Skip to content

docs(EAW-9): teacher-blind chess rediscovery and withdrawal experiment plan #1829

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
AdaWorldAPI merged 1 commit into from
Oct 28, 2026
Merged

docs(EAW-9): teacher-blind chess rediscovery and withdrawal experiment plan #1829

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
57 changes: 57 additions & 0 deletions ...esearch/epistemic-awareness/eaw-9-blind-chess-rediscovery-teacher-withdrawal.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,57 @@
# EAW-9 — Hindsight-blind chess rediscovery: 0+63 Thought-Git × teacher × NNUE/NARS

**2026-10-27. Research plan only, no learning run, implementation, promotion or merge authorization.** Independent main-based branch. Idea provenance [EAW-6 #1826](https://github.com/AdaWorldAPI/lance-graph/pull/1826); read EAW-7 strict world frontier, EAW-8 NNUE feature adapter, EAW-4 [#1824](https://github.com/AdaWorldAPI/lance-graph/pull/1824) and CFA-4 [#1817](https://github.com/AdaWorldAPI/lance-graph/pull/1817). This stage **does not** replace Stockfish-rs D-CHESS-NARS-0.

## Central research hypothesis
Can a chess learner, limited to the exact position and historically available evidence at ply v, use at most 63 active speculative Thought-Git lookahead worlds and sparse NNUE feature deltas to rediscover conditional tactical dependencies or better candidate ordering over time? *Only after deciding* may a deeper independent teacher reveal its evaluation and actual game result. The learner's advantage must survive source-disjoint held-out games and explicit teacher withdrawal. If it fails, do not claim rediscovery or emerging causal understanding.

## Existing shipped / probed spine, avoid duplication
- [D-CHESS-NARS-0](https://github.com/AdaWorldAPI/stockfish-rs/blob/main/.claude/plans/stockfish-harvest-64x64-v1.md): learning flow NNUE teacher → NARS recipe revision → persistent strategy → teacher-withdrawal, with D-CHESS-NARS-0a recipe census; current L6 search shell / policy seam, not complete Stockfish tournament strength.
- [DecisionEpisodeV1](https://github.com/AdaWorldAPI/stockfish-rs/blob/main/docs/DECISION-EPISODE-V1.md): physically separate StrictObservation and RetroLabel, game_id/ply vs exact PositionKey vs approximate recall vs palette; stable deep `TeacherLabel` and rare search events. No Retro data in `CandidatePolicy::score`.
- [expert_iteration_stream.rs](https://github.com/AdaWorldAPI/stockfish-rs/blob/main/examples/expert_iteration_stream.rs): full-window exact per-root candidate teacher, frozen/privileged baselines, deterministic episode encoding, candidate refutation/recovery/PV changes, game-level partition; deep *play-path alpha-beta* values may be bounds, not exact rankings.
- [hindsight_stream.rs](https://github.com/AdaWorldAPI/stockfish-rs/blob/main/examples/hindsight_stream.rs): real Lichess PGN results and version read fence; [rung_hindsight.rs](https://github.com/AdaWorldAPI/stockfish-rs/blob/main/examples/rung_hindsight.rs) was INCONCLUSIVE under final NNUE-only score oracle (mate blindness).
- [temporal_replay.rs](https://github.com/AdaWorldAPI/stockfish-rs/blob/main/examples/temporal_replay.rs) / [planner temporal.rs](https://github.com/AdaWorldAPI/lance-graph/blob/main/crates/lance-graph-planner/src/temporal.rs): independently replayable as-of histories and strictly pinned reading horizon. `TemporalPov::at(v,0)` enforces version range, but full `classify/deinterlace` and data-causal readiness belong to planner.
- [CFA-3 #1816](https://github.com/AdaWorldAPI/lance-graph/pull/1816): ARM co-activation is not causal proof; [CFA-4 #1817](https://github.com/AdaWorldAPI/lance-graph/pull/1817) optional NSL/ONNX ranking, training split and evidence restrictions.
- [EAW-5 #1825](https://github.com/AdaWorldAPI/lance-graph/pull/1825): legal / Odoo to follow *after* chess; no real domain causal claims from synthetic worlds.

## Study design: distinguish THREE time domains
- **Observed past** `Past≤v`: PGN and source records available to the actor at ply v, source witness pinned; Strict reader `QueryReference::at(v,0)`. Historical future positions/results are invisible.
- **Simulated future** `Hypothetical(T_i,depth)`: legal alternative moves generated from current board/learned heuristic within one independent Thought-Git world. Its successor board isn't a source observation and carries no empirical independent evidence root.
- **Hindsight/teacher** `Retro(v)`: full-window independent deeper evaluation, PV/refutations/search labels and actual PGN result. This exists as a separate later evidence lane only. Evaluating a hypothetical continuation for scoring is allowed Strict-time search; reading *actual* continuation/teacher labels is not.
- Multiple candidate rules/heuristics can coexist in one world. A Thought-Git sibling world with a different choice cannot be semantically unioned into one chess position. T0 immutable, T1..T63 live slots with external generation and source seals. All training labels reference correct baseline and source split.

