Finding 14 — reweighting the fair leaf toward achievable clears

Status: exploratory, evidence tier development. Negative result, with a measured mechanism and a dose-response across six arms. Built and measured in this checkout on 2026-08-20/21.

Namespace: approaches/lifetime-objective/leaf-reweight/, run runs/RUN-A527-LEAFW/, seed lease SEEDLEASE-A52-LEAFW = 0xa5278000–0xa527ffff for every tuning, CHECK and screening game. Headline results are on the shared evaluation cohort 0xa51d1000–0xa51d103f (64 paired games), which was never tuned on.

No existing file was created or modified by this work. New files appear only under approaches/lifetime-objective/leaf-reweight/, docs/exploratory/, runs/RUN-A527-LEAFW/ and build/lifetime/.

Why this exists

finding-10 Addendum A labelled 1,024 public positions with their exact 8-move achievable clear optimum, averaged over independent completions of both the hidden board and the future, and asked what predicts that label:

Reweighting recovers 95% of all the signal any structural property can supply, and the cosine between the frozen weight direction and the predictive direction is +0.141. The conclusion drawn there was that the leaf is not missing information — its weights point almost orthogonally to it.

That is a statement about a diagnostic label. This document tests the different and harder claim the programme actually needs: that a leaf which predicts achievable clears better is a better leaf to search with.

It is not. Over 64 paired whole games at depth 4 with seven chance strata, rotating the leaf toward the fitted direction degrades score, lifetime, numbered clears per move, covered reveals per move and mean occupancy monotonically. The full refit loses 237,182 points (one-sided 95% bound −290,406, 7-0-57). The best arm — setting kRoughnessWeight from 0 to +560, Addendum A's first-ranked and cheapest candidate — is an exact tie at +627 (95% bound −73,325, 31-0-33). The frozen weight vector is the best of the six.

1. CHECK gate — the weights become data and nothing else changes

The candidate is approaches/lifetime-objective/leaf-reweight/search.cpp. Its search driver is a copy of approaches/lifetime-objective/risk-calibration/search.cpp, which finding-05 already proved decision-identical to the frozen reference. The only change is that frozen::fairLeaf(state) is replaced by a local dot product over frozen::extractFairFeatures(state) with the nineteen weights held in a struct. The feature extractor itself is the unmodified frozen one, called by name, so the only degree of freedom this program introduces is the weight vector. The accumulation order of the dot product is preserved term for term, because floating-point addition is not associative and bit-identity depends on it.

Five gates, all passed, all on the tuning lease:

936,612 real boards compared, zero differing bits. 481 decisions compared, zero differing columns and zero differing work counts.

A fifth, cross-binary gate closes the loop on the comparator. The frozen-weight arm at the evaluation configuration is runs/RUN-A51D-s7confirm/fresh-s7.json, produced by a different binary (approaches/lifetime-objective/risk-calibration) on the same 64 ordered seeds at depth 4, seven strata, 16,000,000 work. Re-playing the first eight of those games with this binary at frozen weights reproduces them exactly:

8 games × 11 recorded fields — score, moves, work, numbered clears, covered reveals, rises, level points, chain points, maximum chain depth, mean occupied cells, censor flag — 0 mismatches.

Games are deterministic given a seed and a policy, so fresh-s7.json is used as the paired comparator for all 64 games rather than spending another 3.8 hours recomputing a file this binary provably produces. That file's headline figures are the ones the coordinator supplied as reference points and they match exactly: 398,498 mean, 114.66 mean moves, 2.0571 clears per move, 1.1549 reveals per move, 23.15 mean occupied cells.

The work-count identity is the part that matters beyond column identity: it shows the cache, the iterative-deepening ladder and the work budget all behave identically, so a work-limited decision degrades the same way. Nothing past this gate was run before it passed.

