State curriculum

Keep the deep, mature boards that a future-reading planner reaches after fifty or a hundred moves, strip away everything privileged about how it got there, and hand the bare positions to other experiments as practice starting points.

This is infrastructure, not a policy. Nothing here plays a game to win one; it manufactures a dataset that several later experiments consumed.

The intuition

A learner that always starts from an empty board spends nearly all of its experience on easy, early positions. The interesting decisions: the ones that decide whether a game reaches move 200 — happen on crowded boards with tall columns, buried gray discs and a rise about to land. Ordinary play reaches those positions rarely and briefly; the perfect-information oracle reaches them constantly and survives inside them for hundreds of moves.

So: use the oracle as a tour guide to interesting positions, then delete the guide. What is exported is a board, not a trajectory.

Where the privilege stops

The C++ converter states the rule in one line: oracle privilege terminates at the public-state constructor. An exported record contains a canonical board, the visible next disc, and the five-drop rise phase. It does not contain the seed, the future tape, the hidden values behind any gray disc that was still covered, the score, the level, or the move index.

More than that, the diagnostics attached to each record are not the oracle's own outcome. Each exported position is re-played forward under seven independent futures derived from the public state itself, using a fair depth-1 search, over a 25-move horizon. So the numbers stored next to a position — mean moves survived, survival rate, clears and reveals per move, and a coarse flow band — describe what an ordinary player can expect from that board, not what the cheating planner achieved from it.

How it works, step by step

1. Play 64 games with the privileged beam planner (depth 4, beam 128, 500-move cap) on a training-only seed family.
2. From move 50 onward, sample positions along those games.
3. Convert each to the public encoding, canonicalize it by reflection, and deduplicate, keeping at most 4,096 states.
4. Re-score each state under seven public-derived futures with fair depth-1 play, at a 25-move horizon, and attach the flow diagnostics.
5. Write the file. A separate TypeScript generator produces a sibling format that tags each record by where it came from — oracle play, a combined policy, or a safety policy, and stores remaining moves and remaining score as targets.

What happened

The generator worked and its isolation checks passed: 4,096 restart states were produced, and the privilege-isolation tests confirmed that nothing privileged crossed into the exported records. That is the whole retained result. It is a support artifact, recorded as task-record only, with no ledger entry.

What the curriculum was used for is where the outcomes are, and those are all negative so far: a value network trained on these states plateaued well below fair search; a reinforcement learner that began half of its episodes from these positions failed every one of its absolute floors; an option-based policy learner using the same 50/50 mix improved on the shallowest search and stayed far below depth-4; and the curriculum long-outcome network predicted survival usefully while ordering moves worse than exact search. None of those failures is attributed to the curriculum itself; none of them vindicates it either.

Technical recordopen ↓close ↑

• Native converter, oracle-curriculum.cpp: completed; task-record only support; "4,096 restart states were generated and privilege-isolation checks passed" (experiment index). No ledger entry exists for it.
• TypeScript generator, generate.ts: support-only; repository-verified. Emits D7CURR1 records tagged by source (oracle, combined, phase-safety) with remaining-moves and remaining-score targets.
• Read from the source: 64 generation games from 0x3d660000, oracle depth 4 / beam 128, 500-move cap, first sample at move 50, at most 4,096 states, seven restart scenarios at horizon 25, separate reveal and visible-disc random domains for the restart futures, and a static assertion that the level bonus is 17,000, so the restart diagnostics are on corrected Hardcore scoring.
• Downstream consumers and their recorded outcomes are listed in audit-05 §2.1 and §2.4; the corpus fingerprints appear there too. §2.5 of the same audit records that the generated corpora themselves lived under /tmp and are gone from this machine, so any experiment that wants them must regenerate them from a leased training range.

What this taught us, and what is still open

• Isolation can be made structural. Terminating privilege at a constructor, and recomputing every diagnostic from public-derived randomness, is a pattern worth copying: it makes the information boundary a property of the type system rather than a promise in a comment.
• A dataset is not a result. Four thousand positions exist, were used four or five times, and have not yet produced a policy improvement.
• Still open: the one-hour measurement above; and, separately, whether a training pool drawn from oracle-visited boards but labelled fairly helps — which audit-05 identifies as untested in this repository, as distinct from the optimistic labels that have already failed.