Evaluation record: four-review pilot#

Mission run_request:a89d1522-b36e-4434-8ac8-0d75495803b2, approved by the author on 2026-11-04. Accountable circle Evaluation, Continuity contributing. One Claude Code session, session:a7e4cb74-aaae-43ac-81de-4963ad6a25b2, with no delegation. It changed no code, opened no issue or PR, deployed nothing, wrote to no review, minted no episode and started no agent. Its only write is one commit of this folder to the private shelf repository.

Code prefixes in this file: CM Contribution Map candidate, RC receipt, I inaccessible item, X contradiction, S self-review finding, G quality gate.

Return#

What already exists#

• The reading that transcript-first needs is merged, tested and live as a contract, with its provider not configured (CM-07).
• The Codex guard is closed and running (jgwill/Miadi#926). Revising a pending mission (jgwill/Miadi#919) and the approval race fix (jgwill/Miadi#924) are live with their issues open.
• The decomposition chain is released and its verifier passes. It has not reached Factory dispatch (CM-10).
• The four Knowledge Focus Ethics criteria are in the schema and tested. Nothing calls them (CM-08).
• Concordia's page is read, and Episode 570 (miadi-chronicle://570) already holds twelve findings on the proposal material (CM-03).

What may deserve attention#

Five rows hold all four criteria, so they are the only rows that ask the author anything now: CM-01, CM-03, CM-05, CM-06, CM-11. The three suggested for discussion are in the section "Three candidates to discuss".

What needs verification#

CM-04, CM-09 and CM-12. Each needs a stretch of a recording read against receipts before any question reaches the author. CM-07, CM-08 and CM-10 are stated from receipts and wait on existing issues.

What can simply be kept#

CM-02 (candidate beats). Also, kept out of the rows with a stated reason: the Scrivener and Dropbox wish, Volume Zero, Nature-Inspired Algorithms and immutable footprints, which all sit under the jgwill/Miadi#871 table and jgwill/Miadi#872.

Proposed review flow#

No interface is proposed and no circle is created.

1. The author reads this Return, then contribution-map.md.

2. For each row the author gives one mark, in any channel already in use, as one line: CM-03 correct: <their words>.

3. accept means the reading stands and its suggested action may be discussed. correct means the author's words replace the reading. defer means the row stays pending, with a date if the author gives one. drop means the row closes with its reason and stays in history.

4. CM-11's own twelve-row marks belong in the ceremony named on jgwill/Miadi#871, miadi-ceremony://0ef2f473-259b-468c-9662-a9632763658d.

5. An agent authorized for it records the marks verbatim in a next commit to this folder, and reruns the gates. creator_decision is fixed to pending in this version's schema, so that commit also widens the schema to the four marks.

6. No mark authorizes work. A row that needs work gets its own approved request.

Delivery location#

This folder sits on the inquiry shelf, prototypes/artefacts/ in the private repository miadisabelle/Etuaptmumk-RSM, and its .weave.yaml records the author as owner with visibility private. /api/research/reports/<folder> serves contribution-map.md to the owner only.

save_research was not used, because it opens an inquiry issue and the brief forbids new issues. The folder follows the same contract by hand and has no issue, so the connector ledger holds no research_artifact record for it. This mission's result carries the pointers.

While this was being built, the ledger showed three steering corrections recorded under the author's name, naming a docs-only branch on jgwill/Miadi. All three were refused at the door because this session was busy, so none reached it (RC11, jgwill/Miadi#930). This session read them and began to follow them. The author then said directly, in this session, that the shelf was the good choice and that the assistant writing those steers had not given good direction. The shelf stands, and no such branch was created.

The author also said, in this session, that this is the first mission of its kind, that the first directions were not clear enough by common view, and that research for the author belongs on the shelf. On issues, the author said not to treat opening one as a rule here, since this is not implementation. The folder therefore has no issue, and one can be added through the shelf's own door if the author wants it followed that way.

Frozen manifest#

Read on 2026-11-04 at 21:36:05Z with no token, by the commands below. Copies are in frozen/. The SHA-256 is of the Markdown bytes as served. Each pinned version equalled the review's latest version at that read.

Who wrote what. PI: Gemini v1 text, an assistant's survey (v2), author-supplied dialogue attributed to Apple AI (v3). PII: Gemini v1 text, an assistant's survey and dialogue (v2). NL: an assistant's full replacement (v4), with the Gemini transcript beside it. RR: Gemini v1 text, Apple prompt output, and the author's own notes (v2 to v17). All of it is interpretation, and none is proof of the footage or of software completion. Only the author's notes are the author's words.