2. What the fitted direction is, once it is an actual leaf

Addendum A reports the fitted direction in leaf units per one sigma of each feature, normalised to unit length. That is a direction, not a leaf. A search needs raw weights whose scale is commensurate with the −1,000,000 terminal utility and with the 1.0 coefficient on immediate score, because — per audit-02 §5 — the ranking is effectively lexicographic: one sampled death costs 18×–920× the entire legitimate discrimination range between actions, so the search minimises modelled four-ply death first and maximises the leaf second. Shrinking or growing the leaf therefore changes how much modelled death risk the search will accept. That is a different experiment from changing the direction the leaf points in.

refit.py holds that second axis fixed:

1. fit standardised OLS of the target on the leaf's own 19 features, on the same 60% content-hash training split Addendum A used;
2. normalise the fitted per-sigma vector to unit length → u, and the frozen vector expressed per sigma to unit length → f;
3. interpolate, d(α) = unit((1 − α)·f + α·u);
4. convert back to raw weights w_i = d_i · S / σ_i, choosing S so the resulting leaf has the same standard deviation over the corpus as the frozen leaf_value; and
5. add a bias so it has the same mean.

Every Tier-2 arm therefore differs from the frozen leaf in direction only. The bias is a nineteen-term dot product's twentieth parameter; it defaults to exactly zero, so the CHECK gate is unaffected.

The refit reproduces Addendum A exactly, which is the check that this pipeline is reading the same thing: held-out R² 0.7342 for the reweighted 19 features, 0.3962 for the frozen scalar, cosine +0.1406.

Tier 1's anchors come from the same rescaling applied to a single term: the raw weight a same-scale fitted leaf would use for that feature alone.

Substituting all four at once moves the leaf's standard deviation from 32,762 to 34,526 on fair-play positions — a 5% change in scale, so the Tier-1 arms are close to pure direction changes as well.

3. Risk 3 first: how much of the fitted direction is a synthetic artifact

Addendum A A.6 warns that its added cover-geometry block earns its keep off the fair-play manifold: +0.054 R² on synthetic positions but +0.005 on fair-play positions, and −0.015 on lowest-column positions for the gap target. The same question applies to the reweighting itself, and it has to be answered before any gameplay number is interpreted. tier1.py refits separately by position origin and scores every fit on every origin's held-out positions.

Held-out R² on achievableClears, fit on the row origin, scored on the column origin. No position both trains and scores.

Two things in that table change how Addendum A's headline should be read.

First, the headline 0.396 is a mixture number dominated by the synthetic block. On synthetic positions the frozen leaf scalar explains 0.7% of the variance; on the positions a fair policy actually visits it explains 61.8%. The gap the reweighting has to close on the fair-play manifold is +0.168 (0.618 → 0.786), not +0.338. The leaf is a much better predictor of achievable clears than 0.396 suggests, where it is actually used.

Second, the fitted directions genuinely disagree with each other by origin.

The fair-play fit and the synthetic fit point in directions only +0.310 apart — barely more aligned with each other than either is with the frozen leaf. The all fit sits between them at +0.758 and +0.820, so it is a blend of two different answers, weighted 448/1024 toward synthetic positions. The synthetic fit transfers to fair-play held-out positions at R² +0.214, worse than the frozen leaf's own +0.618.

Consequences carried into the arm design:

• the fair-play-only refit is the primary Tier-2 vector, not Addendum A's all fit, because it is the one whose training distribution matches where the leaf is evaluated;
• the fair-play fit is also more aligned with the frozen leaf (+0.176 vs +0.141), so "almost orthogonal" is slightly overstated on the manifold that matters; and
• numbered_cells — Addendum A's most striking sign flip, +0.601 per σ and "the largest fitted term" — is +0.147 per σ on fair-play positions. Most of its size is a synthetic-generation artifact: the synthetic generator sweeps occupancy deliberately, giving numbered_cells a corpus σ of 6.04 against 2.30 on fair-play positions. The sign is still positive both univariately and conditionally on fair-play positions, so the direction of the diagnosis survives; its magnitude does not.

