Experiment log
Follow each investigation from its approach to its reports and dated notes. Proposed work is not evidence; completed work can include a negative result. Entries are ordered by when the investigation began.
| Updated | Investigation | Status | Finding and next step |
|---|---|---|---|
| 2026-09-08 | Protocol and replay validation | Active | The harness completes normal games with Rust and Python policies. Smoke results establish plumbing, not policy strength. Next: Repeat a bounded smoke after protocol changes and retain every failed or interrupted attempt. Log 2026-09-08 |
| 2026-09-09 | Engine arena: paired-seed offline tournaments | Active | The engine arena reproduces the protocol ordering (ETA beats fast on 128 paired boards; the frozen reference loses to both). Deterministic mode now gives both tiers seed-indexed dice, theft, and deck tapes; a protocol smoke confirmed identical rolls across rotations. Next: Use the arena for development sweeps and structure measurements; confirm any frozen candidate with a fresh protocol cohort. Log 2026-09-08 · Log 2026-09-09 |
| 2026-09-09 | Game tree structure | Active | Median legal actions: 8 in the main phase, 22 for the robber, 1 for rolls and discards; the main phase widens from 2 to 9 through a game. Each planned turn multiplies expectimax-v2 work by four to six; depth 3 costs about 0.5 s natively and WASM is 1.3 times slower. Next: Measure how much of the raw legal list the candidate generator prunes, and the cost of the confirmed depth-3 candidate under contention. Log 2026-09-08 · Log 2026-09-09 |
| 2026-09-09 | Exchange-aware ETA | Active | The frozen 80-game comparison did not meet its improvement rule. Retain ETA as the baseline. Next: Record decision traces before proposing a new estimator; port acquisition remains an open question. Liquidity statistics · Liquidity audit · Log 2026-09-09 |
| 2026-09-09 | A tunable turn-level expectimax | Active | The corrected depth-3 candidate (expectimax-v2-plan) won 47 of 80 authoritative games against 11 for the compared ETA slot (p = 1.0e-6), meeting its rule; engine arena: depth 2 wins 50.4% and depth 3 57.4% of paired slot-games against ETA, and 46.9% against the frozen reference. Next: Tune the leaf weights and inclinations on unregistered sweeps against the new baseline, measure WASM cost per depth in the browser, and test a response-aware opponent model. Expectimax v2 protocol · Expectimax v2 corrected · Expectimax v2 plan protocol · Log 2026-09-09 |
| 2026-09-09 | Leaf-guided opponents versus the builder model | Active | Round one, 256 deterministic games: contrast -0.008 wins per game (95% interval -0.111 to +0.096). See the ablation report. Next: Read the confirmation cohort on fresh seeds where registered; otherwise retest under the confirmed depth-3 candidate before changing defaults. Log 2026-09-09 |
| 2026-09-09 | Proposing trades versus never proposing | Active | Round one, 256 deterministic games: contrast -0.027 wins per game (95% interval -0.132 to +0.077). See the ablation report. Next: Read the confirmation cohort on fresh seeds where registered; otherwise retest under the confirmed depth-3 candidate before changing defaults. Log 2026-09-09 |
| 2026-09-09 | Maximal trade openness | Complete | Screen -0.094 then confirmation -0.016 (95% interval -0.075 to +0.043): removing the trade margin makes no measurable difference against builders. Next: Test the margin against another search seat that prices trades. Log 2026-09-09 |
| 2026-09-09 | Hidden hands sampled from event history versus public counts only | Complete | Screen -0.090 then confirmation -0.066 (95% interval -0.125 to -0.008) on 192 fresh seeds: sampling hands from event history instead of counts alone is worth about 0.07 wins per game. Supported. Next: Measure calibration of the sampled hands against sealed true hands offline, then test richer inference (refusals, discards). Log 2026-09-09 |
| 2026-09-09 | One sampled world versus four | Complete | Screen -0.051 then confirmation -0.012 (95% interval -0.072 to +0.048): one sampled world instead of four makes no measurable difference at depth 2. Next: Retest at depth 3, where worlds also shape the opponents' rounds. Log 2026-09-09 |
