Files
Soul_Game/FINDINGS.md
T
mmcghen f9e22e872f Run 13: the full world - four towns, resources, bonds, spatial grid
World scaled 3x (768x768) with all four towns at compass points, each
running its own soul culture: North generational, South nature, West
materialist (new profile), East selfish. Two new resources: Ore (metal
tier) and Herb (medicine - mends health, reason to forage, trade good).
Spatial hash grid replaces O(n^2) proximity scans (soul pressure,
neighbor, teacher) - 3.6x headless speedup at ~80 NPCs. First social
verb beyond work: pair-bonds - settled adults who keep warm company pair
up; warmth now accrues from proximity not only gifts; bonds break at
death. Reproduction deferred (bonds-first per plan).

The four-town thesis fully realized on one engine (240-day run):
North know 1.27 clean, South nourish 0.85 clean, West prosperous but
debris 0.23, East hungriest 0.60 and lowest knowledge 1.08. Positive
cultures stay soul-clean, negative ones carry real debris, all from
behavior weights - nothing scripted. 99 bonds, 80 deaths, four living
societies.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-21 17:04:19 -04:00

16 KiB
Raw Blame History

WorldSim — Soak Experiment Findings (July 2026)

Four long-run experiments against the South (Nature) town MVP. Each run: 15 NPCs, 256×256 world, seed 12345, ~240 simulated days (2 years) at accelerated speed. Data: soak_stats.csv (daily), soak_npc_days.csv (per-villager dominant states), soak_gifts.csv (every gift, with names).


Run 1 — Baseline (no seasons): the dead paradise

Setup: constant regeneration, upkeep + food demand only.

Result: perfectly stable, perfectly inert. 200 days: wood held a ±1% band, shelter condition pinned at 100%, zero trades, no roles — every villager Socializing-dominant ~114 of 119 days, 78% of waking life idle. The village completed all construction and expansion by day 34 and then had nothing left to want.

Lesson: a village is a metabolism, not a project. Equilibrium without pressure produces no society: no differentiation, no exchange, no structure. Abundance makes generosity unobservable.


Run 2 — Seasons (regen throttle only): the buffers win

Setup: 30-day seasons; winter cuts food regen to 5%, wood to 40%. Granary added (clay walls — first clay use), need-based gift trigger.

Result: winter became a worked season — gathering 4,889 ticks/day vs 1,857 in summer, wild stocks visibly drained and refilled. But nourishment ran the identical sawtooth year-round, the granary sat untouched at 150/150 through two winters, and trades stayed at zero across 240 days.

Lesson: three compounding buffers (the sustainability floor acting as a ~360-unit hidden savings account, personal pack reserves, wild food budget exceeding winter need) meant scarcity never reached an individual. Pressure that only raises labor doesn't create interdependence.


Run 3 — Frost: the hungry gap

Setup: winter now decays standing food ~7%/day toward 15% (frost kills the larder); granary draws deepened; anti-redeposit rule.

Result: the simulation came alive.

  • First trades in sim history: 21 (17 deep winter, 4 early spring) — exactly when granary-empty met neighbor-with-food.
  • Granary sawtooth textbook: full through autumn, empty by winter day ~25, refilled within 4 days of spring. Both years.
  • The sim reproduced the "hungry gap" — a real agricultural phenomenon neither designed nor named in the code: worst hunger (village nourishment 0.21) fell in early spring, after stores ran out but before the land regrew past harvestable thresholds.
  • Crisis cascaded: the food emergency pulled labor from heating (firewood reserves dipped to 10, condition below 100% for the first time).
  • Balanced: severe hunger, zero lasting health damage, both years.

Lesson: storage limits + decay + regrowth lag ⇒ emergent seasonality of need. Sharing is a response, not a behavior: 480 days of abundance produced zero exchanges; ten days of genuine need produced twenty-one. Still missing: individual roles — identical souls in a global work queue respond identically (winter gather ticks all within 490553).


Run 4 — The contrast test: 3 selfish souls among 12

Setup: villagers M/N/O run the East (Selfish) soul profile. Same engine, same possibility space; only behavior weights differ: low sustainability (strip nodes to ~15% vs the 50% commons floor), low generosity (refuse gifts; excluded from asking), low community bias (no communal work at all), high accumulation (double carry, standing hoard).

Result — the four-town thesis validated:

  • Selfish villager-days: 0 build ticks (vs 104/day for nature souls), gather 69 vs 269, rest 0 (never tired), socializing 1,369 vs 921. A visible leisure class living off a working village.
  • Gift network: 11 gifts, all nature→nature. The selfish never gave and never needed to receive. First individual differentiation appeared: Villager_A gave 4 of 11 gifts — the village's generous one.
  • The 12 communal souls absorbed 15 souls' worth of burden without a single ruin.