Risk 3 is therefore not hypothetical, and it is not fully avoidable either: even the fair-play block is 384 positions harvested from d2s5 play, not d4s7 play.

4. The screening proxy, and why it is not used as a gate

Depth 3 at seven strata costs about 1/32 the work of depth 4 at seven strata, so it is tempting for screening. It is also exactly the "cheap proxy" gamble that finding-10 §1–§7 just lost: a scenario benchmark scored over nine moves ranked search configurations backwards, because over a short window the score is dominated by a term that is a function of the rise phase rather than of skill.

The proxy cannot be validated here, and one known ordering already refutes it. finding-05 measures the chance-strata effect at +101,171 (95% bound +47,457) at depth 4 and +7,276 (95% bound −45,961, not significant) at depth 3. Depth 3 demonstrably fails to reproduce a depth-4 ordering for that factor. There is no leaf-weight ordering at depth 4 already established in this repository against which depth 3 could be validated, so no validation was possible in advance.

Depth 3 was therefore used only as a response-surface map on the tuning lease, to choose magnitudes within a family, and every verdict-bearing number in §5 is a direct depth-4 run. This is the "screen a small number of arms directly at depth 4 and say that you did" branch.

The map, on SEEDLEASE-A52-LEAFW seeds 0xa5279000–0xa527903f, 64 paired games, depth 3, seven strata, --max-work 16000000. The run was stopped after 8 of 25 planned arms to release the machine for the depth-4 evaluation; arms execute in a fixed order, so what survives is a prefix — the entire roughness family and the first two solid_exposure points.

Two things are worth reading off it before the depth-4 arms are seen.

roughness is monotone in the wrong direction, and the mechanism is visible. The largest weight tested loses 129,554 points and 35 moves. Clears per move fall monotonically 2.0042 → 1.8384, reveals per move 1.1119 → 1.0077, and mean occupancy rises 23.66 → 25.33. Rewarding a rough board makes the policy worse at the two conservation quantities that finding-01 shows are what ends games. This is the opposite of the mechanism Addendum A proposed — that a rough board offers more distinct run lengths for an arriving disc to match. Every one of the five roughness points is negative, so this is a dose-response relationship, not a single unlucky draw.

solid_exposure is the one Tier-1 term that might be real, and its optimum is below the fitted value. +1,250 gains +35,988 with clears and reveals both up, but the 95% lower bound is −25,099 and the split is 32-1-31, so this is a direction to test at depth 4, not a result. +2,500 is already back to a tie, and the fitted anchor is +4,934 — the fit wants roughly four times more weight than the best point measured here.

The proxy inverted, and §5 measures it

Both readings above are stated as depth-3 observations because §5 then falsifies one of them outright:

At depth 3, setting kRoughnessWeight to +560 is one of the largest losses in the whole document — a confident, dose-responsive, mechanistically coherent loss, with the flow rates and occupancy all moving the way a real effect would move them. At depth 4 the same one-line change is an exact tie that slightly improves both flow rates and raises the cohort minimum by 16,898.

Nothing about the depth-3 evidence looked unreliable, and it was wrong anyway. This is a second, independent instance of the failure finding-10 §1–§7 documented — one more reason to treat a cheap surrogate for a whole-game ranking as a hypothesis generator only, no matter how internally consistent it looks. It also retrospectively justifies not using depth 3 as a selection gate: had the arms been screened at depth 3, roughness would have been discarded as the worst candidate rather than reported as the best of five.

It leaves solid_exposure +1,250 — the depth-3 best — genuinely unknown at depth 4. It was queued and not run (§8), and it is the recommended next arm precisely because the depth-3 estimate for it cannot be trusted in either direction.

5. Results — depth 4, seven strata, the shared evaluation cohort

Cohort 0xa51d1000–0xa51d103f, 64 paired whole games, 2,000-move cap, depth 4, seven chance strata, --max-work 16000000, 12 threads. Comparator as in §1.

