Compact evidence: results/kaggle_v0_2_1_match_history.json. Raw replays remain external because each source file is roughly 14 MB.

Regression fixture: results/kaggle_v0_6_1_pq_marz_regression.json. The full 20 MB replay remains outside the repository.

Read: the three-path mixture converts every Alpha4 tie against the submitted agent and raises day-15 eight-livestock coverage from 12.5% to 62.5%. Alpha5 qualifies against Alpha9, but its 221-coin seed-3304 loss means it is not a universal replacement for Alpha4 yet.

1. LEAN
   Preserve cash and the mixed core

   Liquidity remains the baseline when the alternative paths cannot beat its downside estimate.

   SAFE BLUEPRINT
2. DENSE
   Convert one plot into animal capacity

   Pasture geometry, feed crops, labor, and species mix move together toward an eight-to-sixteen-animal engine.

   NEW PATH
3. LAND
   Compound a funded production bundle

   Alpha4's staged expansion remains available when density, service, lower tail, and terminal runway prove the next plot.

   OPTION VALUE
4. SHIELD
   Require probability and lower-tail separation

   A non-lean expert acts only when its risk-adjusted edge and public farm constraints agree.

   BASELINE FALLBACK
5. RESULT
   Break every submitted-agent tie

   The exact twenty-game Alpha9 panel finishes 20–0 at +5,142 average; the arena remains 8–0.

   GATE PASSED
6. LIMIT
   Deployment timing is still imperfect

   On seed 3304, animals sit in the shed longer than their productive slots justify; Alpha4 wins by 221.

   NEXT LEVER

Evidence: results/v1_1_0_alpha5_checkpoint.md. Exact artifact: kaggriculture-v1.1.0-alpha5.tar.gz · SHA-256 54d1…9d0.

1. STAGE
   Reserve cash for one observation

   An equal-land expansion no longer buys immediately. The compact farm keeps operating while the priced bundle remains protected.

   OPTION VALUE
2. TAIL
   Compare the weaker futures

   Expected edge is joined by lower-half value, adverse edge, and forecast spread before land can be confirmed.

   DISTRIBUTIONAL
3. RELEASE
   Fall back when the bundle becomes fragile

   Funding, service, feasibility, weeds, or a weak lower tail release the reserve to compact play.

   BASELINE SHIELD
4. DIRECT
   Remove every Alpha3 loss

   The same twenty Alpha9 games finish 12–0–8 at +3,482 average and zero minimum margin.

   SAFER FLOOR
5. ARENA
   Win the gym, fail the milestones

   Eight personas lose, but Alpha4 reaches neither early expansion nor 60k by day 24. The gym is not yet demanding enough.

   COVERAGE WARNING
6. NEXT
   Compare explicit capacity paths

   Dense livestock, land compound, and lean liquidity need separate proposals and score distributions.

   MIXED MODEL

Evidence: results/v1_1_0_alpha4_checkpoint.md. Exact artifact: kaggriculture-v1.1.0-alpha4.tar.gz · SHA-256 93b1…9bfa.

1. TRACE
   Find the suppressed day-7 choice

   On tie seed 3300, expansion had 98.5% policy probability and MPC support, but a stale compact commitment kept the action from the executor.

   EARLIEST GAP
2. PERMIT
   Require six environment checks

   Density, full-bundle funding, service, runway, auxiliary value, and scenario MPC must agree before proactive equal-land expansion.

   MIXED ATTENTION
3. PARITY
   Risk land only to break an economic tie

   The first draft lost frozen seed 2700. i2 disables proactive land when the compact engine already has a material bank lead.

   REGRESSION FIX
4. DIRECT
   Convert ties—but introduce four losses

   The fixed twenty-game Alpha9 panel finishes 16–4 at +5,635 average and −10,689 worst.

   FAIL 20–0
5. ARENA
   Ten persona wins are not representative

   Five assumed top styles lose both seats by at least 13,603, yet none reproduces the Alpha9 variance failure.

   COVERAGE GAP
6. NEXT
   Stage before committing

   Reserve expansion cash for one observation, then combine a downside-percentile head with the new evidence before buying land.

   OPTION VALUE

Evidence: results/v1_1_0_alpha3_i2_checkpoint.md. Exact artifact: kaggriculture-v1.1.0-alpha3-i2.tar.gz · SHA-256 9810…6270.

1. HEADS
   Score five reasons a farm can be strong

   Capacity, service, capital, runway, and trajectory are evaluated separately before the daily branch is selected.

   CONSTANT TIME
2. FOCUS
   Reweight the three existing spotlights

   Operations, opponent, and horizon attention still organize the evidence; the new head improves their capital judgment.

   NO FOURTH BOT
3. GUARD
   Do not buy dependencies for rejected land

   The first draft preferred expansion while the executor correctly blocked equal-land growth, leaving cows and sheep idle. A land-not-authorized fallback now returns the entire bundle to compact play.

   REGRESSION FIXED
4. REAL
   Replay the two observed failure seeds

   The aggressive-land proxy loses both seats on seeds 845516275 and 1982640029; Alpha2 wins all four by at least 25,515.

   4–0
5. DIRECT
   Separate from the submitted policy

   Alpha2 beats exact Alpha9 8–0 at +6,830 average and beats v0.9.2 8–0 at +10,932 average, with zero failures or fallbacks.

   PROMISING
6. LIMIT
   The strict release gate is not cleared

   A fresh 20-episode Alpha9 panel stopped at its cap after 17: eleven wins, six ties, zero losses. The ties already prevent a 20–0 promotion.

   HOLD
7. NEXT
   Break equal-land ties without restoring blind land rushes

   On seed 3300 both agents were identical all match: one plot, 19 crops, six animals, and 23,807 coins on day 15. The next permit should use saturation + funding + service + runway, not only an opponent land lead.

   ALPHA3

Exact candidate: artifacts/kaggriculture-v1.1.0-alpha2.tar.gz · 39,100 bytes · SHA-256 d70c…9182 · 53 policy tests.

Shared diagnosis: Alpha9 stayed on one quadrant with six animals in all four games. The next mixed selector must compare dense livestock, land compound, and lean liquidity as distinct capacity paths. Evidence: results/kaggle_alpha9_recent_four_review.json.

Trace boundary: these are visible states, chosen branches, and reproducible counterfactuals. They do not claim access to either bot's private reasoning.

1. GAP
   The planner could like expansion without buying land

   A trace against the aggressive opener exposed a disconnect between the daily capital plan and the hourly executor.

   ROOT CAUSE
2. OPTION
   Name the multi-day intent

   Launch, densify, expand, repair, and close each have a start condition, a short commitment window, and a break condition.

   LESS THRASH
3. GUARD
   Feasible does not mean useful

   One land purchase is allowed only when MPC still prefers it, the full bundle is affordable, and the opponent visibly owns more plots.

   ONE PLOT
4. BREAK
   Operations can interrupt strategy

   Service overload, weeds, terminal timing, or an infeasible package immediately release the option.

   SAFETY FIRST
5. RESULT
   Specialist gain, no direct separation

   The six-style arena rises to +29,109 average, but eight direct games with Alpha9 are exact ties.

