Close the dynamic embodied Tiny Scientist loop

[dustinoconnor]

What changed

• adds compact L1 causal records with formal variable-binding repair and strict semantic verification
• adds predictive-GNW experiment selection, bounded and committed Unity guidance, observation isolation, and contamination rejection
• adds dynamic Gemma hypothesis admission into a variable-sized causal pool with a strict discovery/held-out verification cutoff
• runs local Gemma asynchronously so Unity control continues during proposal generation
• adds counterbalanced batch/reset tooling, Unity red/blue feature telemetry, focused tests, frozen protocols, and qualified result reports

Why

The original Tiny Scientist could formulate causal records offline, while the Unity PGNW planner could only test a hand-written five-hypothesis pool. This change joins those components: embodied evidence can now cause Gemma to formulate an L1 candidate, the formal layer admits it without behavioral rule authority, PGNW selects subsequent tests, and only later uncontaminated evidence may verify it.

Main result

In corrected seed 154, the live agent froze four mixed discovery observations, generated L1 c red k blue e + t 10.0 q 0.5, admitted it at prior 0.20, and crossed posterior 0.95 after six strictly later clean updates. Thirty-three held-out updates ended at posterior 0.999999999584 with zero discovery-evidence reuse, critical-hunger exposure, respawns, or survival failures. The claim remains bounded to the existing red/blue/probe vocabulary and one successful corrected live replication.

The committed-versus-passive manipulation also reached the 0.95 red-cause threshold in 3/3 committed runs versus 1/3 passive runs across seeds 150-152. This is a small manipulation check, not a population-level significance claim.

Root causes addressed

• compact DSL failures mixed variable-role binding and syntax corruption
• weak PGNW weights rarely changed Unity behavior
• post-pickup contamination prevented usable causal evidence
• dynamic proposal retries repeated after terminal verifier rejection
• sub-measurement signed drift changed token-level effect selection

The implementation retains strict verification, safety vetoes, MPC regret bounds, terminal rejection, canonical feature ordering, and symmetric fixed-precision evidence serialization.

Validation

• python3 -m unittest discover -q — 226 tests pass
• Python compilation passes for dynamic PGNW modules
• git diff --check passes for the published scope
• raw 59–69 MB Unity JSONL recordings and local model adapters are intentionally excluded; compact derived metrics and full qualified reports are included