[Pentexa]

METARGON NEST v0.2 — genuine interactive conversation + teaching

Redesigns NEST around the real agent:

• Conversation is grounded in real telemetry (sanitized snapshot:
  observations, challenge/status, last action + outcomes, verified/
  unverified hypotheses with confidence/contradictions, telemetry, history).
  Primary mode uses a user-configured OpenAI-compatible LLM; actions are
  only requestable via a validated [[ACTION:<intent>]] tag. Offline
  deterministic interpreter is kept and labeled "command mode" fallback.
• Teaching integration: a teacher hint is classified into one of three
  semantics — exploration guidance ("try switch S1", changes candidate
  availability only), an unverified causal claim ("I think S1 opens door
  D1", injected as a status=hypothesis candidate with sup=0), or a
  learning objective. The student (MetargonAcademyAgent.act()) remains
  the sole action authority; external planners never run, and teacher
  claims are never auto-verified (verification happens only through real
  observations in agent.learn()).
• Memory v0.7 configuration enabled in the session: episode-scoped
  causal memory, atom-level transfer, bank reuse, per-episode probe reset.

Controlled learning experiments (honest)

nest/tools/diagnose_learning.py compares fresh / trained-without-
guidance / trained-with-guidance / trained-with-claim / misleading /
wipe under matched seeds/budgets and separated training-evaluation. No
positive transfer is forced; results are reported as measured.

nest/tools/diagnose_regression.py isolates the post-training regression:

• Repro: fresh 1.0 vs trained 0.0 on T1E1 (same-topology, different rule).
• Isolated cause: memory interference — both content clearing and a
  functional memory mute restore fresh-level success; exploration-state /
  counters / RNG / threshold-2 controls do not.
• Not universal: cross-topology pair T2E2 (TP2→TP1) shows no regression.
• First divergence is now computed on actual environment action sequences.
• Diagnostics are not tuned against the evaluation set; the student's
  cross-task interference is reported as the (documented) mechanism.

Tests

• NEST suite: 33/33 passing (conversation acceptance, teaching
  semantics, claim injection, learning validation, regression diagnosis
  with cross-topology, frontend node syntax + DOM smoke, persistence).
• Repo hygiene: OK. CI contains a NEST job that runs everything,
  including frontend validation.

Known limitations

• No claim of positive general transfer is made.
• Teacher claims remain unverified until demonstrated by observations.
• The student's episode-scoped memory is configured with transfer
  threshold 1; regression diagnosis shows threshold 2 does not change the
  T1E1 outcome.
• LLM use is optional; without an API key the app runs in deterministic
  command mode.

Starting commit: a8b0e4b (origin/main before PR #45); PR #45 frontend
fixes are integrated via merge.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.