# Explainer Native Cubeful K001 Current Handoff

Version: `project-task-manager-explainer-native-cubeful-k001-handoff-v5`
Updated: 2026-08-24 EDT
Status: `GENERATION_3_FROZEN_READY_FOR_IMMEDIATE_CODEX_AFTER_LIVE_PREFLIGHT`

Current task: `test-crawford-state-context-family-ablation-v1`

Normative frozen protocol:
`milestones/explainer-native-cubeful-k001/prompts/003-test-crawford-state-context-family-ablation-v1.md`

Protocol freeze commit:
`222c2a50418082c65fd57792564c1dc1cc1ad0f6`

Implementation branch:
`backgammonsimplified/backgammon-explainer@research/explainer-native-cubeful-k001`

Generation 3 starting implementation:
`75ce3ba6cf9a03db98886b26f1d2dc6b057f6e2a`

Accepted product architecture remains fixed:
`ridge-ranking-hadd-value-explanation-sidecar-v1`.

Pairwise Ridge remains sole recommendation/ranking authority. Compact HADD remains accepted probability/value/explanation authority. Any improved model is only `CANDIDATE_FOR_FUTURE_INDEPENDENT_MODEL_SELECTION`.

## Completed Generation 2 evidence

Generation 2 `diagnose-native-cubeful-context-error-structure-v1` completed at implementation:
`75ce3ba6cf9a03db98886b26f1d2dc6b057f6e2a`.

Result identity:
`b40afe6802c375fcc89d3a34a3cbeef4cec7708576c9c3dce971e44ca11dbddb`.

Artifact package identity:
`d4be9a1562a02b62d4716bcc9cd94bcc3422531f0d483666a81dcc57872143d7`.

Terminal classification:
`CONTEXT_FAILURE_LOCALIZED`.

Deterministic routing cell:
`crawford_state/money_or_zero`.

The primary Generation 1 P3-plus-context stream had eight supported `STABLE_CONTEXT_REGRESSION` cells across all six frozen families. The strongest useful evidence for routing is that the full context candidate's DEVELOPMENT regression remains broad, while the prospectively frozen deterministic localization rule selected the Crawford-state `money_or_zero` cell.

Important limitation: every DEVELOPMENT row is money play. Match-length, score-relation, and Crawford-state therefore each have exactly one populated supported `money_or_zero` cell, each with within-family positive squared-error regression mass share `1.0` and identical RMSE regression `+0.027784317582417117`. The selected Crawford family is therefore a deterministic routing result, not causal evidence by itself.

Generation 2 performed model fits `0`, used modeling candidates `NONE`, accessed PROTECTED FINAL EVALUATION exactly `0` times, reproduced the exact Generation 1 population and fixed prediction streams, and passed package/hash verification. Production, Analyzer, Canonical, Corpus, Ridge/HADD roles, and calculated-cubeful authority remained unchanged.

## Frozen Generation 3 question

The Generation 2 protocol prospectively permits one bounded context-family representation or ablation tied to the selected routing family. Generation 3 therefore performs exactly one causal test:

- reproduce the Generation 1 full P3-plus-context additive specification;
- remove only accepted context field `cubeful_crawford` before additive basis construction;
- retain all other 14 context fields and all P3 position features unchanged;
- reuse exact TRAIN/DEVELOPMENT memberships, four grouped folds, target semantics, additive basis, alpha `100.0`, optimizer, seed, and numerical policy from Generation 1;
- fit exactly one new candidate;
- perform no hyperparameter search, feature search, interaction search, or alternative-family ablation;
- PROTECTED FINAL EVALUATION access budget remains exactly `0`.

The terminal rule is prospectively frozen. `CRAWFORD_CONTEXT_FAMILY_CAUSAL` requires at least `0.005000` global RMSE recovery and `0.002000` global MAE recovery versus the Generation 1 full-context control, stable improvement in at least `3/4` grouped folds, no fold regression above `0.002000`, and no material global-bias regression. Otherwise the result is `CRAWFORD_CONTEXT_FAMILY_NOT_CAUSAL`.

No result in this generation changes production or automatically selects a model.

## Coordinator execution

The existing HFCS lane supervisor should consume this handoff and launch one Codex worker after a fresh host preflight. Exactly one active Codex is permitted for this lane.

Before substantive work, verify fresh CPU/RAM/disk/process safety. Run the Explainer task at nice `19` with conservative numerical threading. Corpus Listener/writer, Post Match, and existing GNU/Sage processes retain priority. Never signal, stop, renice, restart, reconfigure, or otherwise mutate those priority processes.

Execute only the frozen Generation 3 protocol through durable implementation artifacts, tests, result, and Task Management closeout. At completion leave the lane `COMPLETED_WAITING_FOR_RESEARCH_DIRECTOR`; do not invent Generation 4 within the Codex worker.

## Absolute boundaries

Existing accepted data only. New GNU `0`; new Sage `0`; new matches `0`; new labels `0`; generic new 0-ply `0`; Sage/GNU campaign validation data as training `NO`; production promotion `NO`; Analyzer mutation `NO`; Canonical mutation `NO`; Corpus mutation `NO`; calculated-cubeful authority change `NO`.

`CUBEFUL_CALCULATION_AUTHORITY_BLOCKED` remains absolute.