Bounds are one-sided 95% percentile bootstrap over 20,000 resamples of whole games, paired by seed, Mulberry32 seeded 0xb0075eed. No arm was censored, no arm recorded a score-identity violation, and no arm achieved a board clear.

Four of five arms lose, three of them by more than their own 95% bound. The fifth — roughness alone — is an exact tie: +627 points on a 254,414-point standard deviation, 31-0-33.

The dose-response, which is the actual result

Express each arm's weight vector in leaf-units-per-sigma and normalise, exactly as §2 does, then measure how far each arm has travelled from the frozen direction toward the fitted one:

Score, lifetime, numbered clears per move, covered reveals per move and mean occupancy are monotone in how far the leaf has been rotated toward the fitted direction — six arms, five columns, one exception. The exception is t1-rough-560, and it is the exception that confirms the shape: setting kRoughnessWeight to +560 moves the leaf's direction by 0.005 of cosine, which is the smallest dose in the table, and it produces the smallest effect in the table. Roughness is the only feature the frozen leaf multiplies by exactly zero, but in σ-scaled direction terms it is a small term, and giving it the weight the fit wants barely rotates the leaf at all.

This is the sharpest available statement of the result, and it is not a null. The fitted direction predicts the exactly-computed 8-move achievable clear optimum far better than the frozen leaf does — 0.786 against 0.618 held-out R² on fair-play positions, and 0.734 against 0.396 on Addendum A's mixed corpus. Rotating the search's leaf toward it makes the policy worse at actually clearing discs: 2.0571 → 1.6019 clears per move, a 22% fall, moving away from the 2.400 that disc conservation requires, and moving mean occupancy from 23.15 to 28.34 cells and away from the ~18 at which finding-07's clairvoyant planner operates. Predicting where clears are available and being able to sustain them are close to opposite objectives for a search leaf.

Work per move stays within 8% of the comparator across every arm (4.58M–5.05M against 4.96M), so no arm silently degraded to a shallower completed depth and none of this is a work-limit artifact.

Tier 1, term by term

• roughness 0 → +560: a tie, +627 (95% bound −73,325), 31-0-33. Addendum A A.7 ranked this first — "changing one constant is the entire experiment", "the cheapest experiment in the repository". It is not an improvement, but it is also the only arm that is not a loss, and it is the only arm that improves both flow rates (2.0667 clears, 1.1611 reveals) and the minimum (+16,898) and the maximum (1,503,379, the largest single game anywhere in this document) at once. Its variance is higher than the comparator's in both directions. A term the frozen search has multiplied by zero since inception turns out to be worth approximately zero.
• solid_exposure +40 → +4,934 with cracked_exposure +100 → +2,868: −67,719 (95% bound −133,082). These are the two terms the fit wants most after the height block, and solid_exposure has the strongest univariate correlation with achievable clears of all nineteen (+0.627). At the weight the fit wants, they cost 17.3 moves of life. The depth-3 map (§4) found the best point at roughly a quarter of this weight, so the failure here is one of magnitude as well as of direction.
• All four together: −75,214 (95% bound −139,902). The bundle is slightly worse than t1-exposures alone and slightly better than the sum of its parts would suggest.
• numbered_cells −18 → +1,048 was not run alone (§8).

Tier 2, the full refit

• Half way, t2-fair-a05: −107,107 (95% bound −169,589), −28.78 moves.
• All the way, t2-fair-a1: −237,182 (95% bound −290,406), −63.91 moves, 7-0-57. The mean falls below lowest-column-plus-a-bit territory: at 161,316 it is nearer the 100,050 of the trivial lowest-column control in finding-01 than to the comparator.

Interpolation was the right precaution and it did not rescue anything. There is no interior maximum: the response is monotone from the frozen vector to the fitted one, so no blend of the two beats the frozen vector alone.

