R12 ETTR-IL-v2 Phase-1 Architecture Handoff

Protocol: R12-ETTR-IL-v2 Phase: 1, architecture and no-update systems qualification Decision: r12_ettr_il_v2_phase1_architecture_frozen_phase2_requires_explicit_user_authorization

1. What is ready

Shohin now has one complete, trainable architecture for testing architecture-native, source-deleted systematic reasoning:

The architecture maps a raw WORLD into a hard categorical packet, applies a separately disclosed COMMAND through a 64-position recurrent transaction reactor, then answers a separately disclosed QUERY from the sealed terminal packet. Raw WORLD tokens, raw COMMAND tokens, base residuals, and KV state do not cross the stage interfaces.

The fixed neural state geometry is:

• 64 packet slots;
• 8 slot types;
• 16 relation roles;
• 256 categorical values;
• 256 relation edges;
• 64 recurrent transaction positions; and
• exact model-visible transport widths of 192 WORLD, 96 COMMAND, and 48 QUERY tokens.

2. What was implemented

Phase 1 includes executable, fail-closed implementations for:

1. typed Horn closure, typed term rewriting, and guarded resource-process semantic families;
2. independent primary and replay oracles through dependent depths 1-6;
3. strict surface ASTs and exact parsers;
4. six presentation families: base, alpha_reorder, alias_split, relation_reification, type_twin, and execution_semantics_twin;
5. lossless token-native transport under the fixed 192/96/48 widths;
6. exact projection into native ETTR packets and 64-position transaction traces, with independent replay to the terminal packet;
7. deterministic semantic candidate scanning and fail-closed cardinality receipts;
8. exact split, fold, quota, invariant-pair, schedule, source, and materialized batch validators;
9. deterministic binding derangement and leakage/custody controls;
10. five matched causal arms: treatment, state reset, binding derangement, query-only, and a parameter/optimizer-family/static-operation-matched dense controller;
11. exact resume, RNG, parameter, token, objective, and zero-update readiness validation;
12. physical WORLD/COMMAND/QUERY source deletion in separate processes;
13. a locked 75-run, update-6000 evaluator with 292 preregistered statistical endpoints and a 100,000-replicate semantic-core bootstrap; and
14. a no-replace Phase-1 source, test, evidence, tokenizer, parameter, and protected-checkpoint freeze receipt.

The dense control is a real alternative recurrent mechanism, not a label in a receipt. It replaces exactly 27,302,912 reactor parameters while retaining the common transaction shell, preserves the exact Muon/AdamW ownership totals, and executes fixed orthogonal replay work to close the protocol's static loss-connected scalar-product budget.

3. Frozen scientific experiment

Each training fold uses two ontologies and withholds the third. The intended claim-bearing experiment contains five arms, three folds, and five fixed seeds, or 75 fits. Each fit has:

• 6,000 optimizer updates;
• 24,000 invariant-pair exposures;
• 48,000 semantic-rectangle exposures;
• 192,000 causal-rectangle exposures;
• 768,000 query-row exposures;
• 405,504,000 charged encoded tokens; and
• 67,697,771 trainable ETTR parameters with the protected base frozen.

Only update 6,000 is claim-bearing. There is no early stopping, best-checkpoint selection, seed removal, retry, decoder selection, or threshold tuning.

The evaluator checks transfer across unseen rules, composition depths, renderers, combinations of all three shifts, and a completely withheld ontology. A positive result additionally requires causal superiority to all four controls, independent source-deleted execution, and one separately authorized confirmation opening.

4. Frozen evidence

The protected checkpoint remains untouched:

step: 300000
sha256: 211d6b2cddf0c2cf8b12cb0b2d73f9c4440d85f6f531018080c8afd35b2f66a6

The exact tokenizer is:

path: artifacts/shohin-tok-32k.json
bytes: 2309567
sha256: 87532df5c121753de3b29194e1f9e3de47986d3f5359548fdf93606773a233d4

Deterministic Phase-1 evidence:

The end-to-end canary materializes 16 rows and four causal rectangles for each ontology, proves exact 192/96/48 transport, and emits a source-free receiving batch. It is a systems test, not learned-capability evidence.

The canonical Phase-1 freeze receipt is:

path: artifacts/r12/ettr_il_v2_phase1_architecture_freeze.json
sha256: 74eed0408d0328105b4433eadcf818a3ceca07e19edd50ac53f2fb50cb2fbee8

That receipt binds:

source files: 37
source inventory sha256: bfcf7be8e95a510b4856b67afa36850e303f129e5843cf24ac3a6ec4df821dbd
test files: 22
test inventory sha256: b505638e7645ad6c59ebd978c6ddecf4a080de4433fcde1e6720ded34b2b939c
evidence inventory sha256: d2fefbbfac990ed2f2a4869c9497b744ded859cfb7005ee5fcfcd82d1d47cf3e

Final local verification completed with 350/350 tests passing: 337 non-arm checks and 13 arm checks. The arm suite includes a real production-geometry dense-control forward. Ruff, Python byte compilation, specification regeneration, canonical-artifact comparison, hash checks, and git diff --check all passed. Separately, the locked evaluator completed a full synthetic production-path rehearsal of all 100,000 bootstrap replicates and 292 endpoints and returned OPEN_CONFIRMATION; it did not open real scores.

5. Explicit non-claims

Phase 1 performed zero weight updates. It did not:

• continue Shohin pretraining;
• fit ETTR or any control arm;
• modify or rewrite the protected checkpoint;
• materialize the literal 20,736-rectangle production population;
• open development or confirmation scores;
• establish that the newly initialized parameters already reason; or
• establish native general reasoning.

The implementation demonstrates that the architecture, transport, materialization, controls, custody, budgets, and evaluator are mechanically defined and testable. Capability remains an empirical Phase-2 question.

6. Phase-2 entry order

Phase 2 must preserve this order:

1. Generate the literal candidate population and recompute every ontology, presentation, and execution-semantics-twin outcome.
2. Certify cardinality and select the exact quotas without replacement.
3. Materialize all selected batches, freeze them, reload them independently, and replay every transaction trace.
4. Run the complete raw-row, semantic, token-sequence, normalized 13-gram, graph-isomorphism, opaque-name, path, length, mask, ordering, and metadata leakage audits.
5. Validate exact split disjointness, packet sufficiency, invariant pairs, schedules, derangements, parameters, optimizer ownership, token counts, and static operation budgets.
6. Execute one zero-update readiness batch for every arm and verify exact resume/RNG receipts.
7. Request explicit user authorization.
8. Fit only after authorization.

Any cardinality, transport, leakage, source-deletion, parameter, budget, or readiness failure blocks fitting. It is invalid evidence rather than a negative reasoning result.

7. Remaining scientific risks

• The production population has not been enumerated. Some theory/depth/stratum quota may fail and require a protocol revision before fitting.
• Execution-semantics twins require fresh oracle execution after transformation; mechanical surface invertibility is not sufficient.
• The dense matched control is deliberately expensive because it closes the native reactor's static operation budget. Its H100 runtime must be measured before the 75-fit campaign.
• Same-account custody proves mechanics, not independent scientific confirmation. A public positive claim still needs independently owned authorization and immutable opening records.
• A passing synthetic protocol would establish bounded systematic transfer, not unrestricted natural-language reasoning. A failing protocol must be localized using the preregistered component precedence rather than repaired post hoc.

8. Authority

The normative protocol is R12_ETTR_ISOLATED_LEARNABILITY_PREREG_V2.md plus its four exact component specifications. This handoff summarizes those frozen contracts. Where wording differs, the hash-bound preregistration and executable validators control.