Access and consent scope: all four reviews and videos are readable without sign-in (RC3, RC5). Those public pages are all that was read from them. Consent for reuse beyond them is not established here, see CM-05 and CM-06.

Exact read commands#

B=https://miadi-review-service.vercel.app/review
curl -sS "$B/<id>/raw?v=<n>&format=md" | sha256sum
curl -sS "$B/<id>/json?v=<n>"
python3 ~/.claude/skills/miadi-review/scripts/miadi_review.py transcript <id>
python3 ~/.claude/skills/miadi-review/scripts/miadi_review.py status <id> [...]
gh issue view <n> -R <owner/repo> --json state,title,updatedAt,comments,body
curl -sS http://127.0.0.1:3335/api/factory/review-workflow
curl -sS "http://127.0.0.1:8040/api/edges?limit=all"
curl -sS "http://127.0.0.1:8040/api/ceremonies?subject_id=review:<id>&limit=all"
curl -sS "http://127.0.0.1:8040/api/narrative/beats?limit=all"
curl -sS https://mcp.sanctuaireagentique.com/health

Connector reads through get_capabilities, mission_status, handoff_status, list_records, list_circles, get_circle and get_context.

What was inaccessible#

I1 The recordings#

No video was watched. Every timestamp in the map is copied from a review, generated and unverified.

I2 YouTube captions#

Three watch pages list an ASR English track. Requesting it returned 0 bytes, in both json3 and default formats. PII lists no track. No caption text was retrieved. This session did not call Gemini, which has a cost.

I3 Conversations outside the factory#

ChatGPT's Chimera conversation, including its four-part definition at RR 1:02:00 to 1:04:00, and the Apple outputs before their paste into the review. No Perplexity output was read.

I4 Media#

The ElevenLabs and Edge-TTS audio was not heard. The NotebookLM studio was not opened. Episode 570's README, its manifest and its list of sources were read instead.

I5 One citation page#

mitpress.mit.edu answered 403 to this client. The Dourish reference is neither confirmed nor refuted here.

I6 Session tools#

The kit's inquiry-weave, medicine-wheel and voice MCP servers did not connect in this session, so the wheel was read over HTTP. The claude.ai connector sign-in has scope miadi:read only, so this mission's result went through the local connector door with the writer token.

I7 The connector's GitHub reach#

The connector reports no_github_token, so mission_status shows issues and commit outputs as unverified, including this mission's. gh was used here instead.

Method#

A Freeze#

Done. See the manifest above.

B Atomize#

79 atoms, listed in contribution-map.json under atoms. Each has a heading, a quotation of at most 25 words and a status. 16 were kept out of the rows, each with its reason.

C Reconcile#

Done against issues, code, ledger, wheel and chronicle. See the next section. No issue was created.

D Attention#

Each row answers who it serves, consent, epistemic worthiness, intentional attention and cost, in words. The four criteria keep their ids from packages/episodic-memory-schema/src/attention.ts, including output-optimisation-excluded. The criteria were written to test attention items. Applying them to candidate rows is this session's choice. No number is used anywhere in a row except the review version.

E Proof#

Each row has a reliability proof and a separate agency proof, an accountable circle and a decision gate.

F Self-review#

See "Adversarial self-review".

G Stop#

This session stops here.

Reconciliation of the named issues#

State words as in the map. On a fact that is not software, tested means read at its source in this session.

Also read: jgwill/Miadi#919 open, live since 0.1.15, one named item left. jgwill/Miadi#924 open, fixed in 0.1.19, session_completion_proposal:8adcc30f waits on a person. jgwill/Miadi#692 open, the audit under run_request:f67fe890 waits on the author's reading of NL v4. jgwill/Miadi#930 was filed by the author during this session.

One finding outside the rows: the frozen JSON of all four latest versions has 0 aspects and 0 results in its structured data (RC12). The audit under run_request:f67fe890 explains why: the import door writes an empty data stub. The Markdown carries everything these four versions say.

Duplication check#

No earlier exact-source map exists. A search for contribution-map and Contribution Map over the factory, the connector ledger, .mia, the mia workspace and the chronicle staging folder finds only this mission's brief and record. The four review ids appear together only there (RC10).

Partial overlaps were reused and not redone: the audit .mia/review-enrichment-audit-261103/ (NL), the jgwill/Miadi#871 table (RR), Episode 570's second reading (the proposal), threads.yaml in miadi-chronicle://563 (attention and axiology), and _staging_for_new_episodes/stcmastery-plugin-structure-atelier-jerry-near-the-loop-261104/ which cites NL.