Lower tail — and it is not what risk 2 predicted

The eight worst games of the comparator, paired:

Every arm improves most of the comparator's worst eight games, several of them by three or four times, and t1-rough-560, t1-combined and t2-fair-a05 all improve the cohort minimum. What the losing arms destroy is the upper half: the maximum falls 1,341,287 → 989,981 → 977,457 → 903,444 → 342,164 down the ladder, and the standard deviation falls 254,414 → 54,703. Q25 is flat for the Tier-1 arms (−1,424, −555, −6,884) and only collapses at Tier 2 (−19,438, −91,250).

The reweighted leaves are not dying earlier in their bad games. They are failing to build the long ones. Given that mean score is 3,400 × mean lifetime (finding-01), and that the comparator's mean is carried by a heavy right tail — two games over 1,000,000 in 64 — losing the tail is losing the mean.

6. The three risks, answered

Risk 1 — the fitted target is a proxy, over a short horizon

Verdict: confirmed, and it is the whole story. finding-10 §1–§7 was bitten by exactly this once already, and the warning was correct a second time. The label is the exact 8-move achievable clear optimum. Eight moves is one rise cycle and a bit. A leaf that maximises it is a leaf that maximises what is reachable now, and the measured consequence is a policy that clears less per move over a whole game and carries more discs while doing it.

The mechanism is legible in the flow rates rather than inferred. Rotating toward the fitted direction raises the weight on numbered_cells from −18 to +1,957 and on solid_exposure from +40 to +9,216. Both say "a board with more numbered discs and more nearly-poppable covers is a good board" — which is true of this move's opportunity and false of the next fifty, because those discs are also the occupancy that ends the game. Mean occupied cells rises monotonically 23.15 → 25.42 → 25.97 → 28.34 as the rotation proceeds. R² against an 8-move label carries no information about that trade, and it is the trade that decides the game.

The correct reading of Addendum A is therefore the one it gave itself: a hypothesis generator. It generated four; all four were tested; all four were rejected at depth 4. That is what a hypothesis generator is for.

Risk 2 — the leaf does not act alone

Verdict: real, but not the failure mode here, and the evidence says so specifically. audit-02 §5 predicts that a reweighting could destroy the death-avoidance keeping games alive, because one sampled death costs 18×–920× the legitimate discrimination range between actions and the ranking is effectively lexicographic. If that were what happened, the arms would show a collapsing lower tail with the upper tail intact.

The opposite is observed at moderate strength. t1-combined and t2-fair-a05 both improve the minimum (+51, +122) and improve most of the comparator's eight worst games, while the maximum falls from 1,341,287 to 977,457 and 903,444 and the standard deviation falls from 254,414 to 190,684 and 159,133. The reweighted policies die less badly and win less well.

Two design choices are what let this be said rather than guessed. Every Tier-2 arm is sigma- and mean-matched to the frozen leaf over the corpus (§2), so the leaf-to-terminal-utility exchange rate is held fixed and the arms vary direction only; and the CHECK gate proves identical work accounting, so no arm is quietly searching a different depth. Without the scale match, a wholesale weight swap would have changed the risk attitude and the direction at the same time and this question could not have been separated.

At t2-fair-a1 the tail does degrade (minimum −435, Q25 −91,250). But the distribution has collapsed to sd 54,703, a fifth of the comparator's: this is a policy that reliably reaches about 50 moves and reliably stops, not one that dies unpredictably early. Death-avoidance survived the reweighting; long-game construction did not.

Risk 3 — the fitted direction may be a synthetic-position artifact

Verdict: confirmed as a genuine defect in the diagnostic, and corrected for before the arms were built — which makes the negative result stronger, not weaker. §3 measures it directly: the synthetic-position fit and the fair-play fit point in directions only +0.310 apart, and the synthetic fit transfers to fair-play held-out positions at R² +0.214, worse than the frozen leaf's own +0.618. Addendum A's headline numbers — 0.396 for the frozen scalar, cosine +0.141 — are mixture statistics over a corpus that is 44% synthetic, and on synthetic positions the frozen leaf explains 0.7% of the variance while on fair-play positions it explains 61.8%.

