# Explainer Error Robustness K001 Current Handoff

Status: `COMPLETED_WAITING_FOR_RESEARCH_DIRECTOR`

Completed task: `diagnose-win-qwg-residual-coupling-v1`

Normative frozen protocol:
`milestones/explainer-error-robustness-k001/prompts/007-diagnose-win-qwg-residual-coupling-v1.md`

Protocol freeze commit:
`c8709beaf5ac632dcd9a45701f67b352c6846430`

Implementation branch:
`backgammonsimplified/backgammon-explainer@research/explainer-error-robustness-k001`

Generation 7 starting implementation:
`45b5200cd5e0ad5f54cfebef50da29f8a9cbc018`

Completed implementation/result commit:
`92fc89d3fc93429f5b357d1e0f5122c345625bd0`

Host: `carbonated-water`.

## Generation 7 durable result

Terminal outcome: `MIXED_RESIDUAL_COUPLING`.

Before outcome access, the exact inherited DEVELOPMENT authority reproduced:

- full candidate membership: 2,094,039 rows / 100,015 decisions / 3,178 complete-game groups;
- exact midrange membership: 337,043 rows / 35,253 decisions / 3,121 complete-game groups;
- accepted prediction stream: `d8ec82da46f5255bc6eacae88ee90132a331880161567f325923b86f66e26232`;
- ordered eligible candidate identity: `c60f0f44872f5e699fa5f1d5ea9985249e18efc5d9df85517b7f1a5e4147133e`;
- inherited candidate-fold assignment: `5a8296a2d36e9dce47ece392b69f10496304825a3fa828dd99c33567be008018`;
- inherited group-fold assignment: `07fc25432bc542aac10f21fd8d86da64ba6e83f7679afe9f533392fa5a716551`;
- accepted soft-binomial label-mass stream: `7bc94362fd98eef42edea7807ba103053880a64bea93d9bd67386be2dcd7daef`.

The pre-outcome residual-semantics receipt preserved the accepted head-wise soft-binomial representation without constructing a conditional target ratio:

- `r_win = static_win - p_win`;
- `r_qwg = static_win_gammon_or_better - static_win * q_wg`;
- equivalently, each head uses the ratio-free score residual `success_mass - conditioning_mass * head_prediction`;
- Pearson correlation and mean absolute residual use unweighted candidate rows;
- `q_wg` Brier loss uses the accepted success/failure masses divided by accepted conditioning mass.

All 337,043 frozen rows reproduced during the single outcome pass. Pooled diagnostics were:

- Pearson correlation between `r_win` and `r_qwg`: `0.23580432251263408`;
- residual sign concordance: `0.5831511112825366` (`196,547 / 337,043` rows; exact-zero exclusions `0`);
- mean absolute `q_wg` residual: `0.059189042185287706`;
- accepted soft-binomial `q_wg` Brier loss: `0.2438645597321206`.

Inherited-fold correlations were all positive:

- fold 0: `0.21250581334820642`;
- fold 1: `0.2710104282213475`;
- fold 2: `0.17629921924040742`;
- fold 3: `0.2594801218298955`;
- fold 4: `0.24883490059002447`;
- fold 5: `0.33327214715019315`;
- fold 6: `0.2512434691858107`;
- fold 7: `0.1462786900229137`.

All four frozen `abs(r_win)` bins passed the 5,000-row, 200-group, and every-fold 250-row support gates:

- `[0.00,0.05)`: 195,208 rows / 3,104 groups; mean absolute `q_wg` residual `0.05263931754433219`; `q_wg` Brier `0.2436298729593826`;
- `[0.05,0.10)`: 85,642 rows / 2,886 groups; mean absolute `q_wg` residual `0.06127655817893638`; `q_wg` Brier `0.245509145262209`;
- `[0.10,0.20)`: 48,360 rows / 2,405 groups; mean absolute `q_wg` residual `0.07526911657291578`; `q_wg` Brier `0.24265087041296235`;
- `[0.20,1.00]`: 7,833 rows / 986 groups; mean absolute `q_wg` residual `0.10031596744997673`; `q_wg` Brier `0.23855842505093997`.

