R12 Registry-Constrained Opcode Projection Preregistration
Status: implementation complete; fixed 5,000-update diagnostic pending.
Date: 2026-08-02 EDT
Claim boundary: parameter-free inference diagnostic. This experiment can localize a whole-program decoding defect, but it cannot establish native reasoning unless both unchanged source-deleted causal axes improve and the result later survives trained replication.
Causal hypothesis
Contract-v8's 5,000-update syntax-graph compiler improves every local schedule field while exact terminal packets regress to zero. The flat sticky selector then collapses to the dominant opcode skeleton. These results leave a narrower hypothesis: the per-step compiler contains useful conditional opcode evidence, but independent argmax decisions form invalid or incoherent complete programs.
The treatment projects those existing per-step probabilities onto one complete train-registry opcode path. It scores each valid path by mean per-step log probability and makes one global hard selection. Registry order, empirical class counts, targets, and candidate answer quality do not enter the score. This removes both stepwise hybridization and the learned majority-class prior.
Frozen inputs
- Base schedule: completed contract-v8 5,000-update V100 arm
725573. - Base report SHA-256:
7922f9b26a41be8125983149d3d896ba812deb4a65765d5b7a02f4bfad8c8623. - Train-only registry SHA-256:
03fc92829bc4a1c9f9e8381953ac506e04afeef746871a60ebfca1e482cbafcc. - Registry payload SHA-256:
d58185b4a5c7b28e54cd9497215dd8d5f0e52f7339a968f10facbc6669497b4b. - Architecture/data seed:
31/11. - Development batches: 32, exactly 512 rows.
- Learned parameter delta: zero.
The base compiler checkpoint and registry remain hash-bound and read-only. The projection may consume only the base compiler's opcode probabilities and the frozen train-only opcode sequences. It may not consume QUERY, answers, terminal targets, oracle programs, development-derived templates, or candidate scores.
Fixed projection
For a base opcode distribution p[t, opcode] and candidate program z:
score(z) = mean_t log p[t, z[t]]
z* = argmax_z score(z)
Programs longer than the requested rollout are ineligible. Each candidate's terminal opcode is replayed after its end, which is state-neutral because the exact algebra has already committed, halted, or rejected. Source, target, relation, type, and value predictions are unchanged from the sealed base compiler. The selected opcode skeleton is replayed only by the existing exact transaction algebra.
Required report
- projected opcode and joint schedule accuracy;
- oracle-initial and autonomous-initial exact terminal packets;
- fully autonomous factual top-1;
- strict, margin-1, and intervention-DID WORLD and COMMAND gates;
- registry, schedule-run, source, and output hashes; and
- an explicit comparison with the unprojected 5,000-update endpoint.
Decision rule
The projection justifies a trained structured successor only if it restores a nonzero exact terminal packet rate without factual regression, or crosses at least one fully autonomous strict causal axis. It is a native-reasoning candidate only if both strict axes improve, then replicate across seeds and held-out population orderings. A gain visible only with oracle initial state remains an interface diagnostic.
If projection remains zero on exact terminal state and both causal axes, the valid-path decoding hypothesis is closed for the current opcode evidence. The next architecture must change how program evidence is represented, not merely how independent decisions are decoded.