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: v0.6.1 is viable but not dominant. Its clearest reviewed failure was strategic, not computational: it treated a cash lead as safety while the opponent built a much larger productive base.

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

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 00–11
   Early · opening

   Establish cash and wheat, run a bounded melon opener, and probe the opponent’s visible commitment before changing course.

   DEVELOP
2. Days 12–21
   Middle · contest

   Build the recurring-crop and livestock engine. Scale labor, collect fertilizer, and switch crops only on strong evidence.

   ADAPT
3. Days 22–27
   Late A · optimize

   Stop long-payback expansion while preserving feed, care, harvests, and existing production. Improve what can still score.

   CONVERT
4. Days 28–29
   Late B · execute

   Keep enough wheat and hands for the remaining chores, then route output to the shed and sell the true surplus.

   CLOSE

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

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: the replay test proves the new model notices the missed compounding signals. It does not prove that the response wins more games; that requires the completed benchmark and later ladder evidence.

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

Frozen controls: submitted artifacts/kaggriculture-v0.2.1.tar.gz and artifacts/kaggriculture-v0.6.1.tar.gz, plus local v0.4.0, v0.5.0, v0.5.1, and v0.6.0 controls. The 12,785-byte artifacts/kaggriculture-v0.7.0.tar.gz is packaged locally but remains on hold.

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.