R12 Atomic Typed-Edit Algebra Preregistration

Status

Preregistered before inspecting the sealed sparse-residual result. That result failed both required state-isolated axes. The atomic architecture is now implemented and locally qualified before its first H100 launch. Fifty-five focused and related tests pass with clean Ruff, byte compilation, Bash syntax, and diff checks. The trainer fails before optimization unless canonical edits exactly reconstruct the first real training batch's target terminal packet.

Diagnosis

Copy-biased residual transport separates preservation from rewriting, but its value, type, active, root, relation, and status gates can still make mutually inconsistent local decisions. A causal intervention is not a bag of changed fields. It is a small typed edit whose opcode, address, payload, and graph effects must agree. Independent gates can reduce aggregate Brier loss while never producing the exact intervention-specific state consumed by the reader.

Architecture

Compile COMMAND plus the initial typed state into one parallel set of atomic edits, then apply those edits with a fixed differentiable state algebra. The neural module predicts:

• one categorical node action per slot: KEEP, ALLOCATE, WRITE, or CLEAR;
• value and type payloads, consumed only by actions that legally use them;
• one categorical relation action per typed edge: KEEP, LINK, or UNLINK;
• one categorical root action: KEEP, CLEAR, or SET(slot);
• one categorical disposition action: KEEP, COMMIT, HALT, or REJECT.

The executor applies all non-conflicting edits atomically. KEEP is exact identity, ALLOCATE writes active/type/value together, WRITE changes only a legal active value, CLEAR removes the node and every incident relation, and relation actions are masked by post-node active endpoints. Root and status are single categorical decisions. This removes serial exposure while preserving the transaction algebra that independent field gates discarded.

Hard inference uses deterministic argmax actions and emits one valid terminal packet. Soft training uses the same algebra with bounded straight-through categoricals. The runtime receives autonomous initial state and COMMAND bytes only. It receives no QUERY, answer, target, oracle program, host solver, candidate selector, or best-of-K evaluator.

Objective

Retain the full-state Brier anchor and complete 2x2 WORLD/COMMAND terminal- delta loss. Add action supervision derived deterministically from the initial and target terminal packets; this is a canonical state difference, not an oracle transaction trace. Ambiguous edits are rejected during materialization or assigned a deterministic minimal form before the immutable train split is sealed. Loss is balanced by action class and intervention axis.

The first gate uses the same architecture seed 31, data seed 11, 1,000 updates, 3e-4 learning rate, clip 1.0, 32 evaluation batches, protected checkpoint, source-deleted reader, and held-out ordering as the sparse- residual arm. This changes the edit representation and executor, not the data population or evaluation.

The primary action-loss weight is 1.0. Before inspecting any sealed result, an equal-budget 4.0 dose arm is admitted because live training telemetry localizes most residual action loss to the 256-way value payload while node, relation, root, and disposition actions are already low. The primary remains the claim-bearing architecture gate. The dose cannot promote on action NLL, state loss, or interface fields; it must independently move both hard WORLD and COMMAND under the unchanged evaluator.

Promotion And Rejection

Promotion requires nonzero strict WORLD and strict COMMAND under the fully autonomous path and corroborating movement under oracle-program/autonomous- state isolation, without material factual collapse. It must then reproduce on a fresh data ordering and architecture seed. Action accuracy, aggregate field accuracy, training loss, and soft margins cannot promote it.

Reject the mechanism if either causal axis remains invariant under the state- isolated gate or if gains depend on one ordering. A negative result would show that the missing object is not merely coherent editing; the next intervention must alter how COMMAND semantics are bound to state identities before edit compilation.

Decision: compile_one_coherent_typed_state_difference_and_apply_it_atomically_before_scaling_any_dense_terminal_decoder.

Result

H100 primary 725510, V100 replication 725511, and H100 action-weight-4 dose 725513 completed all 1,000 updates cleanly. Canonical edits exactly reconstructed the first real target packet before every run began. The H100 primary learns low final-batch action NLL (0.253; value payload 1.076) and reaches 54.49% factual top-1, but strict WORLD/COMMAND and both margin-1 rates remain zero. The oracle-program/autonomous-state cross has DID exactly zero on both axes. Report SHA-256 is bd9e55a2dfabf632209efe7431d567f01d88add14f9ecccc558a81b40392fd9d.

The V100 result independently returns the same isolated zeros; report SHA-256 is b1c078e22eb2095094494a33288202ef6fa1b72404291b94805ce3185b3ef0c4. Increasing action weight to 4 does not repair the gate. It reaches 54.69% factual but lowers value accuracy to 49.43%; every autonomous and isolated WORLD/COMMAND strict, margin-1, and DID metric is zero. Report SHA-256 is 051385f1fb3dbc7ca1111398d2f1322cfb464c4e6cebdc00ca61547d3d69406a.

Coherent edit mechanics are therefore insufficient when the initial packet is lossy. Training targets derived from an autonomous WORLD state include repairs for information the frozen WORLD compiler already discarded; COMMAND cannot identify those repairs. The next bounded diagnostic trains the same atomic editor from oracle initial packets to isolate COMMAND binding. It is not a reasoning claim. A positive result would require a jointly trainable WORLD compiler plus atomic COMMAND editor; a negative result would localize failure inside COMMAND-to-edit binding itself.

Disposition: retain_fixed_atomic_algebra_but_stop_asking_command_to_recover_world_information_that_the_initial_packet_erased.