SRIC: endogenous semantic witness

SRIC Block 23 Real Latent-Reasoner Pilot Status

Target

The staged real-model target is a source-locked multimode GPT-2 Coconut checkpoint evaluated on PrOntoQA-style logical reasoning tasks.

The adapter captures the embeddings actually placed at latent-token positions and supports deterministic evaluation under externally supplied traces. The audit protocol includes replay, zero-latent, truncation, step permutation, norm-matched noise, semantic-twin transplant, and same-answer different-proof transplant.

Source lock

Repository: connordilgren/are-lrms-easily-interpretable

Commit: 32f413d8d55239d9bc54bb6b6ec37b0630891ed4

Checkpoint repository: connordilgren/gpt2-prontoqa-multimode-coconut

Checkpoint filename: checkpoint_37

Checkpoint SHA-256: not yet committed. This is deliberate. The digest must be computed from the locally obtained file before a real final audit.

Current environment

• status: ADAPTER_VALIDATED_REAL_EXECUTION_BLOCKED
• real checkpoint executed: False
• blockers: transformers_missing, repository_checkout_missing, checkpoint_missing, cuda_unavailable
• Python: 3.13.5
• Torch: 2.10.0+cpu
• CUDA available: False
• Transformers available: False

Package-local protocol evidence

• fake fixture tasks: 72
• fake fixture exact replay maximum: 0.000000

These numbers validate adapter contracts and intervention bookkeeping only. They are not empirical findings about Coconut or PrOntoQA.

Required next execution

1. Obtain the exact repository commit in frozen/SOURCE_LOCK.json.
2. Obtain the declared checkpoint and compute its SHA-256.
3. Update the lock before any final audit.
4. Install the source-locked upstream runtime separately from the package verification environment.
5. Run external_preflight.py and require a ready status.
6. Run a small adapter smoke set and verify exact replay.
7. Freeze donor rules, semantic-twin transforms, task split, thresholds, and intervention seeds.
8. Run the real pilot without changing the frozen gates.
9. Preserve a failure or underidentification result as the canonical outcome if the model does not satisfy them.