progress.md

Original prompt: make it continuous and learning, save its state, and add chat and a 3D environment interface.

2026-10-20 request: "innvoate advance and improve project further look at new and best research to help". v0.10 implements optional zero-initialized all-source context readout, quarter-turn augmentation, expert-only matched collection, no-context inference trials and a predeclared four-arm/three-seed learning study. Default human readout and prior checkpoints retain compatibility. Research links, competing approaches and limits: docs/v10-learning-design.md. The complete 4-arm/3-seed 12-run study (6,144 updates) is fully verified with exact frozen replay (reports/v10-replay-verification.json). Result: all-source residual readout improved raw unassisted policy success from 15/24 (62.5%, human_plain) to 16/24 (66.7%, context_residual_plain) with lowest collisions (73.2%) and fastest arrival (21.8 steps); rotation augmentation degraded navigation performance (dropping human from 15/24 to 10/24) due to teacher tie-breaking directional bias under partial observability. Lesioning the all-source residual drops success from 16/24 to 10/24 (SPL drops from 0.6667 to 0.4083), proving causal reliance on the multi-source representation. Full 595-test suite passed in 378.52s (reports/v10-tests.txt); browser regression passed cleanly (exit 0); isolated wheel parity verified across all 42 files and README metadata (reports/v10-package-verification.json). Live source state is preserved without unauthorized promotion. Full analysis: docs/v10-build-report.md.

2026-10-19 further upgrade: v0.9 adds opt-in legal-action masking and observation-only episodic mapping/planning, controller attribution trials and frozen multi-condition studies. Default raw behavior and legacy reports remain intact. v3 checkpoints save mode/map and validate plausible post-decision pose; v1/v2 read in raw mode with fresh map memory. Browser has observed-only SVG, effective teacher-action status, extended comparisons and exact map save/load checks. No live checkpoint or trained policy is replaced. Mapping gains must be attributed to engineered assistance; architecture bottlenecks are hypotheses, not measured causes of raw-policy failure.

v0.9 final verification: 527 tests passed (222.24s); desktop/mobile browser and bundled action-loop screenshots inspected; isolated final wheel matches all 42 source/static files and README metadata and exactly replays all 192 variant/scenario episodes. Across 24 declared fixture scenarios, observed-map navigation reached 24/24 goals with SPL0.9642 and zero collisions; legal masking reached18/24 and raw policy0/24. Map-only reached24/24 too. Results and limits: docs/v09-build-report.md. Independent review fixes include saved-pose validation and semantic/action-geometry verification of study receipts. Final validation also repaired an unread-body startup rejection that could hide its response via TCP reset; the failed receipt was retained. Original live checkpoint and previous evidence remain unchanged. No model promotion, commit or publication.

2026-10-19 follow-up: "innovate advance and improve whole project further". Implemented v0.8 Cartesian curriculum coverage, finite action-stratified reservoir replay, explicit work budgets and round receipts, archived exact baseline/source/config/progress/results, verify-navigation [--replay], frozen diagnostics captures, transactional live startup and truthful observer labels. Details and validation receipts: docs/v08-build-report.md and docs/v08-upgrade-plan.md. The smoke candidate remains 0/4 on its frozen tests; no promotion or capability claim is warranted. Current boundaries: finite replay may omit historical worlds, candidate exports do not resume curriculum optimizer state, and environmental substrates/odor/gates remain observers rather than policy/physics inputs. Existing live state is preserved. Future research should target measurable unassisted navigation with matched ablations rather than adding more observer names.

Implemented and verified NEUROGRAFT v0.2 against docs/live-session-plan.md while preserving all v0.1 reports and acquired source graphs:

• neurograft/environment.py: deterministic 9x9 multi-beacon navigation world with shortest-path teacher, per-source sensory observations, and JSON state serialization.
• neurograft/live.py: LiveSession wrapping GraphCircuit + 5-way navigation head, online AdamW (lr=3e-3) supervised route imitation, persistent recurrent states, local command & fact memory chat (local-memory-and-control), and atomic weights_only=True checkpointing with bundle SHA-256, ZIP CRC32/header/padding, and tensor finiteness verification.
• neurograft/service.py & neurograft/server.py: thread-safe background loop, periodic autosave, named checkpoints (data/live/checkpoints/<name>.pt), HTTP JSON API, and static asset serving.
• neurograft/web/: interactive Three.js 3D navigation arena, cross-source route matrix, source activity gauges, loss/accuracy sparkline, checkpoint & goal controls, chat drawer, and Playwright hooks (window.render_game_to_text, window.advanceTime).
• neurograft/cli.py & Run-NEUROGRAFT.ps1: added python -m neurograft live and -Mode Live launcher.
• Verified with pytest and headless Chromium Playwright (verify_ui.mjs and web_game_playwright_client.js) with zero console/page errors.