| 2026-09-09 | Maximal aggression | Active | Round one, 256 deterministic games: contrast +0.020 wins per game (95% interval -0.084 to +0.123). See the ablation report. Next: Read the confirmation cohort on fresh seeds where registered; otherwise retest under the confirmed depth-3 candidate before changing defaults. Log 2026-09-09 |
| 2026-09-09 | Opponent progress and production terms | Active | Round one, 256 deterministic games: contrast +0.000 wins per game (95% interval -0.105 to +0.105). See the ablation report. Next: Read the confirmation cohort on fresh seeds where registered; otherwise retest under the confirmed depth-3 candidate before changing defaults. Log 2026-09-09 |
| 2026-09-09 | A maximal development-card appetite | Active | Round one, 256 deterministic games: contrast +0.031 wins per game (95% interval -0.070 to +0.132). See the ablation report. Next: Read the confirmation cohort on fresh seeds where registered; otherwise retest under the confirmed depth-3 candidate before changing defaults. Log 2026-09-09 |
| 2026-09-09 | Development card and knight terms | Active | Round one, 256 deterministic games: contrast -0.035 wins per game (95% interval -0.144 to +0.074). See the ablation report. Next: Read the confirmation cohort on fresh seeds where registered; otherwise retest under the confirmed depth-3 candidate before changing defaults. Log 2026-09-09 |
| 2026-09-09 | Reachable-site value and road contention terms | Complete | Screen -0.074 then confirmation -0.016 (95% interval -0.078 to +0.047): no measurable effect at depth 2. Next: Test the terms where they should matter: a lineup with a contested site race, for example against another search seat. Log 2026-09-09 |
| 2026-09-09 | Ablations: what each mechanism is worth | Active | Eighteen one-change arms plus six fresh-seed confirmations: lookahead (-0.324) and scenarios (-0.410) are decisive; event-based hand knowledge is worth about 0.07 wins per game (-0.066, interval excludes zero); every other term and slider is within 0.1 with an interval crossing zero. Next: Retest the surviving mechanisms under the confirmed depth-3 candidate and against a second search seat, where trading and containment terms could matter. Log 2026-09-09 |
| 2026-09-09 | Placing the opening with the search instead of the builder heuristic | Active | Round one, 256 deterministic games: contrast -0.035 wins per game (95% interval -0.133 to +0.062). See the ablation report. Next: Read the confirmation cohort on fresh seeds where registered; otherwise retest under the confirmed depth-3 candidate before changing defaults. Log 2026-09-09 |
| 2026-09-09 | Follow-ups to the learned leaf | Active | The tables hold against more searches (+0.172; symmetric field +0.094). No learning setting or the corner window separates from a 0.2 noise floor; two four-hour runs do not beat the 96,000-game tables on fresh seeds (+0.031, +0.055). Depth 4 under a 4 s budget with a gap rule is inconclusive in the engine for both leaves (+0.078, -0.051) and weaker through the protocol (-0.200, interval below zero), so the baseline stays ntuple-leaf. Next: Read the equal-time depth-4 screen and the gap-rule-only cohort; then change the representation (more window families under a long budget) or the search around the tables rather than train longer. Log 2026-09-09 |
| 2026-09-09 | N-tuple tables as the search leaf | Complete | The frozen hex plus portfolio tables beat the hand-written leaf on fresh seeds at every depth tested (equal depth 2: +0.219, 95% interval +0.124 to +0.314; one ply against depth 2: +0.164; depth 2 against depth 3: +0.156; depth 3 against depth 3: +0.242) and won 46 of 80 authoritative games against 25 for expectimax-v2-plan (paired contrast +0.263, +0.094 to +0.431); the protocol baseline is now ntuple-leaf. Next: Vary the learning settings and the window families under a longer budget, read the portfolio tables, speed up feature extraction, and test the tables against a second search seat and negotiating opponents. Ntuple leaf protocol · Log 2026-09-09 |
| 2026-09-09 | Cross-play of four search styles | Active | Four inclination presets split 256 deterministic games almost evenly (66, 59, 69, 62 wins); no preset dominates. Next: Cross-play at depth 3 and against the frozen reference, then a heatmap of pairwise contrasts. Log 2026-09-09 |