The central finding: selfishness strictly won. Most leisure, least fatigue, fullest bellies, zero costs, no pushback. This is not a balance bug — it is the empirical case for the game's soul layer. Economics alone cannot make selfishness cost anything; that is the job of imprints, debris, echoes, and social memory (the Unity POC's warmth-echo system). The soul-consequence layer is not flavor on top of the economy — it is the only thing standing between the South town and being farmed by the East.


Run 5 — The soul layer, uncalibrated: everyone a sinner

Setup: warmth echoes (memory replaces omniscience), refusal debris, commons-line harvest debris, soul-to-soul atmospheric pressure, debris metabolic tax (faster drain, weaker meals).

Result: justice arrived — all 108 refusals belonged to the selfish three, whose health collapsed to 3164 while nature souls held 100 (the selfish never rest, so they have no healing loop — laziness blocking recovery is fully emergent). But two pathologies: (1) the whole village ratcheted to maxed others-regard debris, because the 50% commons penalty line sat above the nature souls' own 45% harvest floor and debris had no clearing path — harm without healing; (2) warmth-weighted asking plus near-total gift transfers produced a 2,501-gift winter food-swapping loop, and the universal debris tax turned the hungry gap into famine (village nourishment 0.00).

Lesson: a moral system needs calibrated innocence and a redemption path, or it converges to universal guilt and reads as noise.

Run 6 — The soul layer, calibrated: the ledger works

Setup: commons line lowered to 40% (below any respectful floor); gifts capped (keep 5, give ≤3, ask only when genuinely hungry); giving heals the giver (0.003 others-regard vs +0.02 per refusal — harm ~7× easier than healing).

Result — the design document, in numbers:

  • Nature souls: debris 0.000 after two years. The innocent are innocent.
  • Selfish souls: 0.225 and climbing. The debris gap is total and driven entirely by how each soul lives.
  • Gift ping-pong dead (2 gifts); the contact log is 18 refusals — in the hungry gap the only villagers with surplus are the hoarders, so the starving ask them, are refused, and learn. The village's warmth map is a map of learned coldness pointing at three names.
  • Material consequences within one lifetime are mild — the selfish stay comfortable. The sim's justice is written in the souls, not the outcomes. That is correct: in the full game, the ledger cashes out at death (heavy debris carries forward; the rebirth verdict; the grave nobody tends). The sim now produces exactly the input reincarnation consumes.

Next

Run 7 — Organic construction v1: the stone-lock collapse

Setup: two towns (South nature+selfish, North generational); all construction quotas removed. Shelters founded after three unsheltered nights (beds are real: 3 + modules); granaries founded from surplus.

Result: a construction boom (8 shelters, multiple granaries by day 10), then total civilizational collapse by day 40 — every building ruined, and both towns spent the remaining 200 days homeless among the monuments. Cause: a priority inversion. The homeless-gather-stone rule preempted every other communal priority, so once ruins made everyone homeless, all communal souls locked onto stone forever — foundations got laid, walls (wood) never did, firewood hauling and granary stocking ceased entirely, and the emergency rule prevented the emergency from ever ending.

Lesson: in a needs-driven economy, priority ordering is the constitution. An urgent rule that outranks the work needed to resolve it deadlocks the society. (Also: surplus-founding without scarcity memory produced granary sprawl by day 5 — buildings must be founded by lived need, not by momentary abundance.)

Fixes → run 8: construction sites outrank homeless stone-gathering (finish the walls before founding another foundation); homeless founding defers while any shelter is under construction; granary founding requires hunger memory + surplus, not surplus alone.

Runs 89 — Organic construction v2: fixed forward, broken backward

Headless debugging (runs driven from the CLI) found and fixed three real bugs: wilderness homes (founders built wherever their pack filled — usually at a stone mine 100+ cells out; homes now rise at the founder's habitual sleeping spot, within reach of the community), sleepless nights (low-fatigue villagers socialized all night, so beds went unused and rough nights — the founding trigger — never occurred; night now sends everyone to bed), and floor-sleeping (arrivals at a full shelter counted as sheltered; a bed must exist when you lie down in it).

Verified working: bootstrap founding near town, ~2,200 construction deliveries, first roofs by day 3, three seasons of stability.

Open problem (later resolved in run 10): the loop cannot recover from loss. The winter crisis ruins shelters and rebuilding stalls.

Run 10 — The mineral question: recovery achieved

Demand diagnostics found the recovery blocker was material, not motivational: homeless villagers correctly demanded stone (462/525 villager-days) but could not get it, twice over:

  1. The commons ethic stranded the stone. Respectful souls won't take the last half of any deposit — but stone doesn't regrow, so once all deposits sat at the floor, the world's remaining stone was morally untouchable and the homeless stayed homeless forever. Fix: the harvest floor is an ethic of regrowth — floors and commons debris now apply only to living, regenerating resources. Minerals quarry to depletion, for every soul equally.
  2. Lossy packs milled the mountains to dust. Idle variety-gathering mined stone as busywork, and the make-room rule vaporized cargo on every demand flap — 900+ stone destroyed in packs. Fix: idle gathering touches only what regrows, and packs may only shed what the land replaces. Minerals are mined on purpose and conserved.