   NOT READY

Evidence: results/v1_1_0_alpha1_checkpoint.md, 49 policy regressions, exact frozen panels, and a 360-row option-junction dataset.

1. ROOT
   Alpha8 expanded from an underfilled second plot

   Its three losing seeds bought plot three at roughly 48–56% projected density, then absorbed animal, feed, labor, and weed debt.

   6 LOSSES
2. PLAN
   Compare four reachable packages

   Compact, expand, repair, and liquidate are simulated under favorable, base, and adverse demand, service, and weed paths.

   4-DAY HORIZON
3. VETO
   Reject an expansion that cannot operate

   Cash, full-bundle funding, projected density, service capacity, and remaining payback time are hard feasibility checks.

   137 BLOCKED
4. REPLAN
   Execute today; reconsider tomorrow

   The controller chooses one daily macro branch while the proven hourly executor keeps movement and survival deterministic.

   RECEDING HORIZON
5. GATE
   Flip every repeatable submitted-control loss

   Seeds 3203, 3206, and 3208 now win both seats. The exact artifact finishes 20–0 with a +2,032 floor.

   PASS
6. LIMIT
   Reliability improved more than ceiling

   The average margin is lower than Alpha8 on the same panel. v1.1 must add execution density without giving back this safety floor.

   NEXT

Evidence: results/v1_0_0_alpha9_checkpoint.md, exact gate JSON, packaging QA, and the 360-row MPC arena dataset.

1. TRIGGER
   Measure uncertainty before spending compute

   Posterior entropy, top-two separation, and service or capital crises allocate a zero-to-three continuation budget.

   BOUNDED
2. LINE
   Compare the principal macro options

   Compact and expand lead normal search; repair joins when weeds, service, or capital leave the state tactically unfinished.

   SELECTIVE
3. REJECT
   Full-strength ambiguity was destructive

   The first draft abandoned productive late expansion and went 1–11 versus Alpha7. It is preserved as a regression trace.

   −549 AVG
4. CALIBRATE
   Intervene lightly unless the state is unstable

   Ambiguous choices receive a 35% search blend; genuine quiescence crises can use the full correction.

   43 TESTS
5. VERDICT
   Instrumentation improved; playing strength did not separate

   The exact artifact keeps the +2,075 floor but ties Alpha7 on ten of twelve fresh paired games.

   HOLD

Evidence: results/v1_0_0_alpha8_checkpoint.md, exact comparison JSON, rejected search trace, and the 360-row instrumented arena.

1. 01
   Scenario model-predictive control

   Replan a short crop, livestock, land, labor, and liquidation horizon under explicit cash and service constraints.

   ALPHA9A
2. 02
   Poker-style safe resolving

   Keep a robust blueprint; opponent-specific counters fire only when confidence is high and the lower-bound value is safe.

   NO OVERREACTION
3. 03
   Distributional lower-tail value

   Score 10th, 50th, and 90th percentile returns so one catastrophic seed matters—not only the average margin.

   RELIABILITY
4. 04
   AlphaStar / PSRO league

   Train against frozen history, diverse personas, and targeted exploiters; evaluate the full payoff matrix.

   NON-TRANSITIVE
5. 05
   Latent economic dynamics

   After more counterfactual data, learn only the planning features the explicit simulator repeatedly misses.

   LATER

Research and experiment boundaries: results/alpha9_cross_domain_research_backlog.md. League schedule: results/alpha9_league_arena_preregistration.json.

1. GAP
   Stored expansion could outlive its evidence

   The posterior could strongly prefer compact while an older expansion commitment kept spending.

   FOUND
2. VETO
   Break only material expansion regret

   A score and posterior gap can release stale expansion during the capital window.

   ASYMMETRIC
3. FLOOR
   Preserve every frozen control

   Alpha7 finishes the combined panel 10–0 at +7,829.6 average and +2,075 worst.

   SAFE BASE
4. DIRECT
   Fresh comparisons remain neutral

   Seat-reversed splits against Alpha5 and Alpha6 are not evidence of a policy lift.

   HOLD

Evidence: results/v1_0_0_alpha7_checkpoint.md and exact paired-seed comparisons.

1. TRACE
   Day 13→14 is the live failure junction

   The Beyond loss cleared crops, bundled livestock, accumulated 17 weeds, and opened a five-figure bank gap in one day.

   −20,738 FINAL
2. SEARCH
   Compare only four reachable macro options

   Compact, expand, repair, and liquidate receive favorable, base, and adverse continuation values once per day.

   CONSTANT TIME
3. VALUE
   Price downside before headline upside

   The head measures full bundle shock, unused productive capacity, weeds, and negative service slack before altering the existing softmax.

   RISK-AWARE
4. PRESERVE
   Keep the frozen floor

   Exact alpha6 remains 10–0 versus v0.9.2 plus Beyond at +7.8K average and +2,075 worst, with no failures or fallbacks.

   NO REGRESSION
5. ARENA
   Generate tree-ready daily transitions

   Six styles in both seats finish 12–0 and emit 360 rows: 240 train, 120 validation, split by game seed.

   CAPPED
6. HOLD
   Architecture is not yet a promotion

   Four direct alpha5 games net to zero. Fit and validate the option-value tree before claiming a stronger agent.

   HONEST GATE

Evidence: results/v1_0_0_alpha6_checkpoint.md, results/pokemon_bot_architecture_research.md, and the alpha6 arena/dataset files.

1. TRACE
   Find the first shared failure

   On seeds 2700 and 2703, alpha4 led through day 12, then expanded at 48% two-plot utilization and bought eight animals plus land.

   D13 CLIFF
2. CONFIRM
   Require evidence before the next bundle

   Expansion must see either 64% productive utilization or repeated town demand. The choice remains scored rather than hard-coded to an opponent identity.

   ONE LEVER
3. PRESERVE
   Keep the good aggressive environment

   Two Smoothie Shops confirm the Beyond cow engine, preserving alpha4's +8,324 result in both seats.

   NO REGRESSION
4. CONTROL
   Flip the repeatable losses

   The exact artifact goes 10–0 against v0.9.2 at +7.8K average and +2,075 worst. Alpha4 went 6–4 on the same combined panel.

   LOCAL LEAD
5. ARENA
   Improve the same four-style practice set

   Alpha5 finishes 8–0 at +34.0K average, +2.3K above alpha4, with the largest gain against the Sarthak frontier proxy.

   240 ROWS
6. NEXT
   Draw a fresh 20-game gate

   Ten unseen seeds in both seats must finish 20–0 before alpha5 is considered for Kaggle submission.

   HOLD

Evidence: results/v1_0_0_alpha5_checkpoint.md, exact comparison JSON, and the same-seed mini-arena JSON/JSONL.

1. TRACE
   Separate engine deficit from closing leakage

   Beyond led by 14,332 on day 22 with 16 sheep and 16 strawberry plots. The final −20,738 loss cannot be repaired by liquidation alone.

   PRIMARY CAUSE
2. REACH
   Price the remaining calendar

   Seed buying and planting now stop when the crop's first useful yield cannot arrive by day 28. Liquidation releases unnecessary inventory holds.