| 2026-09-09 | Port value terms | Active | Round one, 256 deterministic games: contrast +0.020 wins per game (95% interval -0.086 to +0.125). See the ablation report. Next: Read the confirmation cohort on fresh seeds where registered; otherwise retest under the confirmed depth-3 candidate before changing defaults. Log 2026-09-09 |
| 2026-09-09 | Learning who accepts | Active | Round one, 256 deterministic games: contrast -0.031 (interval -0.125 to +0.063); no effect against fixed builders. Next: Test reciprocity against an opponent whose acceptance depends on history, for example a second search seat with the same model. Log 2026-09-09 |
| 2026-09-09 | The discard-exposure term | Active | Round one, 256 deterministic games: contrast -0.023 wins per game (95% interval -0.134 to +0.087). See the ablation report. Next: Read the confirmation cohort on fresh seeds where registered; otherwise retest under the confirmed depth-3 candidate before changing defaults. Log 2026-09-09 |
| 2026-09-09 | A maximally risk-seeking style | Complete | Screen -0.062 then confirmation -0.055 (95% interval -0.118 to +0.008): leaning harmful, not resolved by 768 games. Next: A larger cohort or a position-dependent test (trailing versus leading) before changing the default. Log 2026-09-09 |
| 2026-09-09 | Static turn planning versus one round of lookahead | Active | Round one, 256 deterministic games: contrast -0.324 wins per game (95% interval -0.413 to -0.235). See the ablation report. Next: Read the confirmation cohort on fresh seeds where registered; otherwise retest under the confirmed depth-3 candidate before changing defaults. Log 2026-09-09 |
| 2026-09-09 | Two common random scenarios versus eight | Active | Round one, 256 deterministic games: contrast -0.410 wins per game (95% interval -0.489 to -0.331). See the ablation report. Next: Read the confirmation cohort on fresh seeds where registered; otherwise retest under the confirmed depth-3 candidate before changing defaults. Log 2026-09-09 |
| 2026-09-09 | Transposition tables in the turn-level search | Complete | Verified Zobrist-style key and tables shared across depth passes or a turn: 128 registered games identical move for move with 11.7% less decision time at equal nodes; at equal time 15% faster at the same completed depth with no strength change (-0.031, interval crosses zero); protocol cohort inconclusive (27 to 36, -0.113, -0.265 to +0.040), baseline unchanged. Next: Share opponents-round arrivals across afterstates that differ only in the player's own hand, and revisit the time rule so saved time can reach a further pass. Log 2026-09-09 |
| 2026-09-09 | The acquisition-time plan term | Complete | Screen +0.062 then confirmation +0.039 (95% interval -0.021 to +0.099) on 192 fresh seeds: removing the plan term does not hurt at depth 2; the term is redundant once turns are planned. Next: Retest under the depth-3 candidate before removing the term from the leaf; keep it as a documented setting. Log 2026-09-09 |
| 2026-09-11 | Predicting opponents next builds | Complete | Not confirmed. A denial leaf term reached +0.061 wins per game on development seeds 0-63 (95% interval +0.009 to +0.112) and fell to -0.025 (-0.071 to +0.020) on fresh seeds 64-127; the predicted opponent model and predicted space threat never separated from zero. The predictor matched 70% of the fast builder's next builds and 53% of ETA's, but only 28% of another search's, and the search is the rival that wins these tables. The denial seat did lose about 0.1 fewer frontier sites per game in every cohort. The same cohorts measured a slot-order asymmetry in the paired ablation lineup: +0.172 wins per game for slot 0 on seeds 0-63 with identical twins. Next: Keep all three switches off. Future paired contrasts should run mirrored slot orders or report the twin-seat baseline; retesting builder-directed terms needs a table where builders are the real rivals, and any retest of the denial term should use the ntuple-leaf baseline and fresh seeds only. Log 2026-09-11 |
