R12 ETTR-IL-v3 Materialization Contract
Contract ID: R12-ETTR-IL-v3-materialization-v1
Status: implemented before architecture-facing v3 materialization
Purpose: preserve the broad initializer population frozen by
R12-ETTR-IL-v3-initializer without weakening the ETTR causal packet
contract or silently restoring the rejected v2 checkerboard filter.
1. Scope
This contract changes no neural module, tensor width, parameter count, checkpoint, optimizer state, or training authorization. The complete Shohin ETTR system remains 192,779,435 parameters. It changes only the validation, objective, and receipt semantics used when broad v3 causal rectangles are materialized.
The candidate protocol was frozen before this receiving-contract mismatch was found. This document is therefore an explicit materialization-contract repair, not a claim that the original candidate freeze already contained the repair.
2. Defect
The v3 protocol deliberately admits broad compiler, transaction, and composition episodes. WORLD or COMMAND interventions in those episodes must change the terminal packet, but they need not change the answer to every query. The inherited v2 receiver instead required every edge of every 2-by-2 WORLD-by-COMMAND rectangle to change the query label.
That requirement would reject broad v3 rows after selection and recreate the same sparse XOR/checkerboard population that v3 was introduced to avoid. It would also train an impossible contrastive target on a valid no-effect query edge.
3. Factorial geometry
Every architecture-facing causal rectangle still must:
- partition its batch exactly into
(W0,C0),(W0,C1),(W1,C0),(W1,C1); - keep the query prefix and autonomous read position matched;
- use distinct WORLD and COMMAND source renderings on their declared axes;
- keep packet support geometry fixed across the rectangle;
- change the initial packet across the WORLD axis;
- change the terminal packet on every WORLD and COMMAND edge; and
- carry independently executed targets for all four corners.
An equal answer label is permitted only when those packet-level causal gates pass. It therefore means that the selected query is invariant to a real state change, not that the intervention was ineffective.
Strict answer-label checkerboards remain mandatory for separately designated qualification/evaluation boards. They are not a universal initializer-data shape and cannot be inferred from the generic rectangle type.
The legacy CPU materializer keeps strict checkerboards as its default. V3
callers must explicitly set require_query_checkerboard=False; non-Boolean
mode values fail closed. This prevents a broad v3 caller from accidentally
changing v2 behavior or relying on an implicit global relaxation.
The Horn and resource semantic adapters independently recompute every
broad-mode answer from each corner.
The adapters also keep dependent-composition admission strict by default.
V3 explicitly requests require_dependent=False only while revalidating the
four broad rectangle corners, matching the execution mode used to construct
those corners. This does not weaken stage admission: dependent-composition
and closed-loop factual candidates are still generated and reconstructed
under their strict dependent-prefix contract.
Transaction traces are effect-normalized. Every disclosed command atom is written to its own command slot, but a cursor write is emitted only for the first cursor value and subsequent cursor changes. Atomic-deadlock suffixes therefore retain the complete command while avoiding repeated writes of an already frozen cursor. The receiver still rejects every emitted microstep that has no generic state effect and still requires exact terminal-packet agreement after independent replay.
4. Query objective
For every matched intervention pair:
- cross-entropy supervises both factual answer labels;
- if
correct_target != foil_target, the pair receives the original difference-in-differences causal margin; - if
correct_target == foil_target, the pair receives Jensen-Shannon invariance supervision between the two predictive distributions; and - an effect pair can satisfy the margin only when its measured difference-in-differences reaches the configured margin.
The Jensen-Shannon term is finite, symmetric, and bounded. It does not create an oracle route: the loss receives only the same immutable factual targets already used for answer cross-entropy.
5. Receipts
Objective and optimizer-update receipts separately report:
- all classification-supervised query pairs;
- answer-changing contrast pairs;
- answer-preserving invariance pairs; and
- contrast pairs satisfying the causal margin.
For each WORLD and COMMAND arm:
contrast_pairs + invariance_pairs = supervised_query_pairs
margin_satisfied <= contrast_pairs
These equalities are checked on device. The objective schema is
shohin-ettr-composite-objective-v3.
Qualification source custody measures the complete physical pipeline/
tree and the exact five-file train/ runtime closure imported by CPU tensor
materialization. The measured inventory must equal the physical executable
tree. Extra, missing, linked, or unmeasured executable objects are rejected
before candidate rows are read.
6. Acceptance
The repair is accepted only if:
- every existing strict v2 objective and data-contract test still passes;
- all-answer-changing rectangles retain their prior support and margin behavior;
- all-answer-preserving rectangles produce finite differentiable loss, zero contrast support, and full invariance support;
- mixed rectangles partition support exactly;
- packet-no-consequence rectangles still fail closed; and
- strict v2 dependency defaults and explicit v3 broad-corner mode are both regression-tested;
- repeated deadlock cursor writes are absent while the full command remains represented; and
- architecture-facing v3 materialization and its global audit bind this contract and exact source commit in their receipts.
This contract authorizes no model fitting.