Two specific claims in Addendum A do not survive the restriction to fair-play positions:

• the leaf is "nearly orthogonal" to the predictive direction. On the manifold where it is evaluated the cosine is +0.176, and the R² gap the reweighting has to close is +0.168 rather than +0.338;
• numbered_cells is "the largest fitted term" at +0.601 per σ. On fair-play positions it is +0.147 per σ. Most of its apparent size comes from the synthetic generator deliberately sweeping occupancy, which gives the feature a corpus σ of 6.04 against 2.30 on fair-play positions.

Every Tier-1 anchor and both Tier-2 vectors tested here were built from the fair-play-only refit, which is the strongest version of the hypothesis available. It still lost, monotonically, at every dose. The result is not "the fit was contaminated so it failed"; the fit was decontaminated first and failed anyway.

The residual caution belongs to this document too: the fair-play block is 384 positions harvested from d2s5 play, not from the d4s7 policy actually being modified. A refit on d4s7-harvested positions is not tested here.

7. Verdict

Validity: valid. All five CHECK gates passed before any gameplay; the comparator is a proven bit-identical reproduction; every arm played the same 64 ordered public-information games at a declared fixed work bound with identical work accounting; no censoring, no identity violation, no illegal move.

Outcome: fail — a clean negative, with a mechanism. Reweighting the frozen leaf toward the direction that predicts the exact 8-move achievable clear optimum does not improve whole-game play. It monotonically harms it. The best arm is a tie (+627, 95% bound −73,325); the full refit loses 237,182 points (95% bound −290,406, 7-0-57).

Evidence tier: development / STANDARD-shaped. 64 paired whole games per arm on the shared evaluation cohort, which is reusable development data. This rejects the exact tested configurations; it is not a qualification claim, and no protected or final seed was opened or is justified by anything here.

What was actually learned, stated as a positive

The interesting result is not "the reweighting failed". It is what the failure says about what a search leaf is for.

audit-02 §4.4 concluded that the frozen leaf is "a reasonable proxy for the right objective (survival), by accident of its weight vector rather than by construction". Addendum A then measured that this weight vector is nearly orthogonal to the direction that predicts where clears are available, and read that as a defect. This document tested the repair and found that the frozen weight vector is well tuned for play precisely because it does not predict achievable clears.

The mechanism is visible in the flow rates rather than inferred. The frozen leaf spends 78% of its σ-scaled direction on covered_height_risk and rise_pressure — two quantities that carry almost no conditional information about how many clears are available now, and that are exactly the crowding pressure which decides whether the board is still playable in thirty moves. The fitted direction moves that weight onto numbered_cells, solid_exposure and cracked_exposure, which say "a board with more numbered discs and more nearly-poppable covers is a good board". That is true of this move's opportunity and false of the next fifty, because those same discs are the occupancy that ends the game. Measured consequence: occupancy rises monotonically 23.15 → 28.34 cells and clears per move fall 2.0571 → 1.6019.

So the accident in audit-02's phrasing may not be an accident. A leaf that is "a crowding evaluator" is a leaf that estimates future capacity to clear, and that is a different and more useful quantity than present availability of clears, which four plies of exact search can already see for itself. The leaf's job is to price what the search cannot see. Addendum A's regression measured how well the leaf duplicates what the search already sees, and the answer — badly — turns out to be a description of a well-designed leaf, not a defect in one.

This also sharpens the standing recommendation in finding-01 and finding-07: the binding constraint is a slow variable, the accumulation of a flow deficit over eight-plus rise cycles. Any diagnostic label defined over an 8-move window is measuring the fast variable. Two independent attempts in this repository have now been misled by a short-horizon proxy — finding-10's H = 9 scenario suite, which ranked fair policies backwards, and this one.