## Stage ladder (future test-only PRs require owner approval)
**R0: Frozen baseline and leak gate.** Pin game-level train/validation/test splits, board position, legal moves, frozen NNUE model, strict observation and policy; exact teacher score bound and source revision. Copy existing Golden episode record, do not invent a new one.
**R1: Bounded lookahead oracle parity.** EAW-7 compare legal successors and protected path-world receipts to independently replayed shakmaty positions; EAW-8 exact NNUE delta vs refresh (king special cases). Count world slots and budget; do not claim parallel execution.
**R2: Simple *interpretable* rediscovery features.** Without teacher knowledge at Strict-time, expose capture, check, threats, material, king safety, mobility and typed relational support as deliberately declared chess features. Lookahead can learn patterns of `requires`, `prevents`, conditional forks, forced mates, transpositions, tactic refutations. Do not label chess heuristic scores as certified causality.
**R3: Train ranking (optional).** Compare seeded-random, frozen MVV-LVA/butterfly ordering, ARM support/association, NNUE-style incremental feature ranking, candidate rules and privileged deep teacher upper bound. Same CPU, nodes, total work, branch/world budget, move legality, teacher depth, position corpus and per-game split. Train on Retro labels after Strict decisions, freeze weights and feature schema for held-out measurements. Teacher labels with unstable depth/margins remain withheld or separately scored; `NarsTruth` synthetic branches aren't independent empirical observations.
**R4: Teacher withdrawal.** Freeze learned ranker, remove future teacher feedback on held-out matches and examine whether improvement persists over frozen/reset baselines. Compare training-only vs inference-with-teacher. Retrospective label stream cannot alter earlier Strict ranking, no leakage from adjacent plies of same test game into training.
**R5: Conditional explanation not just rank.** For a given scenario, ask which guard/fact/move causes the *tested* deeper outcome to differ. Use protected intermediate traces, independent first and joint ablations, preserve derivation roots and multi-feature alternatives. Emit candidate explanation such as `if king exposed ∧ defender pinned, check sequence wins`, with explicit scope and counterexample; never promote to universal theorem without separate independent proof.

## Falsifiers and negative controls (NOT RUN)
| ID | Perturbation | GREEN only if |
|---|---|---|
| BL-01 | Replace all future game plies/Retro labels for fixed Strict observation | Strict feature IDs, branch selection and ranking byte-identical |
| BL-02 | Select candidates with privileged deep teacher inside policy callback | Hard reject; teacher only in Retro/training |
| BL-03 | Final NNUE score sign disagrees with mate-aware game outcome | PGN result/teacher retains authority, NNUE-only proxy invalid |
| BL-04 | Same PositionKey in distinct episodes across train/test | Game_id-disjoint split rejects historical leakage |
| BL-05 | Identical root appears in many Thought-Git branches | No artificial NARS confidence or ARM independence inflation |
| BL-06 | Equal final eval but differing PV refutation, checkmate trace or path | Protected differences remain, no strong equivalence |
| BL-07 | Learned high score selects illegal move | shakmaty rejects; no execution right granted |
| BL-08 | Schoolbook heuristic coactivation says causal influence despite no ablation effect | Labeled association only, not causal law |
| BL-09 | NNUE feature update stale on king move, book model/feature hash changes | Fresh refresh / refusal; exact parity |
| BL-10 | All 63 slots live / ABA stale recycled W | Explicit bounded queue/refusal; no slot-as-evidence-pointer |
| BL-11 | Training uses same game, future plies, or transposed episode across split | Leak detected; held-out invalidated |
| BL-12 | Deep teacher scores are narrow-window bounds but used as total exact ranking | Refuse; full-window teacher or typed bound treatment |
| BL-13 | NNUE learned arm loses to frozen/ARM when full cost charged | PARK optional learning, keep deterministic path |
| BL-14 | Teacher withdrawal destroys measured gain | Report dependence; no autonomous rediscovery claim |
| BL-15 | Learned chess conditional transplanted as legal norm | Reject cross-domain promotion and any source effect |

## Measurements / reproducibility
Repeat across unlike sources: short forced-mate/pin/tactic corpus, broad game-disjoint real PGN, positions with severe class imbalance/quiet positional play and deceptive sacrifices. Seed splits and world choices. Report teacher instability and precision/recall@k, top-1/top-3, node count to tactical discovery, search wall-clock and cold cost, model training/inference, abstentions, book leak checks, exact delta-vs-refresh rate, memory/slot pressure, branch roots, protected trace bytes and outcome variance. Compare new learned heuristics with carefully matched baseline. Do not declare a single test-game success generalization.

## Ownership and explicit red boundaries
Stockfish-rs owns chess search, `shakmaty` legal semantics, licensed NNUE weights/feature index and teacher-policy seam; lance-graph owns temporal contracts, Thought-Git world receipts, DR64³ research and exact Quack/RelAlg. Existing D-CHESS-NARS-0 is the learning arc, do not launch a rival scheduler or learned truth store. EXO-98/148 durable world refs OPEN; EXO-46 physical source binding OPEN; EXO-109 independent oracle, EXO-133 real data, EXO-135 evidence law and EXO-152 numeric owner. No commit from rejected [#1812](https://github.com/AdaWorldAPI/lance-graph/pull/1812), no production APIs or cross-license code transplant. EXO-50 coordination message ≠ delivery/ACK.

## Handoff to EAW-10
Only after objective chess temporal blind gates, challenge the same framework with a *tiny, fully specified rule AST*: alternative support, conjunction, one qualified exception, effective dates and typed routing. Its law source/authority is separate; the chess experience supplies a **testing methodology**, not transferred legal correctness.