R12 Causal-Delta Terminal-State Preregistration

Status

Implemented and locally qualified on 2026-08-01. Both preregistered doses are complete on H100 and independently replicated on V100. The result is negative.

Diagnosis

The direct terminal-state quotient reaches high aggregate terminal-field and factual accuracy on both H100 and V100, but predicts intervention-invariant states. WORLD and COMMAND strict gates, margin-1 rates, and DID remain zero. The original state objective balances positive and negative classes within each field, but it does not balance the sparse differences between episodes. The optimizer can therefore fit the dominant terminal-state manifold while discarding the small set of coordinates that carry each causal intervention.

Treatment

Keep the 18,520,349-parameter ParallelTerminalStateCompiler, hard state constraints, frozen Shohin backbone, frozen algebraic query stack, and exact source-deleted evaluator unchanged. Add rectangle-level terminal-state delta credit during training only.

Every batch is already an immutable set of complete 2x2 WORLD x COMMAND rectangles. For each WORLD edge and COMMAND edge, the objective computes the predicted terminal-state difference and exact target difference. Loss is assigned only to support-valid coordinates whose target actually changes. Binary fields use squared delta error; value and type use vector Brier delta error. Each present semantic field and intervention axis receives equal weight, independent of raw coordinate sparsity. The original full-state Brier loss remains as an anchor.

This objective sees rectangle membership and terminal packet targets. It does not see QUERY bytes, query targets, answer labels, oracle programs, host execution, candidate scores, or best-of-K selection. Inference remains one deterministic hard terminal state from initial state plus COMMAND.

Matched Gate

Two preregistered 500-update arms use identical architecture seed 31, data seed 11, stream position zero, autonomous initial state, LR 3e-4, clip 1.0, and 32 evaluation batches:

• primary: full-state loss + 4.0 * causal_delta_loss;
• dose control: full-state loss + 1.0 * causal_delta_loss.

The exact protected checkpoint, release, algebraic reader, evaluation ordering, and 200M system cap remain unchanged. Changed-coordinate counts and per-axis delta losses are logged on every metric interval.

Promotion Rule

Training loss, aggregate packet fields, exact packet count, and factual top-1 are diagnostics. The mechanism advances only if a fully autonomous arm moves both strict WORLD and strict COMMAND above the zero matched baseline without a large factual collapse. A margin-only result may justify one bounded extension but is not a win. Promotion requires fresh-ordering and second-seed replication.

Rejection Rule

Reject causal-delta credit as sufficient if both doses remain strict-zero, if only one causal axis moves, if gains require oracle initial state, or if a gain fails fresh population or seed replication. The next architecture in that case must transport a sparse coherent edit object explicitly rather than asking one dense terminal state to represent both invariant background and intervention.

Decision: train_terminal_state_differences_as_first_class_causal_objects_and_require_joint_world_plus_command_transfer.

Result

All four 500-update runs completed cleanly. Weight 1.0 reaches 60.94% factual top-1 on H100 and V100; weight 4.0 reaches 48.24% on H100 and 60.94% on V100. Every run remains strict WORLD 0% and strict COMMAND 0%. More importantly, the oracle-program/autonomous-state cross-check is invariant for every arm: WORLD and COMMAND strict, margin-1, and DID are all exactly zero. The H100 weight-4 fully autonomous stack shows a 5% WORLD margin-1 and DID 0.252, but it vanishes on V100 and under oracle-program isolation while factual accuracy collapses; it is therefore a composition artifact, not state reasoning.

Report SHA-256 values are:

• H100 weight 1: 5effcde6064ad349408e6d9a9b6466b47d002adc851ceacd73a6894f5e677225;
• H100 weight 4: 9ebb1af5437cfdda25329e254847a36ba37c1367206a29524a578951f3ea1bc8;
• V100 weight 1: 99e22ebda447dce748b1017bd11919a819a8adbb3718e4fea01becd6faac171e;
• V100 weight 4: 90755ca4cab9a42dbf51512eda2d1e91a31309f6472ab769c61f1e66fd32f4fd.

This closes loss-only repair. The dense absolute-state architecture can absorb causal-delta gradients by degrading the common state estimate without learning a stable intervention-conditioned edit. The next mechanism must make identity transport and sparse editing distinct architectural operations.

Disposition: reject_loss_only_causal_delta_repair_build_explicit_identity_plus_sparse_edit_transport.