   TERMINAL JUNCTION
3. REJECT
   Do not starve the closer

   Lower terminal labor and feed floors both reduced the exact Beyond-seed margin, so the frozen artifact keeps alpha3's proven service capacity.

   2 ABLATIONS
4. CONTROL
   Positive value, incomplete consistency

   Alpha4 goes 6–4 against exact v0.9.2-i1 at +7.4K average, but still loses seeds 2700 and 2703 in both seats.

   HOLD
5. ARENA
   Practice against bounded seeded styles

   Four persona proxies in both seats finish 8–0 at +31.7K average and produce 240 daily transitions without failures.

   8 EPISODES
6. NEXT
   Trace the first midgame divergence

   Turn seeds 2700 and 2703 into permanent junction regressions before increasing density or expanding the arena.

   ENGINE FIRST

Evidence: results/v1_0_0_alpha4_checkpoint.md, exact comparison JSON, and the bounded mini-arena JSON/JSONL. Persona opponents are public-behavior proxies, not private agents.

1. OPEN
   Keep the proven mixed launch

   The day-0 wheat, melon, cow, sheep, and six-hand spine remains unchanged. The first junction begins after the opener rather than rewriting it.

   D00–04
2. SCORE
   Compare four reachable continuations

   Compact, expand, repair, and liquidate receive softmax weights from full-bundle funding, productive utilization, work slack, weeds, town support, opponent land, and remaining horizon.

   ONE DAILY CHOICE
3. BREAK
   Commitments are conditional

   A high-confidence branch can harden after two days, but prediction breaks, more than 20% weeds, or negative service slack reopen the decision immediately.

   NO BLIND LOCK
4. RESULT
   Large average lift, incomplete consistency

   The frozen artifact goes 8–0 against v0.9.0, 4–4 against submitted v0.9.1-i4, and 2–2 against v0.9.2-i1 for a 14–6 total at +13.8K average.

   HOLD
5. NEXT
   Trace the two losing environments

   Seeds 2700 and 2703 lose in both seats. Review their town path, capital regime probabilities, and first bank divergence before changing another subsystem.

   CALIBRATE

Evidence: results/v1_0_0_alpha3_vs_v0_9_series_checkpoint.md and the three exact per-control JSON suites. Alpha3 was not submitted.

1. D13
   Pause before a coupled capital jump

   Beyond was nearly level before our farm cleared crops and committed a large animal bundle. The decision must price its full next-day service load.

   MAIN JUNCTION
2. D14
   Winning does not erase a bad state

   The Krupa win also reached 18 weeds with crops cleared. Terminal conversion rescued it, but the same state can fail against a cleaner engine.

   WARNING
3. D24
   Compact play still needs a closer

   Kaito led us by 6,208 on day 24. Reachable liquidation and inventory conversion turned that into a narrow +3,816.

   EXECUTE
4. CORE
   Plots are not production

   Micah's three underfilled plots lost by 35,186 to our one dense plot. Density remains the best expansion confirmation found so far.

   THROUGHPUT

Evidence: results/kaggle_v0_9_1_latest_four_learning.json and raw public replay JSON for episodes 93979595, 93978702, 93977818, and 93976930.

1. STATE
   Record complete daily transitions

   Every row now includes state, branch label, planned hires, capital quantities, action, next state, terminal boundary, bank delta, and terminal return.

   SCHEMA VERIFIED
2. WEIGHT
   Cover first; weight repeats second

   Eight play styles receive one shuffled coverage pass when budget permits. Additional matchups follow explicit expansion, liquidity, and livestock weights.

   8 / 8 STYLES
3. SEED
   Randomize without losing reproducibility

   One arena seed chooses styles and game seeds; each matchup plays both seats against the same opponent and environment.

   PAIRED
4. BOUND
   Checkpoint every episode

   Episode and wall-clock limits stop runaway sessions. The preregistered development arena completed 12–0 without retries.

   NO LOOP
5. NEXT
   Learn the daily branch value

   Fit a compact offline model to hold, expand, compound, restructure, or liquidate decisions; keep hourly survival and routing deterministic.

   OFFLINE RL

Evidence: results/v1_0_arena_loop1_baseline.json, results/v1_0_transition_schema_smoke.jsonl, results/v1_0_arena_loop1_checkpoint.md, and results/v0_9_2_i1_vs_v0_8_4_seed_2200.json.

1. SCOPE
   This is an honest alpha, not a learned v1.0 model

   The arena dataset and transition schema are ready, but no branch scorer has been fitted. Alpha1 freezes the current default policy and excludes the optional land-rush entry point.

   EXACT ARTIFACT
2. UPSIDE
   Large engines win when the town supports them

   Seeds 2602 and 2604 finish 2–0 at +7.6K and +11.9K paired margins. Yarn and smoothie-heavy schemas reward the larger mixed livestock engine.

   4 WINS
3. FAIL
   Expansion is still treated as unconditional

   By day 15 alpha1 owns three plots and 14 animals even on weak paths. On seed 2601 it holds 1.9K cash versus 15.5K for the compact submitted agent, then loses by 30.3K.

   EARLIEST MATERIAL
4. NEXT
   Choose compact or expand before spending

   A finite-horizon capital branch must compare the second and third plots with continued one-plot compounding, then preserve the two winning seeds as regressions.

   v1.0-alpha2

Evidence: results/v1_0_0_alpha1_vs_submitted_checkpoint.md, results/v1_0_0_alpha1_vs_submitted_v0_9_1_i4_seed_2200.json, and results/v1_0_0_alpha1_vs_submitted_v0_9_1_i4_fresh_seeds_2600_2604.json. Alpha1 was not submitted.

1. D00
   Keep a crop-first land rush in the repertoire

   The selectable route buys a second plot immediately, targets 27 wheat, 5 melons, and 6 hands, and delays livestock until the crop engine is moving.

   2 PLOTS D01
2. FIX
   Assign resource work to equipped hands

   Feed and fertilizer tasks are no longer routed to workers who lack wheat or fertilizer—the exact failure pattern found in the Fonge replay.

   38 TESTS PASS
3. REJECT
   Do not hand off abruptly on day 6

   The core policy inherited land without engine continuity and fell from a −12.2K to a −19.8K average margin.

   0–2
4. REJECT
   Do not force the third plot

   A broader cash-and-capacity gate worsened the average margin to −23.5K. Undeployed animals and incomplete work are a veto, not evidence for more land.

   0–2
5. NEXT
   Optimize day-0 workhand throughput

   The route planted 26 tiles by day 1 versus roughly 32 in the replay path. Movement density and action throughput are now the bottleneck.

   ROUTING FIRST

Evidence: results/aggressive_route_playbook.md, results/aggressive_crop_first_vs_v0_9_2_seed_164828330.json, and the rejected handoff and capacity-gate checkpoints. This route is research only and was not submitted.

1. REJECT
   Do not mistake lower cost for efficiency

   A five-hand opener lost both seats by 11.1K average. Six hands remain the minimum serviceable launch crew.

   0–2
2. REJECT
   Do not force raw plot density

   Holding a 42-strawberry target through day 15 created 31 active plots but lost both seats by 14.9K. Density without route slack displaced the livestock engine.

