Structural terminal veto

A safety-first way to use the family's structural planner: never let it choose a move, only let it object to one — and only when two independent panels of imagined futures agree.

The intuition

By this point in the research the pattern was established: the constructive planner knows something about long-term structure, and letting it override the four-move search damages the search. The natural response is not to abandon the signal but to make the override rare and expensive to earn.

So: the four-move search picks the move. Its runner-up is allowed to replace it only if a long simulation says the runner-up survives materially longer — and only if that conclusion holds up on two separate sets of imagined futures, generated from independent random streams. A conclusion that appears on one panel and not the other is exactly the kind of noise that made earlier overrides unstable, and the two-panel rule is designed to catch it.

The word "terminal" here means the simulation runs to the end of the game (or a long horizon) and cares about when the game ends, rather than about points collected in the next few moves.

How it works

1. Run the unmodified full-width four-move search with five chance strata. It supplies the chosen column and the complete ordering of the others.
2. Take only its top two columns. If there is only one legal column, or the search is decisive, nothing else happens.
3. Force each of the two columns, then continue with the frozen constructive continuation — the inner rollout policy, deliberately not its depth-shielded outer form — over a 200-move horizon.
4. Do that on two independent panels of 127 event-stratified public chance streams each.
5. Replace the search's choice with the runner-up only if a fixed terminal-classifier gate passes independently on both panels.

Everything read is public. The search is never modified: with the veto disabled the program reproduces exact four-move play.

What happened

Rejected. The recorded verdict is that the aggregate fitting result trailed plain four-move search and could not reach its gate — so no screen was opened.

There are no retained numbers. The status exists only as a task record; audit 04 §B.3 lists this source among the thirty rejections that cannot be re-derived, with no ledger protocol, no artifact hash and no per-game data. This page states the verdict as the experiment index words it and quotes no scores.

The wider pattern it belongs to is well recorded. The strategy landscape collects the whole family of conservative overrides — long-outcome vetoes, denoised-value vetoes, root tail-risk objectives, regenerative panels — and concludes that the gates themselves work mechanically, while the challengers behind them were "too rare, poorly calibrated, or unstable across whole games". Its verdict is worth reading before designing another one: confidence gates are a safety mechanism, not a source of new strategic information.

The technical record

Status: rejected — task-record only. From the experiment index: "the aggregate fitting result trailed plain D4 and could not reach its gate." No entry in the experiment history.

Configuration, from the source: fitting seeds from 0x3d6e4000, 4 games; screen seeds from 0x3d6e5000, 8 games, never opened; 1,000-move cap; two panels of 127 scenarios; 200-move horizon; 64 events per step; separate random domains for panel A, panel B, reveals and visible discs. The frozen gates required, among other things, a 99% lower bound of at least 10,000 points and 2 moves, no material loss beyond −100,000 points or −25 moves, at most a 10% upper bound on downside, ratios of 1.10 on the fitting quartet with 3 joint wins and 1.05 on the screen with 5 joint wins.

Note the size of the fitting stage: four paired games. Those same four seeds are the ones the D4 flow audit is licensed to replay. A four-game aggregate can show that a candidate is not promising; it cannot establish by how much.

Source: d4-structural-terminal-veto.cpp.

What this taught us, and what is still open

• Requiring two independent panels to agree is a sound idea and it did its job: a weak challenger did not reach a screen.
• A gate cannot manufacture a signal. The underlying comparison — constructive continuation versus four-move search on near-tied columns — is the thing that was not good enough, which is the same conclusion the horizon experiments reached.
• Nothing about the size of the shortfall is recoverable from this repository. Re-running it would be a new experiment needing a new lease and a new protocol.