Correction: run 10's original "zero year-2 ruins" claim was contaminated — the headless console wrapper orphaned child sim processes that kept writing to later runs' CSVs. (Instrumentation lesson: kill process trees; distrust out-of-order data.) Clean reruns showed the recovery loop worked but the town cycled — collapse each winter, rebuild each spring — with chronic post-famine hunger.

Run 11 — Enough, and the stone-mules: clean stability

Two fixes closed it: granary satiation (hunger scars motivate storage only until the town's capacity could carry a winter — enough is a real quantity; sprawl of ~10 granaries fell to the sated level), and quarry return (villagers whose packs filled with undroppable stone after rebuilds became starving porters, unable to gather food while carrying a house on their backs; unneeded stone now goes back to the quarry).

Clean verified result: year-1 catastrophe and rebuild as before; then year-2 nourishment steady at 0.77, year-2 winter with zero new ruins and 100% condition, textbook granary sawtooth (1,200 → 0 → refilled). Organic civilization: naive youth, catastrophe, learning, stable maturity — now on trustworthy data.

Open (Phase 3 economy, not a bug): accessible stone ran dry during the rebuild (construction + quarry-return overflow at full nodes + selfish hoards), so housing plateaus below population — a stable, fed town with a permanently under-roofed minority. Wants deeper quarrying/metal and real stockpiles.


Run 12 — Knowledge, mortality, and the North's soul (SPEC §5)

Built: three skill domains (forage/woodcraft/masonry) that pay off (skilled hands harvest more, masons waste less) and grow by practice; proximity diffusion soul-calibrated by absorption × teaching-drive; mortality — the piece that makes knowledge generational: elders die at ~150230 days, everything unshared dies with them, and a youth comes of age knowing almost nothing. The North's whole identity is two numbers (absorption 1.6, teaching 0.9) vs the South's (1.0, 0.3); the East teaches no one, ever.

First finding (instrumentation): headless soaks were being contaminated by orphaned Godot child processes writing to later runs' CSVs — two files disagreeing on the final day was the tell. Fixed with a StopAfterDays clean self-terminate (no external kill, no race).

Second finding (design): passive-only diffusion barely helped the North — youths spawned at the empty town centroid and worked solo, so a death reset a villager to roughly what they'd have self-taught anyway. The missing mechanic was the North's actual culture: the young must seek out elders on purpose. Added deliberate teacher-seeking for high-absorption souls, and youths now cradle beside the town's wisest elder instead of in a void.

Result: the North leads at every checkpoint (day 150: 0.586 vs 0.477; final 0.972 vs 0.951) and both towns reach far higher knowledge than before. But the gap is modest because a stable town approaches the knowledge ceiling regardless of teaching speed — everyone eventually maxes out; the North just arrives sooner. Teaching culture is a recovery advantage, not an equilibrium one.

Next experiment (the real test): a shock that kills the experts, then measure which town rebuilds its knowledge — the North's thesis stated as a falsifiable claim.


Run 13 — The full world: four towns, one engine

Built: world scaled 3× (768×768), all four towns at compass points each running its own soul culture (North generational, South nature, West materialist — new profile — East selfish); two new resources (Ore, Herb — herbs are medicine, mended health; a reason to forage and a trade good); a spatial hash grid so proximity queries stop being O(n²) (3.6× headless speedup at ~80 NPCs, live view viable); and the first social verb beyond work: pair-bonds (settled adults who keep warm company pair up; warmth now accrues from proximity, not only gifts; bonds break at death). Reproduction deferred (bonds-first, per plan).

Result — the four-town thesis, fully realized on one engine:

Town Nourish Knowledge Debris Character
North 0.84 1.27 0.000 wise, whole — teaching wins
South 0.85 1.18 0.000 content, clean, lives with the land
West 0.80 1.24 0.227 prosperous but soul-marked
East 0.60 1.08 0.223 scheming and hungriest — selfishness starves itself

The two positive cultures stay soul-clean; the two negative carry real debris. The North leads knowledge (its purpose); the East trails it and is hungriest (teaches no one, shares nothing). 99 bonds, 80 deaths — a living, pairing, generational population in four distinct societies, all from behavior weights. Nothing is scripted; the cultures are emergent.

Next: reproduction on top of bonds (Phase B); crisis-recovery experiment (North's resilience); save/load; then the player pivot.


Soul layer: done and validated (runs 56). The expansion phase begins:

  1. Towns as spatial structure: multiple settlements with their own centroids, buildings, and granaries; NPCs affiliated with a home town but free to roam — the blend space between towns becomes real.
  2. The North (Generational) town: second settlement; its signature knowledge/teaching system (SPEC §5) is the largest unbuilt spec piece.
  3. Inter-town contact and trade routes (Phase 3); object-level soul residue traveling with traded goods.
  4. Save/load round-trip (SPEC §9 amendment) — still outstanding.