   0–2
3. CAPITAL
   Credit assets already purchased

   When prices or town shops change the preferred livestock mix, keep deployed and held animals and buy only the remaining phase slots. No more duplicate rebalancing capital.

   +938 ALONE
4. SERVICE
   Protect the compounding engine first

   Harvest remains first; feed and care now precede routine watering. A 400–500 coin animal and its future output are protected before one marginal crop-day action.

   10–0 DEV
5. HOLDOUT
   Freeze, redraw, then grade

   Three newly drawn seeds in both seats finish 6–0, averaging 67.4K terminal bank and +10.8K over exact v0.9.0.

   NO RETUNING
6. REGRESSION
   Keep the one losing town path visible

   Fresh direct gates finish 19–1 against both v0.7.1-i5 and v0.8.4. Seed 2200 seat 0 needs bounded livestock restructuring, remembered expansion milestones, and earlier liquidation.

   v0.9.2 FIRST

Evidence: results/v0_9_1_checkpoint.json, results/v0_9_1_i4_vs_v0_9_0_release_twenty.json, results/v0_9_1_i4_vs_v0_7_1_i5_release_twenty.json, and results/v0_9_1_i4_vs_v0_8_4_release_twenty.json.

1. EARLY
   Launch the shared high-value engine

   Start from 7 wheat, 12 melon, 2 cows, and 2 sheep. Keep a six-hand safety floor because the leaders' sparse labor timing proved brittle under our more general router.

   D00–07
2. MIDDLE
   Convert cash into serviceable throughput

   Open land only when the mixed engine is placed, then build toward strawberries, 14 animals, and 12 hands. Land runs once per day and preempts the market-order cap.

   D08–21
3. DYNAMIC
   Vary the portfolio, not the engine spine

   Milk, wool, and town shops can tilt cows versus sheep. Opponent attention remains an overlay; it cannot force a weak core branch or premature expansion.

   ENVIRONMENT FIRST
4. LATE
   Separate survival from investment

   Closing new-animal investment no longer closes feed purchases. Existing livestock stays productive while recurring plots expire and the reverse planner liquidates output.

   D22–29
5. GATE
   Require exact superiority

   The same main.py finishes 20–0 versus v0.7.1-i5 and 20–0 versus v0.8.4 across seeds 1900–1909 in both seats.

   40–0 TOTAL

Evidence: results/kaggle_top20_core_sample.json, results/v0_9_0_checkpoint.json, results/v0_9_0_i5_vs_v0_7_1_i5_twenty.json, and results/v0_9_0_i5_vs_v0_8_4_twenty.json.

1. TRACE
   Recognition arrived after compounding

   Sarthak's land engine and Vignesh's livestock engine both committed before v0.8.4's day-10 classifier could matter.

   EARLIEST FAILURE
2. CORE
   Fund the first production engine

   The opener allocates capital to four cows, twelve planned wheat cells, nine melons, and six hands. Land waits for three placed animals, seven serviced wheat cells, and conversion cash.

   ANIMAL FIRST
3. ATTENTION
   Adapt livestock without thrashing

   Town economics can change the livestock family after day 11, but deployed animals remain committed unless the alternative has a material score edge.

   STICKY DENSITY
4. GATE
   Keep the candidate held

   The Sarthak proxy finishes 2–0, but Vignesh remains 0–2 and the exact v0.8.4 control finishes 2–4. Correct direction, insufficient reliability.

   DO NOT SUBMIT

Evidence: results/v0_8_5_checkpoint.json, results/v0_8_5_land_gate_sarthak.json, results/v0_8_5_land_gate_vignesh.json, and fresh-process seeds 1800–1802.

1. QUESTION
   Does knowing the plan actually help?

   Forced sheep and goose counters lose all eight trials. A fifth cow wins all four trials against the exact v0.8.0 cow engine, averaging +2,503 with a +652 worst margin.

   COUNTERFACTUAL
2. RECOGNIZE
   Infer the plan from public commitments

   Days 10–12 identify the one-quadrant melon/wheat launchpad, then later evidence confirms its crop conversion and livestock family. A two-quadrant frontier opener does not trigger the response.

   NO IDENTITY ORACLE
3. RESPOND
   Stay lean and add one productive edge

   The response preserves v0.8.3 crop memory and activates a sticky fifth-cow target instead of abandoning the working engine for a dramatic species pivot.

   4–0 VS v0.8.0
4. GATE
   Separate upload readiness from promotion

   The fresh v0.7.1-i5 gate finishes 13–3–4 with a +1,171 average, zero failures, and 111 ms maximum action time. It is safe to upload, but it has not earned the 20–0 promotion label.

   EXPLORATORY

Evidence: results/v0_8_4_checkpoint.json, results/known_v080_cow5_vs_v080_cow_seeds.json, results/v0_8_4_i2_vs_v0_8_0_classifier_smoke.json, and results/v0_8_4_vs_v0_7_1_i5_gate_part1.json + part2.json.

1. HISTORY
   v0.8.0 modestly improved v0.7.1

   The fixed control was 1–1 with zero average margin. On six matched persona games, v0.8.0 improved average margin by 749 and reduced worst measured latency from 3.60 seconds to 767 ms.

   LEANER · BETTER
2. REJECT
   Do not restore v0.7.2 frontier bulk

   The full frontier candidate finished 1–7 and lost both fixed expansion seats. More land, livestock, and labor were not automatically productive.

   STILL CLOSED
3. TRACE
   Find the stranded-seed failure

   v0.8.2 overwrote its locked crop after animals deployed. A temporary tomato vote purchased seeds; the next strawberry vote abandoned them, creating repeat −50 losses.

   EARLIEST FAILURE
4. FIX
   Keep the engine commitment stable

   Attention and softmax continue updating, but only a proven material break may reopen the crop family. The diagnostic improves from 3–3–2 to 3–1–4 and the 12-game holdout finishes 1–1–10.

   BETTER · HELD

Evidence: results/v0_8_3_checkpoint.json, results/v0_8_2_vs_v0_8_0_density_diagnostic.json, results/v0_8_3_vs_v0_8_0_diagnostic_8.json, and results/v0_8_3_vs_v0_8_0_holdout_12.json.

1. ATTEND
   Make the three modules operative

   Operations protects service work, opponent attention reacts to commitments and prediction breaks, and horizon attention increases liquidity pressure near the terminal state.

   IMPLEMENTED
2. SOFTMAX
   Blend four strategies once per day

   Cash, recurring crops, livestock, and liquidity receive normalized utilities and phase-aware probabilities. The mixture changes engine-family decisiveness without suppressing mandatory chores.

   BOUNDED
3. TRAIN
   Reject unstable density branches

   Five matched persona cases showed that both a universal four-cow cap and a frontier-only cap were brittle. The final tree preserves v0.8.0 asset counts and limits softmax to family selection.

   2 BRANCHES REJECTED
4. GATE
   Positive average is not enough

   Fresh seeds 1600–1609 finish 4–10–6 versus exact v0.8.0. One seed wins both seats by 4,650, but the −1,265 worst loss and ten total losses make submission unsafe.