| 2026-09-12 | Build prediction under the learned leaf | Complete | Not confirmed. Swapped pairs against the tables baseline put every effect inside its interval: predicted threat +0.047 (-0.004 to +0.098) on the population lineup, +0.014 on seeds 0-63, and -0.006 on seeds 800-863; predicted opponents -0.010, +0.000, and -0.004; denial 0.5 on the quarter blend +0.002, then -0.037 (-0.071 to -0.003) on fresh seeds, refuted a second time, at about double the decision cost. The predictor named a tables search's next build in 20.0 percent of turns (28.5 against the hand-leaf search) and eta's in 59.0. All three switches stay off under both leaves; the baseline and evidence boundary do not change. Next: Keep all three switches off. The predicted space threat is the only arm with a positive-leaning reading anywhere (population lineup, both leaves, both intervals crossing zero); resolving an effect near +0.03 would need several hundred seeds per half on a population of trading opponents, so register that only as a deliberate large cohort. The denial term should not be revisited without a cheaper predictor, since its leaf term doubles decision time. Log 2026-09-11 |
| 2026-09-12 | Bargaining switches under the learned leaf | Complete | Swapped pairs under the learned n-tuple leaf, pooled after the fact over 192 seeds: all bargaining off -0.143 wins per game (95% interval -0.180 to -0.106), counters off -0.102 (-0.136 to -0.069), asks off -0.085 (-0.119 to -0.050), threat pricing off +0.001 (-0.035 to +0.036), and the stubborn seat +0.006 (-0.008 to +0.020) over 320 seeds after its extension to seeds 864-927. Counters and all off lose at every stage and asks at both builder-table stages (inconclusive at the population table). Bargaining costs +59% decision time under both leaves, counters +41%, asks +7%, threat pricing +4%. Development tier, depth 2; every switch stays on and no default changes. Next: Measure counters at equal time rather than equal nodes: a swapped pair of the browser configuration (hand-written leaf, depth 3, one-second budget) with counters on and off, to see whether the 40 percent extra time per decision costs depth where the browser spends it. Log 2026-09-11 |
| 2026-09-11 | Threat-aware pricing, counters, and reasoned declines | Complete | Two identical defaults in the arm design's slots 0 and 1 differ by +0.176 wins per game (95% interval +0.083 to +0.268, run 91ddb42d), so the +0.219 of run 4228f4f7 is mostly seating: +0.043 (-0.082 to +0.168) against that null. Read the same way, counters off costs -0.199 (-0.309 to -0.089), asks off -0.156 (-0.259 to -0.054), threat pricing off -0.008 (no effect), and a partner that refuses every ask loses nothing by refusing. A swapped pair on fresh boards 64-127 confirms counters at -0.096 (-0.156 to -0.035) for the seat without them, with a seating term of -0.010 there; a second null on boards 64-127 put the seating term at +0.051 (-0.046 to +0.148), so its size varies by board set and every two-search contrast needs both seatings. Development tier at a table of two searches and two builders. Next: A follow-up study: rotate the recipient order of multi-seat offers in the arena, re-measure the null in both seatings, rerun the two-search arms of the ablation program as swapped pairs, and test the bargaining against a partner that counters back. Bargaining arms · Log 2026-09-11 |
| 2026-09-11 | Endgame racing and stopping the leader | Complete | Withdrawn as a strength claim. The single-seating cohorts (candidate always in slot 0) measured the arena seating term: +0.129, +0.172, and +0.168 wins per game on seeds 0 to 63 and +0.078, +0.039, and +0.051 on seeds 64 to 127, the size of the measured bias. Swapped pairs under containment/endgame-seating put the seating-corrected effects at +0.018 (95% interval -0.018 to +0.053), +0.006 (-0.030 to +0.041), and +0.012 (-0.018 to +0.041), each crossing zero. No switch is confirmed and the defaults stay unchanged. Next: None for the switches as they stand; any future endgame cohort is registered as a swapped pair from the start. Log 2026-09-11 |
| 2026-09-11 | Endgame switches in both seatings | Complete | Swapped pairs on seeds 64 to 127, both halves on the current engine (it no longer reproduces the original confirmations): race leaf +0.018 wins per game (95% interval -0.018 to +0.053), hidden points +0.006 (-0.030 to +0.041), leader bias +0.012 (-0.018 to +0.041), each with an interval crossing zero. The seating terms were +0.092 (+0.014 to +0.169), +0.041 (-0.047 to +0.129), and +0.059 (-0.029 to +0.146). The single-seating contrasts of containment/endgame-race were seating; the corrected effects are consistent with none and exclude anything above about +0.05. Next: No further cohorts for the switches at depth 2 against this lineup; re-open only with a swapped-pair registration, a different lineup, or the deeper confirmed configurations. Log 2026-09-11 |