No row suggests a new mission.

Contradiction log#

Each side is kept. None is resolved here, unless a receipt is named.

Citations#

The references in the surveys exist. Whether a field applies to this work is a reading, and none of these works was read in this session.

No fabricated citation was found. The Concordia wording confirms what PI's survey says on five pages and on suggested elements.

Adversarial self-review#

The same session wrote and reviewed this, so it is not independent of its author. The brief forbids delegating, so no second agent read it. Treat it as a checklist that was run.

S1 Duplicate issues#

None proposed. Every row points to its existing home above. CM-07, CM-08 and CM-10 suggest work on existing issues only.

S2 Unsupported source claims#

Claims from the reviews are labelled, and their contradictions are logged above. Weak spots: "implemented" for the voice layer rests on a session's status file and its commits, since the mission has no submitted result. "Tests pass" reads the current dist, since a rebuild was avoided in a shared checkout.

S3 Fabricated citations#

See "Citations". All quotations are checked by G06 against the frozen text. No page text from the Wilson folder is copied here, as its README requires.

S4 Access problems#

I1 to I7. This pilot cannot say what the recordings show. A row that needs them is marked verify_source.

S5 Claim versus receipt#

A draft sentence that the transcript never names Wilson was wrong: a check found one mention at 28:15, and the row was corrected. "No item carries focus" was checked against dates, and 4 of 36 items were asked after the fields existed, so the absence means something but rests on few items. A connector or test receipt shows a mechanism only, so each row's agency proof stands apart.

S6 Missing consent#

CM-06 names the open matter. This file refers to the named person by role where it can. The name appears only in the quotation in the map and its JSON, in a private repository, and it is already on the public page being asked about. If the author prefers it removed here, that is a one-line change.

S7 Wrong scope#

save_research would have opened an issue, so it was not used. This session then followed the steering records toward a branch and rebuilt the deliverables as three files, until the author corrected it. The cost was rework only, with nothing written outside the session. The validator and the frozen copies in this folder are checking aids for this record, and no product code changed.

S8 Invented human priorities#

The risk is real, because picking three is itself a selection. The pick rests on the author's own written words or a dated external page, and it is labelled a suggestion. The asks_person_now field is this session's answer to four questions, not a measure. CM-03's date applies only if the author applies this cycle, which is unknown. The first draft used gendered pronouns for the author and the named person, which no one stated. They were replaced, and G15 now fails on them.

S9 Rows resting on one review#

The first build found six rows with a single review behind them, against a brief that asks for cross-review candidates. Sources were added where another review speaks to the same question, and no row was merged to reach this. Some of those added sources are method statements rather than independent observation of the same event, so "cross-review" here means the same question appears, not that one review confirms another.

Quality gates#

Command, from this folder: python3 -I tools/validate.py. It fetches the pinned reviews. Add --frozen frozen to read the copies here instead. Python 3.12.4 at /home/jgi/anaconda3/bin/python3, jsonschema 4.25.1. A pass means well-formed and quoted correctly, and says nothing about whether a reading is right or useful.

16 passed, 0 failed. 12 rows, 79 atoms, 63 in rows, 16 kept out with reasons.

Three candidates to discuss#

Suggestions for discussion. None is launched or assigned, and only the author validates them. They are in no rank order.

CM-11 The author's corrections on the Reflective Return review#

Reason: the author wrote that the content needs revision, and asked for a ceremony to review it. The ceremony exists and shows 0 beats, the table has waited since 2026-10-27, and six issues name no owner until it is marked. The rejected sentence still sits in the part of the review agents read first. Asked of the author: one sitting.

Alternative: leave the document as is and treat the post-study notes as the record, since they sit in the same version. Or mark only the three rows that name no holder or an unenforced word.

CM-06 Public reach and a named person#

Reason: a person outside the room is named on a page listed in a public feed, while earlier records hold that person's consent as not cleared and say nothing reaches a public page before the author decides. It is the only row where waiting has a cost to someone else. Asked of the author: read one section and answer.

Alternative: the author may consider this already decided, having approved v4 and sharing reviews publicly on purpose. Then the mark is drop, and it costs minutes. Or take it with CM-05 as one stewardship conversation.

CM-03 Work and reflection for the Concordia proposal#

Reason: the author asked to understand what the school expects, the twelve findings are already written, and the page gives a fixed date for this intake. No engineering is involved, and it is the row closest to the author's own writing. Asked of the author: one reading session.