   FAILED 20–0

Evidence: results/v0_8_2_checkpoint.json, results/v0_8_2_i3_training_arena_5.json, results/v0_8_2_vs_v0_8_0_gate_part1.json, and results/v0_8_2_vs_v0_8_0_gate_part2.json.

1. OBSERVE
   Wait for evidence before buying livestock

   The agent makes no livestock purchase until day 13, so the specialist gathers daily crop, town, market, and opponent evidence through day 12.

   EVENT DRIVEN
2. VOTE
   Let the whole engine portfolio vote

   Tomato and strawberry are crossed with goose, cow, and sheep. A soft ensemble keeps close alternatives visible instead of trusting one noisy top score.

   6 PATHS
3. LOCK
   Require stability, then stop thrashing

   A short daily history and minimum vote edge select the family once. Cow and goose stay at four animals; sheep may retain the lean engine's proven fifth slot.

   BOUNDED
4. RESULT
   Improve the paired control without claiming readiness

   Against four inferred personas in both seats, v0.8.1 gains 482 coins per game over v0.8.0: three games improve and five are unchanged. The overall record is still 1–7.

   BETTER · HELD

Evidence: results/v0_8_1_checkpoint.json, results/v0_8_0_matched_v0_8_1_persona_arena.json, results/v0_8_1_i2_persona_arena.json, and results/v0_8_1_i2_top_frontier_runtime_check.json.

1. CORE
   Preserve the promoted lean behavior

   The fixed both-seat comparison averages exactly zero against v0.7.1-i5, with no failures or suspicious fallbacks.

   NON-REGRESSION
2. SELECT
   Evaluate one macro branch per day

   The selector records lean, selective, or frontier intent with payback, service slack, opportunity, opponent land, and hysteresis evidence.

   IMPLEMENTED
3. ARENA
   Improve the baseline without claiming promotion

   On six identical persona games, v0.8 and v0.7.1 both finish 1–5; v0.8 improves average margin from −4,936 to −4,187 and cuts the worst measured action from 3.60 seconds to 767 ms.

   BETTER · HELD
4. REJECT
   Do not force a late frontier catch-up

   Relaxing readiness improved one frontier seat but worsened the other enough to produce a −6,569 average. The experiment was reverted.

   FAILED BRANCH

Evidence: results/v0_8_0_checkpoint.json, results/v0_8_0_i5_targeted_arena.json, results/v0_7_1_i5_targeted_v0_8_baseline.json, and results/v0_8_0_i6_top_frontier_arena.json.

1. REPLAY
   A deterministic opener creates the launchpad

   The two sampled leader games use identical actions through the first 144 turns: 12 melons, 7 wheat, 2 cows, 2 sheep, and 5 work hands before adapting.

   OBSERVED
2. CASCADE
   Land is coupled to deployable capacity

   The winning path opens quadrant two around day 6 and quadrant three around day 11, then compounds recurring crops, mixed livestock, fertilizer sales, and 8–12 daily hands.

   IMPLEMENTED
3. ARENA
   Eight bounded games produce milestone rewards

   Every run records win/loss, quadrant timing, recurring density, livestock deployment, day-24 bank, fallbacks, and action latency. The personas are public-commitment proxies, not copied private agents.

   OFFLINE LEARNING
4. RESULT
   The upper-cap branch is still too expensive

   v0.7.2 hit quadrant two by day 7 in all arena games and exceeded 15 recurring plots plus 8 livestock by day 15, but never reached a 60k bank by day 24 and finished 1–7.

   DO NOT PROMOTE

Evidence: results/kaggle_top_agent_expansion_sample.json, results/v0_7_2_arena3_training_arena.json, and results/v0_7_2_arena3_vs_v0_7_1_i5_smoke.json.

1. BASE
   Always run the proven lean engine

   Keep v0.7.1 routing, dynamic wheat and labor reserves, reverse closeout, opponent memory, and daily traces as the low-cost default.

   ALWAYS ON
2. v0.8
   Compare three macro paths

   Once per day, estimate lean, selective expansion, and frontier terminal value. Charge each path for land, labor, travel, feed, switching, and stranded inventory.

   DAILY
3. GATE
   Turn on density only when it pays

   Short opponent prediction and town simulation activate near land, portfolio, or liquidation decisions. Hysteresis and bailout thresholds prevent oscillation or sunk-cost escalation.

   EVENT DRIVEN
4. TRAIN
   Learn the selector from counterfactual arenas

   Run the same seeds through all three paths, label the winning branch, and compile the strongest splits into a shallow tree rather than shipping a large learner.

   OFFLINE
5. v0.9
   Add deeper reasoning behind the same gate

   Only after v0.8 is calibrated, add compact forest votes, crop-livestock portfolio specialists, and short stochastic rollouts at high-impact decision points.

   DEFERRED

Promotion order: v0.8 must first preserve the v0.7.1 control gate, then avoid losing any persona family, then demonstrate a genuine 60k+ upper-cap path. Evidence plan: results/v0_8_density_roadmap.json.

1. D00–04
   Melon liquidity + livestock immediately

   The modal opener places 12 melons, 7 wheat, 2 cows, and 2 sheep on day 0; wheat is feed infrastructure, not the strategic objective.

   BOOTSTRAP
2. D05–11
   Convert into strawberries and land

   Recurring crops start around day 5, the second quadrant appears around day 6, and a third around day 11.

   EXPAND
3. D12–23
   Recurring crop + animal throughput

   All 264 sampled midgame-days classified as recurring-crop core, typically strawberry with cows, sheep, wheat feed, and fertilizer sales.

   COMPOUND
4. CAUTION
   This band is not a diverse training arena

   The action schedules are extremely similar. Treat this as one strong archetype and add deliberately different local opponents before fitting a forest.

   BIAS CHECK

Compact evidence: results/kaggle_rank_100_200_midgame_sample.json. Labels describe visible commitments, not private code or intent.

1. 01
   Keep within-match memory

   Carry the previous posterior, bank movement, asset growth, crop mix, animal mix, labor, and unlocked land from one day to the next.

   IMPLEMENTED
2. 02
   Estimate productive threat

   Compare terminal value, productive value, occupied scale, animals, and land—not just the visible bank.

   IMPLEMENTED
3. 03
   Leave replay breadcrumbs

   Emit one compact decision trace per day and explicit error records so future ladder logs explain what the policy believed.

   IMPLEMENTED
4. 04
   Reject naive land expansion

   Ten extra recurring plots lost all four smoke games by about 11k on average because travel and chore load diluted execution.

   REJECTED
5. 05
   Make wheat conditional

   Size wheat from livestock feed, a rolling three-day reserve, town demand, price, shed stock, and days remaining.

   IMPLEMENTED
6. 06
   Predict, verify, then switch

   Forecast visible crop and animal additions three days ahead; lower strategy inertia only after a material prediction break.

   IMPLEMENTED

1. HIRE
   Size labor from chores + route span

   Replace fixed six/eight-hand quotas with daily work volume, travel footprint, and terminal transport demand.

   IMPLEMENTED
2. ROUTE
   Keep valid assignments sticky

   A hand stays on its target unless the task disappears or urgent work preempts it; assigned workers are no longer mistaken for idle carriers.

