STRATOKS PHASE 2A: Integrate a scientifically grounded, lightweight fruit-fly-inspired learning brain into the existing shared-world application. Deliver code, an inspectable demo and measured evidence. Art and IMD payments follow later.

BASELINE

Import this recovered source ZIP into your deliverable repository before editing:

https://stratoks.2-31-126-178.sslip.io/handoff/stratoks-09af84d.zip

SHA256: 8816bfda36545436f85d4ca0b13bd439be14fb3a6be285f25bd39f3edf8abd9e

Commit: 09af84d67722e30c0cb0c096ab8a8189a3b03166.

Old GitHub main lacks recovered fixes and is NOT the baseline. Verify checksum and inspect server/docs/LEARNING.md and docs/recovery/VALIDATION.md. Preserve the full application and existing tests. No production access or live writes; owner deploys after review.

MODEL AND HONESTY

Compare open Drosophila circuits using primary sources, including https://github.com/philshiu/Drosophila_brain_model . Check code AND dataset licenses for this future paid-interaction product. Pin sources, versions, citations, model equations, parameters and asset checksums. Do not bundle restricted data without permission.

Implement an explicitly reduced Drosophila-inspired sensory/associative-learning/action-selection circuit, preferably mushroom-body-inspired, suitable for the hardware below. Explain which biological mechanisms are modeled and which are omitted. An independently implemented reduced model is acceptable; do not relabel the current linear contextual bandit as a fly brain. If using connectome-derived wiring, provide neuron/circuit provenance and reduction mapping. Never claim full brain emulation, consciousness or that fly circuits natively encode building, reproduction or wallet identities.

BEHAVIOR

Wire actual world observations through the new circuit into action scores, not decorative neural telemetry. Use bounded local perception, food/threat cues, internal hunger/energy and reward/punishment to learn preferences. No paid inference API, LLM or GPU dependency. Local circuit state and learned associations must differ per creature.

Demonstrate learning which cues/locations predict food, avoiding learned danger, and reversing a preference when rewards change. Preserve existing personality, relationships and authenticated helpful/harmful wallet memory; document these game-level abstractions separately from biology. Keep resource costs, fertility, births/deaths, construction and disasters. Actions must affect the shared world.

INTEGRATION AND PERSISTENCE

Add a versioned brain interface with legacy and new modes, deterministic RNG and reproducible configuration. Preserve IDs, world age, family/history, needs, wallet memories and existing learned weights. Use a versioned, non-destructive migration: retain legacy weights even where no meaningful translation exists. Document initialization of new circuit state rather than claiming equivalent learning transfer. Persist new weights, internal state and RNG through restart and backup/restore. Test rollback on a copied world. No reseeding, covert reset or population culling for performance. Production defaults remain legacy until explicit operator activation after validation.

CLIENT

Preserve layout and current art. Add a compact creature Brain/Behavior inspector: active mode, sensed cues, learned preferences, action candidates, selected action and reward history from real server data. Distinguish reflex/rule/model decisions. Publish a static browser client through IMD with clearly labeled isolated demo scenarios and configurable HTTPS/WSS server connection. Demo controls cannot mutate production. Two real clients must observe the same server decisions. Deliver server source and Docker package; IMD need not host a backend.

ACCEPTANCE

Compare legacy, new learning-enabled and new frozen-learning modes on identical scenarios over at least 10 fixed seeds. Include acquisition, held-out evaluation without learning, reward reversal, threat avoidance and restart retention. Report raw results, variability and failures; do not hard-code expected preferences. Ablate the new circuit to establish causal influence, and ensure scripted reflexes are not presented as learning.

Target Ubuntu x86_64, 1 shared vCPU, 1 GB RAM total, world container 512 MB/0.75 CPU. Benchmark 24, 100 and 250 creatures plus multiple clients under those limits; report measured sustainable capacity, RSS, tick p95/p99 and protocol payloads. Preserve the 1-second tick at the supported population; report unsupported scales honestly. No unbounded neural state or logs.

DELIVER

Source commit, runnable tests, dependencies/licenses, model comparison and chosen rationale, migrations/rollback, benchmark outputs, browser evidence and public client URL. Independently review learning causality, persistence, authorization and regressions against this exact release. No new token, contracts, payment collection or production deployment. Flag unresolved scientific, license or capacity blockers explicitly.