What this does not say

• It does not say the leaf's weights are optimal. It says the gradient does not point toward the achievable-clear direction. A refit against remaining lifetime — the target finding-01 argues for, dense and untailed — is a different experiment and is untouched by this result.
• It does not refute Addendum A's regression, which is reproduced here exactly (held-out R² 0.7342 reweighted vs 0.3962 frozen, cosine +0.1406). It refutes the inference from that regression to a better search leaf.
• It does not test Addendum A's Tier 3, the genuinely absent cover-cracking geometry (coversInWave, landingOnCover). Those are new features, not reweightings. Given that §3 shows their measured gain is +0.005 R² on fair-play positions against +0.054 on synthetic ones, and given that reweighting the existing basis toward the same target loses, the prior on them should now be lower than A.7 implies.

What to do next, ranked by information gain per unit of compute

1. Refit against remaining lifetime, not achievable clears. Same machinery, same CHECK gate, one different label. This is the experiment finding-01 § "Why this matters" actually argues for, and it is the one that would distinguish "the leaf's weights are right" from "this target is wrong".
2. Sweep solid_exposure alone at depth 4 near the depth-3 optimum (+1,250, not the fitted +4,934). It is the one point in this document with a positive estimate at any depth, it was not run at depth 4 (§8), and it is one constant.
3. Do not spend a week on coversInWave / landingOnCover until 1 or 2 produces a positive.

8. Limitations

• 64 paired games per arm. Score standard deviation is 34–64% of the mean, so a 64-game mean carries roughly ±7,000 to ±32,000 at one standard error. The three significant losses are 2–7× that; the t1-rough-560 tie is not resolved and its 95% interval spans roughly ±74,000.
• t1-numbered and t1-solidexp-1250 were not run. Both were queued at depth 4; both were dropped when the shared machine's load average reached 68 against a back-off threshold of 26 and the remaining arms would have taken a further five hours. This is recorded as a no-run, not as a null result. The numbered_cells sign flip is exercised inside t1-combined and both Tier-2 arms, all of which lost; solid_exposure at +1,250 is genuinely untested at depth 4 and is the recommended next arm.
• The depth-3 sweep was stopped at 8 of 25 arms for the same reason. What survives is a prefix in a fixed execution order, so it is an unbiased subset of the planned arms in the sense that no arm was dropped for its result — but it covers only the roughness family and two solid_exposure points.
• The refit corpus is not the policy's own manifold. The fair-play block is 384 positions harvested from d2s5 play. The policy being modified is d4s7. §3 shows the fitted direction is sensitive to position origin at exactly this scale (cosine +0.310 between the synthetic and fair-play fits), so a refit on d4s7-harvested positions could differ again.
• One interpolation path. t2-fair-a05 is the only interior point tested between the frozen and fitted directions, and the path is a straight line in σ-scaled space. A different parameterisation could in principle contain an interior maximum that this one misses, though the monotonicity across six arms makes that unlikely.
• Scale and level were held fixed by construction (§2). This is what allows the risk-2 question to be answered, but it means the experiment says nothing about whether a rescaled version of the fitted direction — which would also change the search's risk attitude — behaves differently.
• These are this repository's simulator semantics, including the two rise-boundary scoring discrepancies in audit-01.
• The machine was shared throughout with other contributors' sessions at load averages of 22–68 against 16 physical cores. Wall times below are therefore not a clean performance baseline; work per move is the comparable quantity and it is reported per arm.