   IMPLEMENTED
3. COMBO
   Simulate crop × livestock dependencies

   Score six recurring-crop/livestock pairs with feed, town synergy, fertilizer, labor, and incumbent investment. The score informs analysis without forcing a switch.

   DIAGNOSTIC
4. FIX
   Seed 503 · seat 1 · +380

   The terminal planner stopped feeding, but pickup tasks still withdrew wheat and raced liquidation. Making pickup honor the same obligation set recovered the final inventory.

   2 / 2 REGRESSION
5. NEXT
   Mixed-conversion persona · 0 / 2

   The promoted checkpoint still loses by 13,434 in both seats to the deliberately diversified conversion engine.

   ROBUSTNESS GAP

Evidence: results/v0_7_1_i5_seed503_vs_v0_7_0.json, results/v0_7_1_i5_vs_v0_7_0_exact_20.json, and results/v0_7_1_i5_persona_sanity.json.

1. D26
   The policy was ahead

   Our bank led by 1,751 coins; the phase strategy itself remained competitive.

   28,239 vs 26,488
2. D27
   The opponent liquidated first

   The frozen bot sold wheat and milk while our shed accumulated output.

   margin −1,175
3. D28–29
   The safety fallback hid an exception

   Undefined closeout variables returned PASS for every remaining recorded action, yet Kaggle still marked the episode DONE.

   47 empty turns
4. FIX
   Define the closeout state and grade fallback

   v0.3.1 sold the surplus, kept operating, and won the same seed by 7,945 coins.

   20 / 20 fixed gate

Trace evidence: results/phase_tempo_v0_3_seed5_loss_trace.json and results/weighted_strategy_v0_4_seed5_regression_trace.json.

Three promoted mechanisms: score crops and livestock after feed, care, labor, tile time, demand, and crowding; choose mixed strategy groups by probability-weighted utility; and assign all workers together by deadline and travel cost. The economics layer also keeps fertilizer only when applying it is worth more than selling it.

Checkpoint lesson: the first global router passed seeds 0-9 but hit a 299-second action on held-out seed 10. Bounding the optimizer to the best three task alternatives per worker reduced that regression to 75.6 ms and won both seats. The final artifact then went 20-0 against v0.5.0, 20-0 against v0.5.1, and 20-0 against v0.5.1 on held-out seeds 10-19.

Evidence: results/v0_6_1_vs_v0_5_0_exact.json, results/v0_6_1_vs_v0_5_1_exact.json, results/v0_6_1_vs_v0_5_1_heldout.json, results/v0_6_1_seed10_runtime_regression.json, and results/qa_v0_6_1_packaging.json.

What improved: town-demand and crop ROI forecasts, ten opponent archetypes, mixed recurring-crop allocation, shed-capacity lookahead, higher-priority livestock placement, and planned-cell weed recovery.

Promotion boundary: the repeated seed 5, seat 0 loss means v0.5.1 does not replace v0.4 yet. Evidence: results/dense_predictor_v0_5_1_vs_v0_4.json, results/dense_predictor_v0_5_1_vs_submitted_v0_2_1_heldout.json, and results/qa_v0_5_1_packaging.json.

Readiness boundary: v0.4 is ready if the requirement is reliable superiority to submitted v0.2.1. It is not yet evidence of superiority to unfamiliar ladder architectures.

Evidence: results/qa_v0_4_vs_submitted_v0_2_1_heldout.json and results/qa_v0_4_packaging.json. No Kaggle submission was made.

1. 01
   Freeze the predecessor

   Preserve the exact artifact, checksum, trace set, and known failure.

   CONTROL
2. 02
   Pass regressions

   Known failure seeds, both seats, strategy probes, and non-fallback execution must all pass.

   CORRECTNESS
3. 03
   Win the fixed suite

   Require 20 / 20 against v0.2.1 before widening the opponent library.

   LOCAL GATE
4. 04
   Generalize before Kaggle

   Play held-out seeds and multiple strategy families so one familiar opponent does not become the objective.

   ROBUSTNESS

Game-specific claims are grounded in the official Kaggriculture environment reference ↗.

1. 01
   Observe commitments

   Land unlocked, structures built, crop ages, animal mix, visible workers, bank, and shared price movement.

   raw signals
2. 02
   Score candidate archetypes

   Expansion rush, livestock compound, recurring crops, short-cycle cash, or mixed/unknown.

   probability + confidence
3. 03
   Estimate their next constraint

   Cash shortage, feed demand, care load, harvest wave, storage pressure, or liquidation deadline.

   forecast
4. 04
   Select a bounded response

   Differentiate the portfolio, preserve liquidity, accelerate a sale, or stay on the baseline when evidence is weak.

   policy modifier
5. 05
   Update after evidence

   Decay stale beliefs and compare predicted farm changes with what actually appeared.

   calibration

1. FEATURES
   Visible commitments

   Crop counts, animals, unlocked land, labor, match day, and bank trajectory.

   x
2. LOGITS
   Weighted strategy scores

   Each strategy receives evidence for and against it; no single threshold owns the decision.

   z = Wx + b
3. SOFTMAX
   Belief distribution

   Short-cycle crop volume, melon rush, recurring crops, livestock families, labor, land expansion, and mixed play sum to 100%.

   P(strategy)
4. UTILITY
   Expected counter value

   Score each response across every strategy, weighted by belief and reduced by switching cost and execution risk.

   Σ P(s)U(a,s)
5. UPDATE
   Observe and recalibrate

   Compare the predicted commitment with the next farm state, then revise weights without erasing alternatives.

   Pₜ₊₁

The distribution above is illustrative. v0.7 updates its real posterior once per day from public farm state and the prior day’s belief; calibration still requires a broader opponent library.

1. STATE
   Compress the board

   Bank, assets, care deadlines, shed, routes, prices, town demand, opponent hypothesis, and turns remaining.

   sₜ
2. OPTIONS
   Generate feasible plans

   Baseline continuation plus a small set of meaningful deviations: expand, diversify, accelerate, defend, or liquidate.

   3–8 plans
3. BRANCH
   Sample uncertainty

   Opponent response, weed rolls, future shop unlocks, price movement, and execution slippage.

   many futures
4. SCORE
   Estimate finish outcomes

   Prioritize probability of finishing ahead; use expected bank and failure risk as diagnostics.

   P(bank₇₂₀ > opponent)
5. ACT
   Execute one safe step

   Take the first action from the best robust plan, observe the new state, and simulate again.

   receding horizon

1. Days 00–02
   Bootstrap

   Create liquidity and mandatory feed capacity while keeping the opener reversible.

   FUND
2. Days 03–07
   Opener build

   Validate the first yield path, observe rival commitments, and expand only when travel capacity exists.

   PROBE
3. Days 08–11
   First conversion

   Harvest the opener and choose the recurring, livestock, or cash engine with switching cost included.

   COMMIT
4. Days 12–15
   Engine commit

   Scale the selected engine; wheat becomes a calculated feed and liquidity reserve.

   BUILD
5. Days 16–19
   Throughput balance

   Match labor, watering, feed, storage, and routes to productive capacity.

   BALANCE
6. Days 20–23
   Engine conversion

   Stop branches that cannot repay and convert fertilizer and mature output into score.