Mean absolute `q_wg` residual was non-decreasing across the supported bins, and the highest-minus-lowest difference was `0.04767664990564454`.

`UPSTREAM_RESIDUAL_COUPLING` did not pass because pooled correlation was below `+0.30` and only `6/8`, not at least `7/8`, folds reached `+0.20`. Its all-positive-fold, non-decreasing-bin, and `0.010000` bin-range gates passed.

`INDEPENDENT_CONDITIONAL_RESIDUAL` did not pass because absolute pooled correlation exceeded `0.10`, only `1/8` folds was at most `0.15` in absolute correlation, and the supported-bin range was not below `0.005000`.

The prospectively frozen fallback for a supported result satisfying neither complete rule is exactly `MIXED_RESIDUAL_COUPLING`.

Model fits were exactly `0`. TRAIN outcomes used were exactly `0`. Candidate models were `NONE`. Feature search and hyperparameter search were exactly `0`. PROTECTED FINAL EVALUATION access was exactly `0`; protected membership, predictions, labels, and outcomes were not opened.

Generation 7 identities:

- input/prediction/fold receipt: `c4b57db539ae2255a822a9d8da129e932d8ef2837d35802a20c1111c068a6cd2`;
- residual-semantics receipt: `5e8ab76e21d20b7c76b0591222a3ff119a0bc2c8b7bf1ff2b9706e9835c0b1c7`;
- global residual diagnostics: `8810d822f640a4af5f875d83c39581e327580e6fcec6179e2baf6ff4766c43dd`;
- inherited eight-fold residual diagnostics: `ff3c36604342854e71d2b7a3be27650718e91cfa6655c9a64c16d5f483ca1781`;
- fixed residual-bin diagnostics: `1ab950ca9af091f033049b3d1f987a3c9f6fd179c915b18f753df6dd4ce67e8f`;
- residual-bin support table: `b7f9786ee658c5f55bb59c31a7e0c1c4b45d2ac118f82d8f4802e1c099183165`;
- terminal routing record: `bd90176ec61cd76862ba933fe349d7a09fe965601fa4747dd6e96f6f7d2a197e`;
- protected-access log: `031b6aa8ad80afd374dccc3b5f829b88652b253373c258e72f8f663f64f889bd`;
- deterministic verification: `8eb11efe1d2a840c6685231b296f43dc391382b6700dc77a41e06be62f7c2136`;
- validation/test evidence: `0dfdc86ee32ebfbd7e13bc61095c6743afa32aabbdf89be670b8eab6ceeb1b44`;
- artifact package: `249d8000aced9873a7ff436f59b6e6b12ebdafda4249025e678f0894179bf003`.

Independent artifact verification passed. All 56 focused and inherited Generation 1–7 tests passed, including exact residual and bin boundaries, support gates, all deterministic terminal routes, accepted soft-binomial hierarchy semantics, inherited folds, portable scorer parity, and manifest tamper detection.

## Authority and routing boundary

Accepted product architecture remains unchanged: `ridge-ranking-hadd-value-explanation-sidecar-v1`. Pairwise Ridge remains sole recommendation/ranking authority. Compact HADD remains accepted probability/value/explanation authority. No research model was fit, proposed, selected, or promoted.

Per the frozen `MIXED_RESIDUAL_COUPLING` routing, a later Research Director pass may choose one new factual diagnostic axis justified by this completed evidence. This worker does not choose that axis, reopen the failed scalar-correction or low-candidate-transfer families, fit the separately reserved Mannitol win-head experiment, or invent Generation 8.

There were no new GNU or Sage computations, matches, labels, generic 0-ply generation, Sage/GNU campaign validation data consumed as training, production promotions, Analyzer mutations, Canonical mutations, Corpus mutations, Cloudflare/public deployments, or calculated-cubeful authority changes.

The implementation and Task Management commits are local only and must not be pushed by this worker. The supervisor may verify the clean state and push the owned branches.