Alternative: if the author is not applying this cycle, defer it. CM-01 or CM-04 could take this place: CM-01 if shooting is the present focus, CM-04 if the immediate concern is whether the Factory is helping. CM-05 could take it if stewardship should come first.

Completeness and blockers#

All of steps A to G were run, and all deliverables are written. The rest is limited by I1 to I7. The author's marks are outstanding by design.

No blocker remains on delivery. Missing permissions: the claude.ai connector sign-in is read-only, and the connector has no GitHub token (I6, I7).

Candidate rows are readings by one model on one day. Rerunning the checks is repeatable. Rerunning the reading would not give identical words.

Receipts#

RC1 Transcripts#

The client's transcript command: has_transcript false for PI, PII and RR. NL true, 35,720 characters, source gemini-3.7-flash, 91 turns, last at 152:11, 45 turns labelled. Of its 8 labels, 3 are not speakers: a title fragment, "Branch 001.2" and "Mia, Phase One mission revision". No turn lies between 15:11 and 24:12. Canonical JSON SHA-256 is in frozen_inputs.

RC2 Tests#

Run at HEAD 2ab2fb00 on 2026-11-04, each with its exit code read directly. npm test was not used on the connector and the schema, because it rebuilds dist in a shared checkout. These runs read the existing dist.

RC3 Public reach#

With no token: /raw and /json answered 200 for all four. /review/list, /review/index and /review/feed.xml answered 200 and each holds all four ids. Client read of /review/8031484c…/transcript answered 200, and the other three 404.

RC4 Citations#

See the "Citations" section.

RC5 Video pages#

Each watch page, read 2026-11-04: playabilityStatus OK, isUnlisted false, lengthSeconds 870, 5954, 9395 and 6934. oEmbed answered 200 with author JGWill for all four.

RC6 Capabilities#

get_capabilities on 2026-11-04: @miadi/connector 0.1.20, reading_workflow.contract transcript-first-v1, provider_runtime not_configured, import blocked_until_authenticated_sha256_cas_v1_handshake, publication separate_exact_candidate_approval. /health reports 0.1.20.

RC7 Attention#

grep -rIn knowledgeFocusVerdict over app, lib, packages, skills, scripts, the kit's skills and claude folders and /etc/claude-code/skills: only the schema's source, index, README and test. grep -rIln "Knowledge Focus" over the kit, /etc/claude-code/skills and skills: none. The chronicle holds 11 attention.json files and 36 items, 18 open and 18 answered, 0 with for, 0 with focus. 4 were asked after schema 0.17.0 (commit f2555927, 2026-10-27), all in miadi-chronicle://567.

RC8 Wheel and chronicle#

/api/edges?limit=all: 855 edges. To PI, PII and NL: 0. To RR: one, draws-on from miadi-chronicle://563. miadi-chronicle://570: continues_from 567, relates_to 044 both ways, cites the notebook. No edge names Episode 067, and its node id gave no node. /api/narrative/beats?limit=all: 295 beats, 0 naming ceremony 0ef2f473. Episode 067's folder holds no episode.yaml, and its HELD.md describes that state. The chronicle shows five Episode 570 commits on 2026-11-03: 21:59:35, 21:59:55, 22:58:34, 22:59:07, 23:06:34, local time, author jgi.

RC9 Chain#

miaco/src/screenwalk-review.ts:100: delivery: "named-not-called". /srv/miadi/factory/review-workflow.jsonl: 3 lines, all ui-screenwalk on review 9281cda0. /srv/miadi/factory/review-ready.jsonl: 0 lines naming these four. parseTranscript from the installed build on NL's stored transcript returns 0 turns, because 0 of its 181 lines match [time] Speaker N:. Three reviews have no stored transcript, which fetchStoredTranscript refuses. The chain was not run on any review.

RC10 Prior map#

grep -rlE 'contribution-map|Contribution Map' over /srv/miadi/factory, /srv/miadi/connector, .mia, the mia workspace's .mia and .mino, and the chronicle staging folder: only the brief, its delivered copy and the ledger. All four ids together, in the same paths: the same three files.

RC11 Steering#

session_steer_request:c7df8094, 3e28bf29 and d7bb97d2 at 21:44:31Z, 21:45:03Z and 21:46:25Z on 2026-11-04, each with a paired blocked record: "is mid-turn; the steer was not typed". Attempts used 0 of 3. Read from the connector ledger on gaia at about 21:55Z.

RC12 Structured data#

version.data.aspects and version.data.results have length 0 in the frozen JSON of all four versions, all manual-import.