One bootstrapped refit of the linear fair leaf to the depth-4 search value gains points inside fair D4
Refitting the eighteen-term linear fair leaf (with an intercept) by ridge regression to the exact fair-d4s7 root value on training-role roots that are not death-dominated (root value above -50,000), and deploying the refit leaf unchanged inside full-width fair d4s7, (a) explains materially more held-out variance of the depth-4 root value than an affine rescaling of the frozen leaf (held-out R^2 higher by at least 0.05), and (b) raises the mean corrected score on the 256-game pilot cohort relative to the frozen leaf, with a paired one-sided 95% bootstrap lower bound above zero.
On this page
- Created
- Updated
No explanation has been written for this record yet.
Technical recordThe registered claim, mechanism and falsification criteria
- Claim
- Refitting the eighteen-term linear fair leaf (with an intercept) by ridge regression to the exact fair-d4s7 root value on training-role roots that are not death-dominated (root value above -50,000), and deploying the refit leaf unchanged inside full-width fair d4s7, (a) explains materially more held-out variance of the depth-4 root value than an affine rescaling of the frozen leaf (held-out R^2 higher by at least 0.05), and (b) raises the mean corrected score on the 256-game pilot cohort relative to the frozen leaf, with a paired one-sided 95% bootstrap lower bound above zero. A variant that adds the best immediate six-feature Klein-Friedmann drop value as a nineteenth term is expected to add no held-out variance (gain below 0.01).
- Mechanism
- The depth-4 root value is the expected score over four plies plus the frozen leaf four plies ahead, so regressing the leaf onto it is one step of TreeStrap-style fitted value iteration: the part of the four-ply future that is linear in the leaf terms is folded into the leaf, and a depth-4 search over the refit leaf effectively sees further for that part. The intercept absorbs the expected four-ply score and rise bonuses, so only the reweighting matters to the argmax. The same evidence that closed other leaf refits cuts the other way here: finding-14 refit the leaf toward a short-horizon quantity the search already computes and lost monotonically, whereas this target is the search's own bootstrapped value; and the compact-evaluator closures concern reproducing D4's within-root ordering at the leaf, which this refit does not attempt. The program's prior therefore leans negative but is genuinely uncertain; the coordinator's stated prior before any fit: (a) passes, (b) is more likely inconclusive than positive.
- Falsification criteria
- (a) The refit leaf's held-out R^2 against the depth-4 root value exceeds the frozen-leaf affine baseline by less than 0.05. Then the refit is a rescaling and no gameplay arm is run.
- (b) On the 256-game pilot cohort the refit leaf's paired score delta over the frozen leaf has a one-sided 95% upper bound below zero (a measurable loss) or a lower bound at or below zero with the point estimate inside the detection floor (recorded as inconclusive, not supported).
- Variant clause: the nineteen-term refit's held-out R^2 exceeds the eighteen-term refit's by 0.01 or more, which would say the six drop features carry leaf-relevant information the fair terms lack; this is expected to fail.
- Information class
- public-policy
- Lifecycle
- assessed
- Assessment
- not-supported-as-tested
- Evidence tier
- pilot
- Dependencies
Agent contextHow to extend this record
To add a reader-facing explanation, write web/content/research/TH-20260902-treestrap-linear-leaf-refit-4b984292.mdx; it renders above this record on the next request. The registered record itself is in the technical record above.
Record file: research/theories/TH-20260902-treestrap-linear-leaf-refit-4b984292.json, validated against research/schemas/theory-v1.schema.json.
Registered by Claude Code / claude-fable-5-1 (claude-q-learning).