# R12 State-Bound Operation-Family Arbiter Preregistration **Status:** completed and rejected at the frozen gate; local family line closed **Contract:** `shohin-ettr-parallel-terminal-state-contract-v20` **Promotion gate:** at least 90% independent held-out operation-family exactness **Reasoning gate:** unchanged source-deleted WORLD and COMMAND strict pairs ## Measured reason for the mechanism The corrected full-corpus audit covers 494,480 training and 60,512 development operations. Its development conditional-oracle family accuracies are: | Interface | Oracle accuracy | |---|---:| | Public operation syntax | 80.1692% | | Syntax plus preceding-state topology | 80.4435% | | Syntax plus exact preceding-state values | 94.0756% | Exact state values therefore contain enough information to cross the frozen 90% family gate, but a smoothed additive factor model remains at 73.7110% in all three modes. LINK is perfectly separated; the residual ambiguity is NONE versus WRITE. Exact composite signature lookup is not viable because only 18.7864% of development operations have a train-seen exact syntax-plus-state signature. The audit report file SHA-256 is `819aa0bdb57b3e46fcf1488323d17e18fd3668db277d4404ed685f76a69b75c5`; its payload SHA-256 is `0a5574bf499440fe4a3ca47d80b08d39e9177e76343ecb8f6a3304f0247ea517`. ## Hypothesis The public operation roles identify *what should be compared*, while exact typed-state values determine whether that operation is a semantic NONE or WRITE. Pooled addition can represent each side independently but does not force their binding. A role-conditioned bilinear lookup should expose the missing interaction without supervising answers, terminal states, or programs at inference. ## Architecture `OperationStateBoundFamilyGatedWriteLinkCompiler` preserves v19's compiler, recurrent operation boundary, hard NONE/WRITE/LINK gate, and typed payload rails. It changes only the family arbiter: 1. Re-encode the exact current typed state into one memory vector per slot. 2. Project each public semantic-role anchor into multihead queries. 3. Project current state slots into keys and values. 4. Compute role-to-slot bilinear compatibility and attended state values. 5. Concatenate role, attended state, and their elementwise product. 6. Pool only valid roles and predict exactly one NONE/WRITE/LINK family. The production compiler has 50,594,556 parameters, 1,576,448 more than v19; the complete system is 206,533,450 parameters. The user has removed the old 200M ceiling. This increase is nevertheless tied to one measured causal purpose and is not evidence by itself. ## Information boundary Inference may read only the autonomous current typed state and public COMMAND tokens/residuals. It may not read QUERY, answers, target packets, successor states, transaction programs, assessor traces, candidate scores, or a host semantic solver. Oracle preceding states are allowed only as family-island training labels. The immutable evaluator and fixed state algebra remain unchanged. ## Training isolation The first v20 arm is family-only. WRITE/LINK count, pointer, value, relation, and payload-rail parameters are bypassed, and the runtime gradient guard must prove they receive no gradient. Training remains class-balanced over NONE/WRITE/LINK. Terminal and causal outputs from random payload rails cannot promote the family island. ## Fixed route 1. Complete corrected v15 and its independent seed-13 evaluation first. 2. If v15's family accuracy is at least 90%, do not run v19 or v20; route the measured payload failure. 3. If v15 family fails, run exactly one v19 family-island fit. 4. If v19 reaches at least 90%, preserve its weights and release joint rails; do not run v20. 5. If v19 misses 90%, run exactly one matched v20 family-island fit. 6. If v20 reaches at least 90%, preserve its weights and release joint rails. 7. If v20 misses 90%, reject standalone NONE/WRITE/LINK control and compile typed effects through direct latent expert competition. Do not repeat a wider, longer, or reseeded pooled/bilinear family classifier. Only a later joint system that improves both WORLD and COMMAND strict paired gates on independent source-deleted populations can count as native reasoning. ## Mechanical evidence before GPU admission - Exact-value counterfactuals change family probabilities with fixed syntax. - Family loss sends finite nonzero gradients through role queries, state keys/values, the multiplicative projection, and typed value embeddings. - Family-island end-to-end tests leave all WRITE/LINK payload parameters gradient-free. - Train, evaluator, router, and Slurm launcher agree on contract v20 and its fixed pass/fail route. V20 was authorized only after v19 failed and was run exactly once. This preserved single-writer scientific attribution and the preregistered endpoint. The immutable Newton runtime is `scratchpad/shohin_ettr_state_bound_family_runtime_898591c_r1`, bound to source `898591cf5861cbff2aef2a293a89347c7fea8bcd`. Its 3,658 files pass complete checksum replay with zero links and zero writable entries. The SHA256SUMS SHA-256 is `73b82ec26467de405ccd7d32ff4031cb0fbacd394c2708ebba355d86cec3d484`. ## Measured closure The finite campaign selected v19 and then v20. Both completed all 1,000 updates in bfloat16 and both failed the independent seed-13 gate identically: operation-family exactness `464/1856 = 25%`, 100% predicted `NOOP`, exact terminal packets `0/512`, WORLD strict `0/48`, and COMMAND strict `0/144`. V19 final compiler SHA-256 is `9b741d392c149b1e32bc4040621d603d93dd05c5a25759ae7df580150d445457`; v20 final compiler SHA-256 is `0aba95f47bc4c0263872dd1e512798fae26c18fbee2f5b92d9295a2f71d1834f`. The final planner action is `stop`; v21 was correctly not launched. Postmortem auditing found that the 94.0756% conditional oracle and the neural fit did not share an equivalent input or optimizer distribution. The oracle uses an exact symbolic resolver over public COMMAND structure plus exact preceding-state factors. V20 instead learns from language residuals, pooled public role anchors, and a learned typed-state encoding; equivalence was never measured. More decisively, every one of the 100 logged checkpoints in both fits omits at least one target family, 33% contain exactly one family, and LINK appears in only 33% of logged batches. At update 499 an all-LINK batch is fit essentially perfectly; the final batch is 75% NONE, 25% WRITE, 0% LINK, and the final model emits NOOP everywhere. The exact trace audits are `artifacts/r12/v19_operation_family_stream_drift.json` and `artifacts/r12/v20_operation_family_stream_drift.json`. This does not reopen v20. Standalone family control remains rejected under the frozen rules. It changes the capability-floor launch requirements: the new unified trajectory must first prove symbolic-to-neural interface sufficiency and use deterministic replay-balanced optimizer windows so one semantic regime cannot erase another.