---
title: Public regenerative B0
family: terminal-policy-iteration
summary: Eliminate weak columns cheaply, test the survivors against the reference, and override it only with strong evidence.
status: rejected
evidence: ledger-recorded
reads: public
---

Spend the measurement budget where it matters: eliminate weak columns cheaply,
test the survivors hard against the reference, and override the reference only
on proof.

<EvidenceLabel status="rejected" evidence="ledger-recorded" reads="public" />

## The intuition

The family's earlier programs gave every column the same expensive treatment.
That is wasteful: at most positions, four or five of the seven columns are
obviously poor, and the entire question is which of the remaining two or three
is best. So run the comparison as a tournament instead of a survey: a short
look at everything, a longer look at the survivors, and a final independent
test between the winner and the reference's own move.

The word "regenerative" refers to the cycle structure of the game. The board
rises every five moves, and a policy is really trying to arrive at each rise in
a state it can recover from. The staged horizons here — 25 moves, then 50, then
75 — are five, ten and fifteen of those cycles.

There is one more piece of discipline, and it is the reason this experiment is
worth reading even though it failed. Each stage uses **independent** chance
scenarios for the decision it is making. Reusing the same imagined futures to
first choose a winner and then confirm it would let a column that got lucky in
the first round keep that luck in the second; independent panels make the
confirmation a valid test.

## How it works, step by step

1. At a stored public position, evaluate **every** legal column under seven
   shared 25-move continuations.
2. Keep the three best challengers, plus whatever the exact four-move reference
   would play, and re-evaluate all of them under 21 nested 50-move
   continuations.
3. Put the survivor through an independent 21-scenario, 75-move panel. It may
   be admitted only if it beats the reference on *all three* of a paired return
   lower bound, the number of discs cleared, and the number of covered discs
   revealed.
4. Compare the admitted column with the reference on one final, independent
   35-scenario, 75-move panel.
5. Every continuation is played by a completed, full-width public two-move-ahead
   search; the exact four-move reference remains the fallback whenever no
   challenger is admitted.

The policy is handed only the board, the visible next disc and the rise phase.
Origin, score, history, scenario identity and any future disc or reveal are
excluded from it; origin identity is used only by the coordinator, to split the
positions into eight whole-game folds.

## What happened

It came out very nearly identical to simply keeping the reference, and the
"very nearly" is the result.

Out of 477 positions, the whole apparatus changed the reference's mind eleven
times. Of those eleven changes, six helped. Score and survival ended up about a
tenth of a percent above the reference, which sounds like a tie and, once the
confidence bounds are computed, is worse than a tie: the lower bounds on both
the score advantage and the lifetime advantage were **negative**, meaning the
data cannot rule out that the overrides made things worse. Its clear and reveal
rates were a hair below the reference's. When its own scenario halves were
compared, it agreed with itself about half the time. Five of eight origin games
did not regress, where six were required.

Every frozen gate except the ordered-halves check failed, and the experiment
was recorded as a valid negative. The ledger's own reading is worth repeating
because it resists the comfortable explanation: the missing signal is score,
survival and flow, not merely commitment stability. The
problem is not that the method was too timid about overriding. It is that when
it did override, it had not actually found anything better.

Two limits belong with this. This is an offline root-panel proxy: only the root
action was improved and every subsequent action reverted to the two-move-ahead
policy, so it is not a complete-game result. And the 477-position panel had
already been read by earlier work, so it is diagnosis rather than fresh
model-selection evidence.

<TechnicalDetails title="The technical record">

Rejected, ledger-recorded. It opened only the locked, already-consumed corpus
of 477 public roots from eight whole-game origins; the stored H200 outcomes
were not used for selection, no origin game was replayed, and zero gameplay,
protected or final seeds were opened.

The frozen gate required at least 70% split stability, 5% override coverage,
75% beneficial-override precision, 1.10× D4's raw score, 1.05× its restricted
mean survival, positive one-sided 95% whole-origin lower bounds for utility and
moves, non-regressing clear and reveal flow, at least six of eight
non-regressing origins, and non-regression in both ordered halves.

| Measure | Candidate | Gate or comparator |
| --- | ---: | --- |
| Overrides | 11 of 477 (2.3061%) | ≥ 5% coverage |
| Beneficial overrides | 6 (54.5455%) | ≥ 75% |
| Score ratio vs D4 | 1.000718 | ≥ 1.10 |
| Restricted mean survival ratio | 1.000621 | ≥ 1.05 |
| Mean utility delta | +695.637 (lower-95 −169.897) | lower bound > 0 |
| Mean move delta | +0.02660 (lower-95 −0.01365) | lower bound > 0 |
| Clears / reveals per move | 1.966311 / 1.053280 | D4: 1.966329 / 1.053383 |
| A1/A2 split stability | 49.6855% | ≥ 70% |
| Non-regressing origins | 5 of 8 | ≥ 6 of 8 |
| Ordered halves | both passed | pass |

Cost: 2,509,342 synthetic transitions, 2,436,507 D2 calls, 4,625,296,860 D2
work units and 572,887,018 D4 work units, in 1,867.000 effective four-thread
root seconds at 35,831,808 bytes peak RSS. Source, result, checkpoint and
preflight SHA-256 values are all recorded. The ledger adds a requirement for
any successor: bind a combined transitive-source hash rather than only the
top-level source hash, before a passing result can be trusted.

Source: "Public regenerative policy-iteration B0 (rejected)" in the
[experiment history](/docs/research/history); the terminal-policy-iteration
table in the [experiment index](/docs/research/experiment-index).

</TechnicalDetails>

## What this taught us, and what is still open

- **"Statistically indistinguishable from the reference" is a real result.**
  A staged, careful, well-instrumented comparison could not find eleven
  reliable improvements in 477 positions. That constrains how much easy
  headroom sits next to the reference's choices.
- **Coverage and effect size have to be reported together.** An override rule
  firing at 2.3% of positions cannot move a mean even if every firing were
  right, and only 55% of them were.
- **The staging machinery is reusable.** Cheap elimination followed by
  independent confirmation panels is a sound design; what it was given to
  measure, a two-move-ahead continuation over 75 moves, is the part that did
  not carry enough signal.
- **Still open:** the same architecture with a stronger continuation policy, on
  fresh successor-closed data rather than a panel that has already been read.