R12 ETTR Isolated Learnability Preregistration

Protocol: R12-ETTR-IL-v1
Status: preregistered design only; no dataset materialization, training, or scored access authorized by this document
Question: can completed ETTR acquire a synthetic, source-deleted learned capability while the step-300k Shohin base remains frozen?
Absolute prohibition: no continuation pretraining, flagship shard access, flagship optimizer access, or write into any flagship checkpoint/output path

1. Claim boundary

This is the first isolated learned-capability experiment for the completed Endogenous Typed Theory Reactor (ETTR). It is not continuation pretraining and does not test natural-language general reasoning. It tests whether one shared ETTR can learn to:

1. compile a raw WORLD into its categorical packet;
2. lose all access to the raw world, residuals, and KV state;
3. execute a separately disclosed COMMAND;
4. lose all access to the command package;
5. answer a separately disclosed QUERY; and
6. transfer across unseen rules, longer compositions, unseen renderers, combinations of those shifts, and an entirely withheld ontology/task family.

A pass establishes bounded synthetic systematic reasoning by ETTR. It does not establish unrestricted intelligence, natural-language transfer, or a justification to claim that Shohin already reasons generally.

2. Frozen system identity

The base is loaded read-only, put in evaluation mode, and frozen. All base gradients and parameter deltas must be exactly zero. The 67,697,771 ETTR parameters are initialized independently for each declared model seed; no ETTR weights or optimizer tensors from another experiment may be loaded.

3. Absolute isolation contract

The experiment launcher must fail before CUDA initialization unless all of the following are true:

• the only model input is the immutable protected checkpoint above;
• the only data inputs are the frozen synthetic manifests defined below;
• no path containing a flagship shard, flagship_out, a continuation manifest, or a prior optimizer state is mounted;
• no optimizer file from step-300k or any continuation run is opened;
• outputs are confined to a new train/ettr_isolated_learnability_v1/<fold>/<arm>/<seed>/ root;
• the protected checkpoint and every source file remain hash-identical;
• filesystem-open accounting reports zero reads from flagship data/shard roots and zero writes outside the isolated output root; and
• no code path calls a pretraining data loader or continuation trainer.

Any violation invalidates the entire run. It is not repaired by deleting the offending artifact afterward.

4. Frozen semantic board

4.1 Ontologies

The board contains exactly three assessor-side task families:

1. typed Horn closure;
2. typed term rewriting; and
3. guarded resource processes.

Candidate bytes contain no family name, ontology label, theory index, split name, renderer index, oracle output, or expected answer.

The exact fit and score-only theory pools are:

No score-only theory, behavior signature, or exact rule combination may appear in fitting.

4.2 Leave-one-ontology-out folds

The withheld ontology contributes no fitting row, auxiliary label, language model target, renderer-consistency pair, or optimizer update in that fold.

4.3 One semantic rectangle

The atomic unit is a 2x2 WORLD x COMMAND rectangle:

• two worlds share one theory but differ in initial state;
• two commands share one command grammar but differ in the operation or ordered operation sequence;
• all four terminal packets are distinct;
• both WORLD edges and both COMMAND edges change at least one answer;
• each packet receives two semantic queries;
• each semantic query has two meaning-preserving paraphrases; and
• the four query rows for one packet have an identical terminal packet.

Thus one rectangle contains four packets and sixteen query rows. Rectangles that do not satisfy every contrast are rejected before splitting.

Composition depth is the number of dependent command operations. Fit depths are exactly 1, 2, and 3. Held-out depths are exactly 4, 5, and 6.

4.4 Renderers and structural presentations

Renderer IDs 0 and 1 are fit-only lexical encodings. Renderer 2 is development-only. Renderer 3 is development and sealed-confirmation score-only. They are respectively canonical JSON, prefix S-expression, record-delimited infix, and reverse/postfix encodings of the same anonymous typed syntax tree. Complete renderer IDs are disjoint; no scored renderer template or literal separator occurs in fitting.

Fitting uses only base, alpha_reorder, and alias_split presentations. The all-axes score cell additionally uses score-only relation_reification, type_twin, execution_semantics_twin, and ambiguity_deleted_twin presentations. Invariance variants retain the same answer; semantic and ambiguity twins use their independently computed changed answer or ABSTAIN.

4.5 Exact split sizes

For each fit ontology, training contains 1,152 rectangles: 384 at each depth 1, 2, and 3. Every rectangle contributes all sixteen rows.

Development and confirmation each contain eight strata for every ontology:

seen_id, rule, composition, renderer, rule_composition, rule_renderer, composition_renderer, and all_axes.

