Context#

On 2026-10-10 Guillaume sent one session (miadi-episodic-memory-schema-and-mw-261010, Mia on Fable 5.1) a five-part request: relate a v0 prototype (jgwill/Miadi packages/ep120-factory-interface-analysis/app, jgwill/Miadi#757) to miadi-chronicle://120 with its published address for readers and its editing address for admins; check whether Deep Diver's notebooks in miadi-chronicle://550 are typed; connect Langfuse traces and prompts to episodes per domain (CEREMONY_, STORY_, ENG_LANGFUSE_*); relate tide sessions to an episode; and hold RISE specs in an episode as foundations, reviews and stateloom drawings are held, with the RISE text touched up so the cycle the prototype implies closes.

The session surveyed the schema, the wheel, the two episodes, the binding line and the RISE text, ran three Sonnet read lanes and one Opus second opinion formed without its draft, and published a proposal page, "What an Episode Holds" (a Claude artifact, private; its HTML is kept beside this inquiry). The page carries 23 coded cards: six steps A1–A6, seven RISE edits E1–E7, six decisions D1–D6, four questions Q1–Q4.

Guillaume then asked two things the first prompt could not carry:

1. Score the proposal with the decision engine (Jev, the System One model the Episode 076 review is about), and split it into what the agent is certain enough to execute and what needs him. His words: "here is what I would do that I'm certain about, and there would be another section where you're not certain about."
2. Treat that as developing the system: "I want you to self-evolve and be able to use that, and other new instances in the repositories and inside the factory will be able to reuse that intelligence." And: "by using that engine, you'll be able to take decisions and come back to me just when this is actually needed."

What existed for Jev before this day: the review 80f50d5b (System-1 Decision Engineering with Jev and TypeSafe AI) held by Episode 076 with its circle circle:1790987877584:llr51t; lane-check in jgwill/smcraft examples/decision-gate/lane-check (typed questions, the DecisionGate machine, a ledger, calibration from rulings); the first runs on Mia's returns in Episode 120's branches/jev/ (2026-10-09); and Guillaume's rule of 2026-10-09 in session miadi-jev-fork1: act when his word and the engine agree at 0.8 or above, come back otherwise.

What was done on 2026-10-10:

• skills/proposal-visualization/scoring/score-proposal.mjs, questions.json and policy.json in this kit: every coded card on a proposal page is sent to Jev with the owner's request in his words, the page's own Today facts and the Decision Gate's D1 rule, and asked whether the request implies it, whether it is reversible, whether the facts ground it, whether only he may decide it, and obvious_now (do now, do after his go, needs a decision). Bands: needs you (D1 rule, whatever the number), would do now (do_now at or above 0.8), after your go.
• The first run, Jev 1.13.0, 23 calls, about 53k input tokens: would do now, none; after your go, A1–A5, E1–E6, D3, D4, Q1; needs you, A6, E7, D1, D2, D5, D6, Q2, Q3, Q4. A1 and E1 read do_now at 0.70. What held every step under 0.8 was evidence_sufficient (0.14 to 0.40 on every card): the engine reads the page's Today facts as measuring what exists, not yet grounding the proposed shapes. reversible was high for the schema steps and low where a step lands in the vessel, the wheel and an issue (A2 0.24, A6 0.16). implied_by_request was high only for E1 (0.81) and A2 (0.74); A1 at 0.18 is a reading a person would rule against. Ledger: ~/.miadi/decisions/proposal-score.jsonl.
• The page was revised: a chip on every card, one section with the three lists, a Revision 1 band.
• The practice was written into skills/proposal-visualization/SKILL.md ("Certain and uncertain: score the cards before publishing").
• The three messages were decomposed (miaco decompose run) into Episode 076's .pde/, beside this inquiry.

Desired State#

An agent that writes a proposal in this factory ends it with three lists the engine helped produce: what it will do now and report, what it will do once the owner says go, and what only the owner can answer. The owner reads the third list and the page, and nothing else reaches him. His one-word answers on the "needs you" cards become rulings in the ledger, the thresholds are calibrated from them, and the next proposal's first list grows. Any instance, in any repository, runs the same script with the same questions, and the kit is where it finds them.

Action Steps#

• Guillaume rules on the nine "needs you" cards of the first scored page (D1, D2, D5, D6, Q2, Q3, Q4, A6, E7), in his words; each ruling is recorded against its decision id in the ledger (lane-check.mjs rule shape).
• Guillaume rules on A1 and E1 (0.70): were they obvious? Those two rulings say whether 0.8 is the cut for obvious_now on proposals or whether it sits lower.
• A rule and calibrate command for the proposal ledger, or lane-check's own once its decider moves into this kit (lane-check/ISSUE-DRAFTS.md, draft 1).
• The questions get a second version from the first rulings: needs_owner read E7 (consistency edits) as the owner's at exactly 0.50, which a person would rule against; the wording that produced it is revised.
• The scoring step is named in the proposal-visualization skill's "Before you publish" list (done) and in the dispatch-discipline survey, so a lane that writes a proposal scores it before handing it up.
• A second proposal page, by another instance, is scored with the same script, and its run is woven into this inquiry.

Structural Tension#

The owner routes some fifty sessions and wants only real decisions to reach him; today every proposal reaches him whole, because the agent that wrote it has no instrument that says which of its cards are obvious. The desired state is a proposal whose obvious part is already done when he reads it. The engine gives the number, the D1 rule keeps names and audiences his, and the rulings he gives on the first pages are what turns the number into a cut he trusts.

Related#

• miadi-chronicle://76 (the Jev review and its circle), miadi-chronicle://120 (first Jev runs on Mia's returns, branches/jev/), miadi-chronicle://551 (learning between sessions)
• jgwill/smcraft examples/decision-gate/ (the machine, lane-check, decision_record.erdf.json)
• jgwill/miadi-orchestration-kit#53 (the proposal-visualization skill)
• jgwill/Miadi#757 (the prototype the proposal starts from), jgwill/Miadi#767 (schema 0.14.1, the base the proposal now names)
• miadi-review:80f50d5b-d5f3-4c82-a6b5-5302db559f5a
• The proposal page's HTML and scores.json, in this inquiry's folder