| 2026-09-12 | Robber placement under the learned leaf | Complete | All four switches are inconclusive under the learned leaf. Tested as swapped pairs through three stages (development seeds 0-63, confirmation and population on seeds 800-863) with a preregistered extension to seeds 864-927 for need and block_leader, 36 cohorts of 256 games all complete, the seating-corrected effects pooled over stages are leader +0.007 (95% interval -0.035 to +0.049) and +0.031 (-0.030 to +0.092), need +0.016 (-0.008 to +0.041) and +0.009 (-0.018 to +0.036), threat -0.005 (-0.042 to +0.033) and +0.043 (-0.005 to +0.091), block_leader +0.011 (-0.005 to +0.027) and +0.015 (-0.007 to +0.036) wins per game, builders and population lineups. The block_leader population extension is the only cohort whose interval excluded zero (+0.033, +0.007 to +0.060) and it does not survive pooling. Combining the two positive arms read +0.035 (-0.015 to +0.085) in the builders lineup and -0.039 (-0.090 to +0.012) in the population lineup. No switch is worth the browser; all stay off. Next: None for these switches in either leaf. The hand-leaf refutation of the need rule does not reproduce under the tables, so any future robber work should name which leaf it prices; block_leader is the only arm to revisit, and only with a different mechanism or a lineup with distinct retaliators. Log 2026-09-11 |
| 2026-09-11 | Robber placement and victim choice | Complete | All three arms beat the default shortlist with the arm in slot 0 on seeds 0-63 (+0.125, +0.160, +0.117) and read near +0.035 on fresh seeds, but swapped pairs under containment/robber-targeting-bias show those gains were the arena seating term: bias-free the effects are leader -0.023, need -0.045 (the only interval excluding zero, on the wrong side), and threat -0.033 wins per game on seeds 64-127. No switch helps, the need rule is refuted at this table, and all three stay off by default. Next: None for these arms. Any future two-search cohort in this arena needs a swapped pair or a null arm; the arms could be revisited against opponents that retaliate, where the leader and threat orderings might matter more than against fixed builders. Log 2026-09-11 |
| 2026-09-11 | Robber targeting against the seating bias | Complete | Three swapped pairs on seeds 64-127 cancelled the arena seating term for the robber-targeting arms: bias-free effects leader -0.023 (95% interval -0.073 to +0.026), need -0.045 (-0.086 to -0.004, the only interval excluding zero), threat -0.033 (-0.084 to +0.018) wins per game, with a seating term of +0.059 to +0.084 for identical searches in adjacent slots in this binary. The slot-0 contrasts of the development cohorts were mostly seating term plus seed selection. Next: Every future two-search cohort in this arena should carry a swapped pair or a null arm; the seating term itself is owned by reports/arena-seating.mdx. Log 2026-09-11 |
| 2026-09-11 | When to buy and play development cards | Complete | No. Three one-switch arms against the unchanged depth-2 control, 256 deterministic games each: single-seating contrasts of +0.145, +0.043, and +0.098 on boards 0-63 and +0.074 and +0.066 on boards 64-127, all the size of the turn-order seating term between adjacent search slots. The knight_play_value arm re-measured as a swapped pair on boards 64-127 corrects to +0.004 (95% interval -0.061 to +0.069); the pair's seating term is +0.070 (+0.005 to +0.136). The diagnostics show the mechanisms work (the tempo rule halves purchases, holding knights forgoes army awards) without winning. All switches stay off by default. Next: Retest contextual card valuation under the ntuple-leaf baseline if a card question opens there; any future two-search contrast in this arena is registered as a swapped pair. Log 2026-09-11 |