Each ontology/stratum cell contains exactly 96 rectangles, 384 packets, and 1,536 query rows. Therefore each scored split contains:

Development and confirmation use independent seed domains and share no semantic world, command sequence, opaque symbol, raw row, or 13-gram.

5. Deterministic generation and hash custody

The master data commitment is:

3b796eef284f523a125a18f5c94ae01b1d8305723751c9710086d670d35867aa

It is SHA-256 of the ASCII string:

R12_ETTR_ISOLATED_LEARNABILITY_V1|2026-07-26|source-deleted

The canonical split-specification SHA-256 is:

d98a6895ec11a52bce5625b8c7aa85c7be755d52d4d5082ab7d3a34b2dedff10

The fold commitments are:

For every domain, candidate tuples are ordered by:

SHA256(master_commitment_bytes || "|" || fold || "|" || split ||
       "|" || ontology || "|" || stratum || "|" || canonical_tuple_bytes)

The lowest admissible unique tuples are selected until the exact cell count is reached. Canonical JSON is ASCII, sorted-key, compact-separator JSON followed by one newline. File SHA-256 is over literal bytes. Dataset SHA-256 is over the sorted list of {path, bytes, sha256, row_count} records in that same encoding.

Before any model seed is instantiated, a CPU-only freeze process must produce immutable train, development, and encrypted confirmation files plus one root-signed manifest containing their literal SHA-256 values. The generator source hash, tokenizer hash, split-spec hash, row counts, per-cell counts, and all overlap audits are part of that manifest. This document deliberately does not invent hashes for files that do not yet exist: absence of the signed literal-file manifest keeps training unauthorized. The manifest may report hashes; it may not change any split rule above.

Three already frozen boards are score-only and must never enter training:

These boards are opened only after all arm checkpoints for all seeds and folds are immutable.

6. Model seeds and optimization budget

The model-seed root is 4d2666338e6a8d080f31058b73e7ccb6be7b1b6686db207dba0b0057630b893e. The five exact signed-63-bit seeds are:

827771697280926998, 9160563446168054265, 5619173084519213573, 2431337583064323711, and 8750822315343322697.

Every arm receives exactly:

• 6,000 optimizer updates;
• eight rectangles per global update;
• two rectangles per microbatch and four accumulation microsteps;
• the same SHA-ordered rectangle schedule, token rows, precision, and loss evaluations;
• BF16 forward/backward with FP32 loss accumulation;
• Muon for eligible matrices and AdamW for remaining architecture tensors;
• train_base=False;
• architecture Muon LR 0.005, architecture AdamW LR 0.001;
• Adam betas (0.9, 0.95), weight decay 0.0;
• 500 linear warmup updates, linear decay over the final 20% to 0.1x LR;
• gradient clip 1.0; and
• the frozen ETTR objective weights: token LM 1.0, packet 1.0, WORLD intervention 1.0, COMMAND intervention 1.0, WORLD query binding 1.0, COMMAND query binding 1.0, transaction 1.0, equivariance 0.25, commit/halt 0.5, sparsity 0.01, anti-bypass 0.1.

There is no early stopping, adaptive curriculum, retry, best-checkpoint selection, or threshold tuning. Updates 1,000 and 3,000 are diagnostic only. Only update 6,000 is claim-bearing.

7. Matched arms and ablations

All arms use the same five seeds in all three folds.

1. ETTR treatment: completed architecture unchanged.
2. State-reset ablation: identical parameters and computation, but the recurrent typed state is reset to the sealed initial state before every command step.
3. Binding-deranged control: identical architecture and schedule; WORLD and COMMAND supervision is deterministically reassociated within the same ontology/depth/renderer/answer stratum.
4. Query-only control: identical architecture and schedule; the terminal packet supplied to the reader is the canonical empty packet. Compiler and reactor still execute, preserving parameter/update/FLOP budget.
5. Dense-state control: compiler and reader are unchanged; typed sparse message/update layers are replaced by a dense recurrent block. Before source freeze it must be width-adjusted to match active trainable parameters exactly and measured forward-plus-backward FLOPs within 1%.

Arms 1-4 have exactly 67,697,771 trainable parameters. Arm 5 is inadmissible unless its signed preflight receipt proves the same count and the FLOP bound. No dormant padding parameters count toward matching.

Two noncompetitive ceilings are also reported:

• gold initial packet with learned reactor and reader; and
• gold terminal packet with learned reader.

They receive no optimizer advantage and cannot satisfy the treatment claim.

