clarifications before that rerun. It preserves the existing publication interface and passed all 723 tests. Its independent exercise reproduced oh-my-pi's defects and identified two further learning assessment gaps in Maka, while preserving valid uncertainty and weaker bases. Implementation and verification are recorded in this workshop's Git history.

C11 — Quote-grounding check

Status: retired into main-review publication. The standalone instruction had no automated caller, freshness baseline, or skill projection. Its required input was a live worktree, while the main workflow already owns the exact result, generated candidates, frozen source identity, and prospective complete state at one publication boundary. Repository inspection found no quote-anchored blocks in the current generated Maka or oh-my-pi outputs. The unmarked Semantic Engine review and several historical legacy reviews contain them, but they are retained artifacts rather than inputs to this migration.

Owned files: main-review analysis checks and tests, the producer and result contracts, the standing shape validator's explanation, the legacy review type and quote-anchor ADR, the removed instruction, its public redirect, this queue, the earlier contract snapshot's retirement notice, and the acceptance record. No generated review content or historical citation was changed.

The prospective bundle verifier now extracts every blockquote ending in a > --- attribution from the exact result, compact review candidate, and conditional legacy candidate. For Git sources it requires a matching repository and full revision, reads the attributed blob with git cat-file, and compares normalized quote text. The worktree and current HEAD cannot supply evidence. For immutable captures it compares against the captured bytes whose SHA-256 is already part of run state. Any failure blocks both prepare and publish before public writes. Standing validation still checks retained shape, and semantic gates still decide whether an occurring quote supports a claim.

Focused checks cover both main outputs, local and GitHub attributions, changed worktree content, wrong revisions, capture resolution, and candidate rejection inside prepare. The implementation and documentation passed the full test and validation checks recorded in the acceptance file. Historical reviews were not retroactively attested. C11 adds no consumer-facing secondary analysis and no legacy-format adapter.

C12/C14 — Memory specialist and legacy production retirement

Status: done for implementation and bounded acceptance; production rerun pending.

The operator authorized delegating memory analysis into a typed report and removing all legacy paths without compatibility. This supersedes C14's earlier ordering prerequisite. Pending historical survey and citation migrations remain pending; they do not justify keeping the legacy producer callable.

The specialist owns source-native memory findings and proposed comparison classifications. The coordinator assigns canonical IDs, resolves shared-record conflicts and incorporates the report into one self-contained main result. Only that result supplies public reviews, comparisons and later consumers. Local input and report files support completion verification, without becoming public evidence dependencies. Historical results are not silently upgraded to the new completion contract.

See memory specialist acceptance for the bounded source trial, validation checks and remaining coverage limits. The rerun acceptance covers metadata isolation, exposure recovery and replacement of unchanged uncommitted generated reviews. The Prime Agent follow-up covers copyable source initialization, required quoted evidence and complete identifier checks.

C16 — Autonomy evidence before normalization

Status: pending consumer migration; prose requirements implemented. The operator authorized finishing the prose requirements and deferring normalized comparison fields until several regenerated reviews demonstrate useful, consistent distinctions. Inspect materially distinct revision mechanisms, reuse specialist findings, and separate decision roles from answer-oracle access. This adds no autonomy grade, registry or generated-table support.

Next, exercise these requirements in source-based main-review runs and check whether their evidence supports the supplement's actual claims. Reconsider normalization only after comparing those records; no old-format fallback or compatibility path is authorized. The supplement and existing generated Maka outputs are unchanged by this step.