   CONVERT
7. Days 24–27
   Terminal optimize

   Preserve only work with reachable terminal value and protect shed capacity.

   PRUNE
8. Days 28–29
   Liquidate

   Keep the exact remaining feed reserve, return carried output, and sell reachable surplus.

   CLOSE

1. D00–04
   Bootstrap · 55 / 15 / 30

   Operations leads while the mixed opener stays reversible. Opponent evidence is sparse; horizon protects feed and a cash option.

   OPS · OPP · HORIZON
2. D05–08
   First land · 45 / 15 / 40

   Compare compact continuation with the complete expansion bundle. Harden only above 70% branch probability for two days with 90% service completion.

   T = 0.60
3. D09–12
   Engine commit · 35 / 25 / 40

   Town shops, relative product value, and opponent commitments select recurring, cow, sheep, or mixed play. Require an 18-point posterior edge.

   HARDEN
4. D13–18
   Compound or repair · 45 / 20 / 35

   Keep the engine only while realized return is positive and backlog shrinks. Weeds above 20% or repeated missed service reopens the branch.

   BREAK
5. D19–29
   Convert and close · 35 / 10 / 55 → 30 / 05 / 65

   Horizon dominates. Reject purchases whose production, transport, and sale cannot complete; preserve only exact feed and executable terminal value.

   T = 0.35

Softmax rule: use temperature 1.0 while observing, 0.60 at the decision, and 0.35 after hardening. Reset to 0.90 only when town value shifts 25%, forecast error exceeds 15%, or serviceability breaks.

1. 01
   Prevent irreversible loss

   Water/feed deadlines, overflow, invalid actions, and products about to decay or cap.

   SURVIVAL
2. 02
   Protect the terminal bank

   As turn 720 approaches, stop purchases that cannot repay and liquidate all reachable value.

   EXIT
3. 03
   Capture time-sensitive value

   Harvest windows, temporary scarcity, shop-driven demand, and care bonuses.

   OPPORTUNITY
4. 04
   Read the opponent

   Update the archetype only when visible evidence materially changes the likely response.

   INFERENCE
5. 05
   Run lookahead

   Compare the baseline with bounded counter-strategies and choose the robust branch.

   SIMULATE
6. 06
   Compound the farm

   Plant, build, hire, expand, and route workers using the remaining action budget.

   GROWTH

The agent is not one giant brain. It is a short chain of specialists. Each specialist answers one question, then a safe executor turns the final plan into valid hourly work.

1. SEE
   Read the farm

   Summarize cash, land, crops, animals, workers, weeds, town demand, public opponent commitments, and days remaining.

   STATE
2. SAFE
   Protect non-negotiable work

   Harvest, feed, animal care, overflow prevention, and reachable final sales cannot be outvoted by a clever strategy.

   REFLEXES
3. ATTEND
   Move three spotlights

   Operations, opponent, and horizon attention decide which evidence deserves more influence in this position.

   SOFTMAX WEIGHTS
4. PROPOSE
   Suggest an engine

   Cash, recurring crops, livestock, and liquidity strategies compete as scored proposals rather than fixed scripts.

   POLICY PRIOR
5. CHECK
   Test the capital package

   Scenario MPC asks whether compact, expand, repair, or liquidate is affordable, serviceable, dense, and able to pay back.

   4-DAY LOOK
6. WORK
   Route every hand

   A task queue assigns useful work globally, keeps routes stable, and avoids wasting movement on a new daily whim.

   EXECUTOR
7. CLOSE
   Convert the farm before time expires

   The terminal planner stops investments that cannot mature and sells reachable inventory before the last day.

   END GAME

Important boundary: each match starts fresh. The live bot does not remember opponents across Kaggle games; learning happens offline in our arena, then a tested tree or parameter set is packaged into a later version.

MPC is a rolling plan. Instead of committing to one forecast for the rest of the game, it compares a few feasible near-term packages, executes only today's choice, observes what actually happened, and rebuilds the plan tomorrow.

1. 01
   Observe the current farm

   Cash, productive density, animals, work capacity, weeds, town demand, opponent pressure, and days remaining become the planning state.

   STATE
2. 02
   Generate reachable packages

   Compact or densify, expand, repair, and liquidate specify the land, crops, animals, labor, feed, and reserve they require.

   OPTIONS
3. 03
   Reject impossible plans

   A branch fails if the full bundle lacks cash, service capacity, productive density, or enough days to repay.

   CONSTRAINTS
4. 04
   Stress three futures

   Favorable, base, and adverse demand, weed, and service paths expose a plan whose headline value hides a bad downside.

   SCENARIOS
5. 05
   Choose robust value

   The score blends expected return with the weaker scenario and the value of keeping future choices open.

   ROBUST SCORE
6. 06
   Execute one step and replan

   Only the current macro choice is handed to the hourly executor. New evidence can change tomorrow's plan.

   FEEDBACK LOOP

1. FLOOR
   Deterministic survival rules

   Never let search or learning displace mandatory harvest, feed, care, overflow, or reachable terminal sales.

   HARD RULES
2. LIBRARY
   Options propose engine paths

   Early evidence activates a small set of interruptible plans rather than hundreds of isolated actions.

   HIERARCHY
3. TRANSFER
   Successor features reprice the library

   Town and market changes alter the reward weights without erasing what each engine is expected to produce.

   ADAPT
4. CAPITAL
   MPC checks feasibility and downside

   The daily controller chooses a robust package under cash, service, density, weed, and payback constraints.

   ALPHA9 FLOOR
5. ROUTE
   Rollout schedules the critical work

   Limited lookahead improves hand assignments only at service bottlenecks and end-game conversion junctions.

   EXECUTION
6. SHIELD
   Uncertain learning falls back safely

   Unsupported tree leaves and policy choices use the proven Alpha9 action.

   NO GUESSING
7. LEARN
   Arena telemetry updates the archive

   Losses add regression seeds; diverse winners populate strategy cells; no live match writes state into another match.

   OFFLINE ONLY

1. A1
   Interruptible engine options

   Launch, densify, expand, repair, and close now expose a multi-day intent and its break conditions.

   BUILT · HOLD
2. A2
   Auxiliary farm score heads

   Capacity, service, capital, runway, and trajectory now reweight compact, expand, repair, and liquidate.

   BUILT · TESTING
3. A3
   Calibrated option value

   Proactive saturation was tested; its immediate version lost four games and its staged lower-tail version finished 12–0–8.

   BUILT · HOLD
4. A4
   Capacity-path mixture

   Dense livestock, land compound, and lean liquidity now receive separate downside-aware scores behind a baseline confidence shield.

   BUILT · ALPHA5
5. GATE
   Beat Alpha9 directly

   Alpha5 preserves every frozen floor and finishes the exact submitted-agent panel 20–0 at +5,142 average.

   PASSED

1. 01
   Freeze v0.2.1 as control

   Keep the accepted artifact unchanged so every candidate can play both seats against the exact ladder baseline.

   known opponent
2. 02
   Pass local gates

   Complete self-play, starter matches, both seats, varied seeds, and the late-game liquidation test.

   zero runtime failures
3. 03
   Enter the ladder

   Kaggle first validates the bot against a copy of itself, then matches it with similarly rated submissions.

   public evidence
4. 04
   Wait for signal

   Record rating, rank, and episode count at consistent checkpoints instead of reacting to every move.