| 2026-09-11 | Development-card timing under the learned leaf | Complete | The two candidate rules lose under the tables and the leaf term does nothing on a usable blend. Swapped pairs, 512 games each, seating-corrected effects: buy_tempo -0.061 (development, refuted), -0.006 (confirmation), -0.053 (population); hold_knight -0.029, -0.025, and -0.051 (population, refuted); knight_play_value on the quarter blend -0.002, +0.000, +0.004 (all inconclusive), and +0.109 (+0.055 to +0.164) on the full blend's development pair, a configuration itself 0.2 wins per game below the tables alone. After-the-fact pooling over the 192 stage seeds: buy_tempo -0.043 (-0.076 to -0.009), hold_knight -0.035 (-0.062 to -0.009), knight_play_value (quarter blend) +0.001 (-0.016 to +0.017). All twenty runs completed 256 of 256 games. The blended leaf was fixed to honor knight_play_value, with a test. No default changes; no switch is worth the browser. Next: Nothing for these switches. If a blend becomes the default leaf, re-read knight_play_value at that blend weight; the effect scales with the hand-written share (zero at 0.25, +0.109 at 1.0 on one stage). Log 2026-09-11 |
| 2026-09-11 | Road blocking and longest-road commitment | Complete | Measured with the candidate in slot 0, the development cohorts gave +0.168 and +0.113 wins per game and the fresh-board confirmations +0.031 each. Those numbers are the size of the turn-order seating term between adjacent searches, and the swapped halves under expansion/road-blocking-swapped show they were that term: seating-corrected, the blocking term gives +0.008 (95% interval −0.053 to +0.069) and the production gate −0.004 (−0.061 to +0.053) on boards 64-127. Both consistent with no effect; the defaults stay off. Trail-breaking settlements are already enumerated and credited; a probe test pins this. Next: No further work on these two switches at depth 2. The corrected estimates and intervals live in the follow-up study expansion/road-blocking-swapped. Log 2026-09-11 |
| 2026-09-11 | Road-blocking switches as swapped pairs | Complete | On boards 64-127 with the control in slot 0, the swapped halves gave +0.016 (−0.080 to +0.111) for opponent_expansion 0.5 (run c6201b2c) and +0.039 (−0.057 to +0.135) for road_contention_gate 1.0 (run 86fa73ec). Halving the contrast differences against the candidate-in-slot-0 runs 2fd1cb5e and ee7d78c7: the blocking term's corrected effect is +0.008 wins per game (95% interval −0.053 to +0.069), the gate's is −0.004 (−0.061 to +0.053). Both are consistent with no effect; the single-seating numbers were the seating term. Next: No further work on these two switches at depth 2; the intervals rule out effects beyond about 0.07 wins per game. Trail breaks are already enumerated and credited. If blocking is retested, use a lineup with a contested site race and a registered swapped pair from the start. Log 2026-09-11 |
| 2026-09-12 | Road-blocking switches under the learned leaf | Complete | No. Twenty-eight swapped-pair cohorts (7,424 games, all complete) across three stages, two lineups, and two leaf contexts: opponent_expansion 0.5 under the tables alone is -0.003 wins per game (95% interval -0.011 to +0.005) over 128 builders seeds and +0.002 (-0.007 to +0.011) in the population lineup; on the 0.25 blend it is +0.010 (-0.017 to +0.037) over 192 builders seeds after the gated extension and -0.033 (-0.060 to -0.006) over 128 population seeds. road_contention_gate 1.0 is -0.020 (-0.045 to +0.004) pooled over 192 tables-leaf boards and is refuted on the blend at -0.060 (-0.095 to -0.024) over the builders seeds. The full-blend registrations were withdrawn before running when a screen showed hand 1.0 far weaker than the tables alone; the blend arms ran at hand 0.25. Both switches stay off, the browser included: under the hand-written leaf they already measured +0.008 and -0.004, and the blocking weight costs about a third more decision time. Next: No further work on these two switches in any leaf context; the intervals rule out useful effects and the gate is harmful under the blend. Blocking ideas need a lineup with a contested site race and a mechanism that actually finds blocks, which neither weight did (under 0.1 blocking builds per game). Log 2026-09-11 |