8. Anti-leakage and custody gates

Every gate is mandatory:

1. zero exact raw-row, semantic-world, theory-hash, command, opaque-name, graph-isomorphism, token-sequence, and normalized 13-gram overlap across train/development/confirmation;
2. zero score-only theory or renderer material in training;
3. exact 50/50 target balance per ontology, stratum, query semantic, and paraphrase, with ABSTAIN scored separately;
4. every wrong-WORLD, wrong-COMMAND, and shuffled-state donor changes the assessor target;
5. no candidate-visible family/split/renderer/theory/variant identifier;
6. a metadata-only classifier using byte length, mask length, row order, and package size remains at or below 52% binary accuracy and chance macro accuracy for multiclass targets on development;
7. no target, oracle product, answer suffix, evaluator output, or donor index exists in a candidate stage mount;
8. physical WORLD -> delete WORLD package and residual/KV state -> COMMAND -> delete COMMAND package -> QUERY -> delete QUERY package -> assessor execution;
9. only the seven allowlisted categorical state tensors and bound metadata cross stage boundaries;
10. post-seal source poisoning, source replacement, and source absence produce bit-identical terminal packets and answers;
11. the external verifier validates the exact WORLD, COMMAND, and QUERY launch receipts and their parent/output hash chain; and
12. development is opened once only after immutable endpoints exist; confirmation is opened once only after every development gate passes.

A custody or leakage failure invalidates evidence. It does not count as an architecture failure until the custody defect is repaired and the entire run is repeated from untouched initial seeds.

9. Metrics and promotion thresholds

Primary accuracy is exact joint correctness: initial packet, complete transaction sequence, terminal packet, disposition, and query answer must all be correct. Packet fields, transactions, halt/commit, and answer are also reported separately.

The treatment must satisfy every condition for every seed:

• fit joint accuracy at least 99%;
• seen_id joint accuracy at least 95% in each fit ontology;
• single-axis rule, composition, and renderer joint accuracy at least 85% per ontology;
• every two-axis joint cell at least 80% per ontology;
• all_axes at least 75% per ontology;
• completely withheld-ontology macro joint accuracy at least 80%, and its all_axes cell at least 70%;
• exact invariant-variant agreement at least 99%;
• noncommuting/order-twin and semantic-twin separation at least 95%;
• ambiguity/contradiction ABSTAIN or REJECT correctness at least 95%;
• WORLD and COMMAND counterfactual answers at least 90% exact;
• post-seal deletion/poisoning invariance exactly 100%;
• treatment exceeds every qualified matched control by at least ten percentage points overall and five points in every ontology/stratum cell; and
• all five seeds have the same positive treatment-control direction.

The dense control is qualified for comparison only if it reaches 99% fit and 90% seen_id. Failure of that control does not waive treatment thresholds. The gold-terminal reader ceiling must be at least 99.5%, and the gold-initial reactor/reader ceiling at least 99%.

After development passes, the unchanged gates are applied once to sealed confirmation and to the three pre-existing score-only boards. No rescore, prompt change, decoder change, or checkpoint choice is allowed.

10. Architecture-rejection and localization criteria

The completed ETTR architecture is rejected for this learned-capability target if any of the following occurs with valid custody:

1. fewer than four of five treatment seeds reach 99% fit by update 6,000;
2. gold-terminal reader accuracy is below 99.5%: reject or redesign the query reader;
3. gold-initial reactor/reader accuracy is below 99%: reject or redesign the reactor/transaction interface;
4. gold ceilings pass but learned initial-packet exactness is below 95%: reject or redesign the WORLD compiler;
5. fit passes but treatment fails any single-axis threshold in three or more seeds: reject the claimed systematic inductive bias;
6. treatment does not beat both state-reset and binding-deranged controls by the required margins: reject recurrence/binding as the causal mechanism;
7. dense-state control matches or exceeds treatment in every transfer cell within two points: reject the claim that ETTR's typed sparse state provides a material generalization advantage;
8. withheld-ontology macro accuracy is below 70% or its all-axes accuracy is below 60% in the five-seed mean: reject cross-family generality;
9. packet/query intervention effects do not match independently computed counterfactuals: reject architecture-native causal use; or
10. confirmation reverses any required treatment-control direction: reject the result.

After an architecture-rejection outcome, more updates, larger synthetic data, threshold relaxation, or selective seed removal are forbidden under R12-ETTR-IL-v1. A successor must name the failed component, change the architecture or objective before accessing a new sealed board, and receive a new protocol identifier.

11. Decision rule

Only a complete pass yields:

ettr_isolated_synthetic_learnability_and_systematic_transfer_confirmed

Any scientific failure yields:

ettr_isolated_learnability_rejected_at_<localized_component>

Any custody, leakage, hash, budget, or equalization failure yields:

ettr_isolated_learnability_invalid_no_capability_claim

This preregistration authorizes no execution by itself.