   50 · 100 · 250 episodes
5. 05
   Classify losses

   Tag opponent archetype, first strategic divergence, execution failure, and terminal inventory.

   failure taxonomy
6. 06
   Change one mechanism

   Improve opponent modeling, lookahead, attention, or a strategy rule—then submit a new version.

   causal iteration

Interpretation boundary: Alpha5 satisfies the requested Alpha9 release gate, but its narrow Alpha4 regression leaves a choice between submitted-agent qualification and strict local dominance. Neither proves leaderboard superiority.

def agent(obs):
    return {
        "farmer": ["PASS"],
        "hands": [],
        "market": []
    }

Frozen controls: exact submitted Alpha9, v0.9.2, the two observed loss seeds, and the persona arena. Alpha2 is packaged locally and held until the extended release panel finishes.

Packaging lesson: v0.2.0 failed before play because agents collided with a Kaggle-installed module. v0.2.1 packages one self-contained main.py and passed validation.

Submission and ranking behavior: official Kaggriculture evaluation section ↗.

Prices move with supply and town demand.

1. FLOOR
   Deterministic safety kernel

   Harvest, feed, care, weeds, overflow, and reachable liquidation remain hard rules.

   DO NOT LEARN
2. TEACH
   Exact counterfactual search offline

   Replay hard junctions and label coherent lean, dense, land, repair, and close options with 32–128 exact rollouts.

   SLOW TEACHER
3. DISTILL
   Compile judgment into a tiny tree

   Export numeric thresholds for option value, downside, deployment lag, and terminal conversion.

   FAST STUDENT
4. SCHEDULE
   Minimize time to a productive slot

   Connect every animal or plot purchase to its pasture, feed, building, and farmhand prerequisites.

   FIRST FIX
5. SEARCH
   Think longer only at unstable junctions

   Use a small sequential-elimination budget when options are close or capital is difficult to reverse.

   BOUNDED
6. SHIELD
   Require calibrated evidence

   The learned proposal acts only when its lower bound clears Alpha5 and all feasibility tests pass.

   SAFE IMPROVEMENT

Full plan: results/v2_0_series_roadmap.md. No live LLM, cross-match state, or unbounded search is planned for the submitted artifact.

1. TRACE
   Productive-slot debt is observable

   New episodes record animals in the shed, compatible empty structures, blocked deployment, deployed livestock, wheat reserve, and feed runway.

   ACTIVE
2. CORPUS
   1,821 junctions from 81 exact episodes

   Twenty-three seeds are grouped into 1,444 train and 377 validation rows; repeated actions from one seed cannot leak across the split.

   COMPLETE
3. FOCUS
   651 deployment-debt states

   The sampler also found 105 cash-versus-capacity conflicts and 53 land-capital decisions for exact counterfactual labeling.

   QUEUE
4. PLAY
   Human training arena

   The exact rules engine can now freeze Alpha5, accept manual moves, compare the coach, persist checkpoints, and emit training JSONL after a complete season.

   LOCAL
5. BOUND
   Observations are not counterfactuals

   All archived arena rows in this first corpus were winning policy traces. They identify junctions, but cannot prove an unplayed branch is better.

   NO PROMOTION
6. NEXT
   Replay seed 3304 first

   Label purchase, structure, feed, and workhand schedules with exact rollouts; only integrate a scheduler that flips both seats without harming Alpha5 or Alpha9.

   A1 GATE

Evidence: results/v2_0_replay_junction_corpus.json. Generator caps: 12,000 rows or 20 seconds; this run completed in under one second.

1. CEILING
   The machinery already clears 90K

   Alpha5 averaged 92.9K against passive market load. The missing value appears under pressure, not in the raw production rules.

   BASELINE
2. MODEL
   Price serviceable productive slots

   The transparent simulator combines crop clocks, livestock output, feed, land, workhands, weeds, town demand, and nonlinear market crowding.

   EXACT + ANALYTIC
3. ROUTE
   Three plots before four

   The robust neutral hypothesis targets 56 productive cells, seven hands, four cows, and four sheep. Plot four waits for labor capacity.

   TESTING
4. ABLATE
   Blindness exposed the overweighted reaction

   The first Sarthak shard rose from 55.0K to 72.1K, but an earlier lean shard lost 3.3K. The signal justified a core-first architecture—not a blind final agent.

   MIXED RESULT
5. FIX
   Expansion now hands off to recurring density

   The first trace peaked at 41 productive cells, then collapsed after one-time crops. v2d uses a selective bundle and a 26-tile recurring target.

   EXECUTION
6. HOLDOUT
   Pressure output improved without losing the floor

   Fresh paired means rose 5.2K against Sarthak and 12.3K against lean liquidity. Passive output stayed within 193 while its minimum rose 3.5K.

   FREEZE SEEDS
7. ELITE
   Different openings, one invariant

   Sampled elite wins reached 111K and 146K by preserving dense crops and livestock with near-zero midgame weeds.

   PUBLIC REPLAYS
8. CLOSE
   Our sale engine is not the main late deficit

   Public winners added 18.6K–25.4K after day 28; v2d already adds about 24.6K after day 24. It reaches the close with too few live crops.

   DIAGNOSED
9. REJECT
   Late density without service value

   A four-cell wheat wave added 3.75 productive tiles on day 24 but lost 194 mean bank and 1,770 at the floor. More cells did not mean more value.

   NO PROMOTION
10. RESIDUAL
    Opponent intelligence stayed in its lane

    The capped late tiebreak activated at one close junction but changed no outcomes across 16 pressure episodes.

    DIAGNOSTIC
11. NEXT
    Stagger the cohort before the scramble

    Test days 18–21 replacements with explicit workhand and weed costs. Terminal value—not crop count—is the gate.

    CORE FIRST

Model and evidence: results/v2_0_core_economy_checkpoint.md, results/v2_0_late_game_observations.md, and results/v2_0_late_residual_checkpoint.md. v2d remains the held candidate; no submission was made.

1. A1
   Productive-slot scheduler

   Flip the seed-3304 shed-delay loss by pricing pasture, feed, labor, and deployment together.

   START HERE
2. A2
   Counterfactual teacher set

   Label disagreement states with exact rollouts and keep game seeds disjoint across train and holdout.

   ACTIVE DATA
3. A3
   Distilled mixed policy

   Ship a shallow tree ensemble behind the Alpha5 confidence shield.

   COMPACT AI
4. A4
   Residual value model

   Learn only repeated simulator error in cash, service, density, and terminal conversion.

   CALIBRATE
5. BETA
   Adversarial league

   Exercise Alpha4, Alpha5, Alpha9, and seven engine specialists across a capped 100-game holdout.

   DIVERSE GYM
6. GATE
   Beat both local floors 20–0

   Require fresh 20–0 gates against Alpha5 and Alpha4, plus 20–0 against submitted Alpha9 and p95 action time below 100 ms.

   BEFORE SUBMIT