| 2026-09-11 | Why two identical searches in adjacent slots differ | Complete | They still differ, +0.180 (+0.075 to +0.285, run 689a1947), so the recipient order was not the cause. The term is adjacency in turn order, a property of the game: two identical seats with bargain off differ by +0.102 (+0.020 to +0.183), with two identical ETAs beside them by +0.141 (+0.051 to +0.230), with the builders swapped by +0.133 (+0.040 to +0.226), and with an ETA between them (slots 0 and 2) by +0.039 (-0.048 to +0.126); one search with three ETAs leaves the builders within 0.03 of each other. Proposals off (+0.074), counters off (+0.090), and asks off (+0.008) are inconclusive or contradictory one by one; pooled they sit at +0.068 (+0.015 to +0.122), about 0.11 below the default null, so full bargaining doubles a term it does not create. The seat directly before another acts first after three of the four rolls in a round. No arena code changed; docs corrected. Development tier, boards 0 to 63. Next: Keep measuring two-search contrasts as swapped pairs. If the bargaining's share of the term matters for a design, trace which mid-game trades or races the earlier seat wins, on more than 16 boards. Log 2026-09-11 |
| 2026-09-11 | Placing the opening with a draft-aware portfolio planner | Complete | Swapped-seating pairs on 128 boards (runs b698064e, da0b1591, 5e3c4985, d89ea4de): planner effect -0.017 wins per game (95% interval -0.075 to +0.042) after removing a seating term of +0.063; the earlier single-seating gains of +0.06 to +0.12 were seating. The calibrated planner matches the previous placement in strength, and the pip-only first draft lost by 0.30. Discrete strategies hurt at every strength and stay off. With the planner and bargaining together, depth 2 wins 66.4% of slot-games against ETA, fast, and ETA (run 42f60c41). Next: Test the coverage, balance, and expansion weights against a second search seat that blocks, in swapped pairs, where balanced pairs may matter more than against builders; then the same at depth 3 under the browser budget. Log 2026-09-11 |
| 2026-09-12 | The opening planner under the learned leaf | Complete | Under the learned-tables leaf, the search-ranked opening ("opening":"search") beats the default planner against builders: pooled L1 effect +0.061 wins per game (95% interval +0.009 to +0.113) over 192 seeds across three swapped pairs, but the only cohort excluding zero is the development boards 0-63 (+0.135, +0.045 to +0.225), and on the 128 fresh L1 seeds the effect is +0.024 and crosses zero. Against the population lineup the pooled effect is +0.001 (−0.059 to +0.061). The coverage 0.4 / balance 0.5 arm is a null (pooled L1 +0.020, −0.014 to +0.053; pooled L2 +0.004), and removing the scarcity markup is a null (pooled L1 −0.005; L2 −0.012), so both weights stay at their defaults. The search arm costs 25 to 33 percent more mean decision time, concentrated in three or four setup decisions of about 2 s each; it fails the browser rule and the browser keeps the planner, whose hand-leaf comparison was a match anyway. Next: If the server default is to change, register a protocol cohort of the search-ranked opening against the ntuple-leaf baseline, since the engine gain rests on the reused development boards and is zero against search tables; the alternative mechanism is recalibrating the planner's weights to the tables' site values rather than routing placement through the search. Log 2026-09-11 |
| 2026-09-12 | Blending the learned tables with the hand-written leaf, and the endgame switches under the blend | Complete | The hand-written terms at quarter strength improve the learned tables: +0.074 wins per game on seeds 0-63 (+0.011 to +0.138), +0.088 on the extension seeds 864-927, +0.080 there at the population lineup, pooled +0.066 (+0.024 to +0.107) over 192 L1 seeds and +0.064 (+0.021 to +0.108) over 128 L2 seeds, at about 40 percent more decision time. Half strength is noise (+0.016 pooled over 192 L1 seeds, the population lineups negative), and full strength loses to the tables alone (the withdrawn half lost from the favorable seat, an unregistered screen read -0.207). Scale 500 is refuted (-0.097 pooled, it stops the seat trading), scale 2000 is a cost-free nothing (+0.029 pooled over 192 L1 seeds). The endgame race leaf is -0.037 pooled over its L1 stages under the blend (refutation) and hidden points is +0.003 pooled over 192 L1 seeds, both after reading +0.018 and +0.006 under the hand-written leaf, so both stay off under either leaf. Next: Register a protocol-arena cohort seating the quarter-strength blend inside ntuple-leaf's exact search against ntuple-leaf before changing any server-side default; the browser build cannot load the tables and keeps the hand-written leaf with both endgame switches off. Log 2026-09-11 |
