REAL_LRM_PILOT_STATUS.md
SRIC: endogenous semantic witness
REAL_LRM_PILOT_STATUS.md
1 # SRIC Block 23 Real Latent-Reasoner Pilot Status
2
3 ## Target
4
5 The staged real-model target is a source-locked multimode GPT-2 Coconut checkpoint evaluated on PrOntoQA-style logical reasoning tasks.
6
7 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.
8
9 ## Source lock
10
11 Repository: `connordilgren/are-lrms-easily-interpretable`
12
13 Commit: `32f413d8d55239d9bc54bb6b6ec37b0630891ed4`
14
15 Checkpoint repository: `connordilgren/gpt2-prontoqa-multimode-coconut`
16
17 Checkpoint filename: `checkpoint_37`
18
19 Checkpoint SHA-256: **not yet committed**. This is deliberate. The digest must be computed from the locally obtained file before a real final audit.
20
21 ## Current environment
22
23 - status: `ADAPTER_VALIDATED_REAL_EXECUTION_BLOCKED`
24 - real checkpoint executed: `False`
25 - blockers: `transformers_missing, repository_checkout_missing, checkpoint_missing, cuda_unavailable`
26 - Python: `3.13.5`
27 - Torch: `2.10.0+cpu`
28 - CUDA available: `False`
29 - Transformers available: `False`
30
31 ## Package-local protocol evidence
32
33 - fake fixture tasks: 72
34 - fake fixture exact replay maximum: 0.000000
35
36 These numbers validate adapter contracts and intervention bookkeeping only. They are not empirical findings about Coconut or PrOntoQA.
37
38 ## Required next execution
39
40 1. Obtain the exact repository commit in `frozen/SOURCE_LOCK.json`.
41 2. Obtain the declared checkpoint and compute its SHA-256.
42 3. Update the lock before any final audit.
43 4. Install the source-locked upstream runtime separately from the package verification environment.
44 5. Run `external_preflight.py` and require a ready status.
45 6. Run a small adapter smoke set and verify exact replay.
46 7. Freeze donor rules, semantic-twin transforms, task split, thresholds, and intervention seeds.
47 8. Run the real pilot without changing the frozen gates.
48 9. Preserve a failure or underidentification result as the canonical outcome if the model does not satisfy them.