9. Reproduce

# build (clang++ explicitly; the Makefile's CXX ?= clang++ loses to make's
# builtin CXX=g++, which trips a false -Werror=array-bounds in
# src/core/native/public-behavior.hpp)
./approaches/lifetime-objective/leaf-reweight/build.sh
B=./build/lifetime/leaf-reweight
RID=runs/RUN-A527-LEAFW

# CHECK gates - all five, before any gameplay
$B --leaf-check --depth 3 --chance-samples 7 --max-work 16000000 \
   --seed-start 0xa5278000 --check-games 24 --check-moves 400
$B --leaf-check --depth 4 --chance-samples 7 --max-work 16000000 \
   --seed-start 0xa5278100 --check-games 3 --check-moves 45
$B --reference-parity --seed-start 0xa5278200 --check-games 4 --check-moves 60
$B --self-parity --depth 4 --chance-samples 5 --max-work  3200000 \
   --seed-start 0xa5278400 --check-games 3 --check-moves 60
$B --self-parity --depth 4 --chance-samples 7 --max-work 16000000 \
   --seed-start 0xa5278300 --check-games 3 --check-moves 22

# the refit, and the cross-origin transfer that answers risk 3
python3 approaches/lifetime-objective/leaf-reweight/refit.py \
    runs/RUN-SUITE-9c41ab7e2d10/structure-h8-j4.csv --origin fair \
    --alphas 0.25,0.5,1.0 --out-dir approaches/lifetime-objective/leaf-reweight
python3 approaches/lifetime-objective/leaf-reweight/tier1.py \
    runs/RUN-SUITE-9c41ab7e2d10/structure-h8-j4.csv

# depth-3 response-surface map on the tuning lease (NOT a selection gate)
./approaches/lifetime-objective/leaf-reweight/sweep.sh $RID/sweep-d3s7 3 64 8 0xa5279000
python3 approaches/lifetime-objective/leaf-reweight/sweeptable.py $RID/sweep-d3s7 --sort

# depth-4 evaluation on the shared cohort
./approaches/lifetime-objective/leaf-reweight/evaluate.sh $RID/eval-d4s7 12 0xa51d1000
python3 approaches/lifetime-objective/leaf-reweight/compare.py \
    runs/RUN-A51D-s7confirm/fresh-s7.json \
    $RID/eval-d4s7/t1-rough-560.json $RID/eval-d4s7/t1-exposures.json \
    $RID/eval-d4s7/t1-combined.json $RID/eval-d4s7/t2-fair-a05.json \
    $RID/eval-d4s7/t2-fair-a1.json

Artifacts

Under runs/RUN-A527-LEAFW/:

Each cohort artifact is drop7-lifetime-cohort-v1: configuration including the full nineteen-weight vector and the bias, cohort summary, score decomposition, flow rates, and one row per game with seed, score, moves, censor flag, rises, level/clear/chain points, numbered clears, covered reveals, maximum chain depth, mean occupied cells, wall time and logical work.

Frozen source hashes are recorded at build time in build/lifetime/reference-sources.sha256. fair-only-depth4.cpp ends in a real int main and is compiled from a build-tree copy whose only changed line is the entry point; the build refuses to proceed if more than that one line differs.

Seed lease

SEEDLEASE-A52-LEAFW = 0xa5278000–0xa527ffff, a sub-range of the SEEDLEASE-A52 reserve in lease-map.md. Role: exploratory development; once read these seeds are development data permanently. lease-map.md was not edited by this work — it is a shared file and other sessions were writing to it concurrently. The coordinator should add these rows when merging.

The shared evaluation cohort 0xa51d1000–0xa51d103f was read for the reported arms only. No weight value, magnitude, arm or interpolation coefficient in §5 was chosen by looking at evaluation-cohort output; the arm list and the weight values were fixed from the tuning lease and the regression before the first evaluation game was played, and every arm that ran is reported.

Environment

AMD Ryzen AI MAX+ 395, 16 physical / 32 logical cores, single NUMA node, 125 GB RAM, Linux 6.18.35. AMD clang 23.0.0git, -O3 -std=c++20 -pthread -Wall -Wextra. At most 12 threads were used by this work, on a machine shared with other contributors' sessions at load averages of 22–68 throughout; the depth-3 sweep and two depth-4 arms were cancelled when load exceeded the back-off threshold.

Wall time is not comparable across arms as a performance measurement: the arms ran at different times against different background load, and a stronger arm plays longer games. Work per move (§5) is the comparable quantity.