| 2026-09-11 | Hand-size risk: spend-down, projected discards, and the hand_risk weight | Complete | Single-seating contrasts on boards 0-63 were +0.066 to +0.129 wins per game, the size of the turn-order seating term; the best arm (discard_projection) re-measured as a swapped pair corrects to -0.029 (95% interval -0.073 to +0.015). No effect anywhere; defaults unchanged. The diagnostics move as the mechanisms predict. Next: No default change. Any retest belongs at depth 3 or under the learned-tables leaf, registered as swapped pairs from the start. Log 2026-09-11 |
| 2026-09-12 | Hand management under the learned-tables leaf | Complete | All three arms correct to zero in swapped pairs. spend_down: -0.006 (95% interval -0.042 to +0.030) pooled over the 128 builders seeds and -0.047 (-0.098 to +0.004) in the population lineup, at 1.19 times the decision time. discard_projection: +0.012 (-0.010 to +0.034) over 192 builders seeds after the triggered extension and -0.002 (-0.023 to +0.019) over 128 population seeds, at no cost. hand_risk 0 on the quarter blend: -0.001 (-0.029 to +0.028) over 192 seeds. Every per-stage rule lands inconclusive; no arm meets the +0.03 browser threshold. Under the tables the search holds 4.8 cards at the roll and discards half as often as under the hand-written leaf, so the situation these mechanisms price comes up half as often. Next: No default change: both switches stay off and hand_risk stays at 1. Nothing here transfers to the browser. If the leaf or the discard rule ever changes, the question deserves one retest on the new configuration. Log 2026-09-11 |
| 2026-09-12 | Search settings under the learned tables leaf | Complete | Nine switches, 54 swapped-pair cohorts, every game complete. Four confirmed harms: scenarios 4 at −0.392 wins per game, samples 1 at −0.109, counts-only beliefs at −0.143, no proposals at −0.172 (pooled L1, 128 seeds, intervals excluding zero); reciprocity is refuted at −0.073 (−0.115 to −0.031) and −0.082 at the population table. Eight sampled worlds gain a confirmed +0.045 against builders (192 seeds) for 1.8 times the decision cost and nothing against a table of searches; leaf-guided opponents read +0.031 and +0.033 at the two lineups with no interval excluding zero at 1.8 times the cost; scenarios 16 buys nothing for 2.7 times the cost. inner_scenarios cannot bind at depth 2: its halves replayed identical games. No switch is worth the browser by the registered rule. Next: Measure inner_scenarios and the cost-saving settings (samples 1, scenarios 4) at depth 3 in the browser budget where they bind, and revisit eight sampled worlds if the tables leaf ever runs where decision time is not scarce. Log 2026-09-11 |
| 2026-09-12 | Every search switch under the learned leaf | Complete | Nine studies, 145 seating-corrected swapped pairs, 74,240 games, every reported outcome matched by recomputation from the archived runs. Bargaining holds under the learned leaf (counters off −0.094 and −0.119, all off −0.165 and −0.098 at the builder table and the table of searches) and the search's information inputs matter more than under the hand-written leaf (four scenarios −0.392, one world −0.109, public-count beliefs −0.143, no proposals −0.172). The quarter-strength hand-written blend improves the tables at both tables (+0.066, +0.064) at 1.4 times the decision time. Eight sampled worlds and the search-ranked opening read positive at the builder table only; every robber, timing, hand, road, prediction, endgame, and planner-weight switch shows no effect that survives both tables. Next: Register a protocol-arena cohort of the quarter blend inside ntuple-leaf's search against ntuple-leaf before any server default changes; if the tables reach the browser, measure the blend's depth reached under the one-second WASM budget. Keep the browser's hand-written configuration unchanged. Log 2026-09-12 |