---
title: Score and admit the miss-mechanism decision model
type: task
status: completed
description: >-
  Score Clef-Flash against several weeks of filed miss-mechanism judgments and admit it at an
  operating point or record its rejection. Waits on weekly reviews filing judgments; admission is an
  operator decision.
base_uri: user:task/league/score-and-admit-miss-mechanism-decision-model.md
created_at: '2026-10-20T14:56:16.798Z'
entity_id: b61f80fd-b422-4bde-ac25-ef5125a29266
finish_by: '2026-11-25T00:00:00Z'
finished_at: '2026-11-04T16:23:42.319Z'
observations:
  - ████████████████████████████████████████████████
  - ████████████████████████████████████████████████
  - ████████████████████████████████████████████████
  - ████████████████████████████████████████████████
owner_identity_uri: █████████████████████████
priority: Medium
public_read: true
relations:
  - follows [[█████████████████████-████-███████-███-███████████████-██-█████-████████]]
started_at: '2026-11-04T15:15:08.381Z'
tags:
  - user:tag/league-xo-football.md
  - user:tag/nfl-betting.md
updated_at: '2026-11-04T16:23:42.319Z'
visibility_analyzed_at: '2026-11-05T23:50:21.187Z'
---

# Score and admit the miss-mechanism decision model

Split from [[█████████████████████-████-███████-███-███████████████-██-█████-████████]] on 2026-10-20, whose measured and judged layers are deployed. This task waits on data. Each weekly process review ([[████████████████████████████████-███-██████-███████-██████████]]) files `review_session` judgments into `miss_mechanism_judgment` with `cli/nfl/file-miss-mechanism-judgments.py`. On 2026-10-20 the store held 231 rows, all from week 6. Scoring needs several weeks.

Clef-Flash scores on structured measured fields only (operator ruling 2026-10-19). Admission is an operator decision.

## Plan

- [x] **Confirm coverage**: `base database query miss_mechanism_judgment --json --limit 1000000`, grouped by `judge_kind` and game week (from `esbid`). Verify: at least three settled weeks carry judgments. If fewer, snooze this task and its continuation to the day after the next weekly review.
  - State: 2026-11-04, 584 rows; 573 latest keys after operator-over-session precedence. Weeks 6 (239), 7 (159) and 8 (175) all carry review-session judgments, so the snooze branch does not apply. Weeks 6-7 are registry v2; week 8 is v4.
- [x] **Decision-model scoring**
  - File: `cli/nfl/score-outcome-decision-model.py`
  - Purpose: score Clef-Flash against the session and operator judgments, then admit it at an accepted operating point or record its rejection.
  - Verify: a `[spike]` observation with accuracy and calibration per choice, and an operator `[decision]` observation; when admitted, one week's `decision_model` judgments appear with `judge_model` set.
  - State: the `miss-judgments` fixture is committed (user-base `f5d0bdad93`) and both modes are scored; the `[spike]` observation is filed. Nothing beats the agree-with-measured baseline. The operator rejected the model as a judge on 2026-11-04 (`[decision]`), so no `decision_model` rows are filed.
- [x] **Drift re-review of weeks 6-7**
  - Purpose: re-judge every weeks 6-7 key under registry v4 against the current classifier, before any model work on the measurement step.
  - Verify: every latest `review_session` key carries v4 and a measured class the current classifier reproduces.
  - State: 2026-11-04, 398 judgments filed (11 overrides), 573 of 573 keys at v4, 0 drift. The 340 rows with no in-game exit were re-filed mechanically, since the v3 and v4 change touches availability only; the 58 with an exit or availability class were read against exit timing and play-level splits. The three week 8 drift rows are operator rulings the classifier now reproduces.
- [x] **Availability evidence-extractor pilot**
  - State: skipped. Rescored on the re-reviewed labels, Clef-Flash's availability flags are 0 of 12 right and it missed the one real availability override, so no signal remains beyond drift (`[finding]` 2026-11-04).

## Key locations

- Judgment store: [[█████████████████-████-█████████-███████████]]
- Publisher of the judged layer: `cli/nfl/publish-betting-selection-reads.py`