Skip to content

0051 — Complete native synthetic encoder graph

Date: 2026-10-21. W18 / issue #19 preparation. This is a new synthetic software experiment, not W12 completion or an admitted backend. The protocol was recorded before implementation in runs/w18-complete-encoder-1792600222383037000/protocol.json. The baseline is 4e3258194b2484cfe69614f100587aee4459e946.

The hypothesis is that the existing nine integer operations can express and execute a complete one-layer encoder, including both normalizations and every input unit. The earlier attention and normalization proofs are separate relations; neither proves this new graph. Their fixtures and evaluators stay unchanged.

Use an authored two-token, two-feature, single-head graph with byte embeddings, fixed positions, query/key/value projections, attention, residual, normalization, a three-unit ReLU feed-forward layer, second residual/normalization, masked sum-plus-max pooling and two logits. The scalar unsafe-minus-safe score uses strict score < threshold; equality blocks. Constants are authored, not trained. There is no floating parent or task-quality estimate.

The versioned tokenizer recipe maps every original UTF-8 byte to the identical integer ID in 0..255, without normalization or remote code. Its exact bytes and SHA-256 are bound by the program, input profile and classifier. The program also binds constant-file SHA-256, arithmetic identity and named token/mask roots. ClassifierManifest.tokenizer is the raw recipe file hash in this profile, just as its weights field binds the raw constant file. This specifies one tokenizer; other recipes reject even when their new hashes are supplied consistently.

Original content is partitioned contiguously, with fixed prefix/suffix tokens on every chunk, including empty input. Padding has a separate Boolean mask. Empty input creates one chunk. Maximum content is 32,768 bytes, measured independently of token padding; the input commitment uses all original units and the existing classifier-specific salt derivation. Prefix and suffix leave at least one content slot. Special-token and padding IDs must also be bytes. A byte token can split a multibyte character; concatenation reconstructs the original UTF-8 exactly.

The graph is frozen in fixtures/wire/integer-encoder-v1.json, with a separately retained SHA-256. Attention uses the explicit table over scores -16..32: 1 for negative scores, otherwise 2**(score//8). This is a declared synthetic piecewise kernel, not a floating softmax-equivalence claim. Masked denominator reciprocals are round(256/n), with zero mapped to zero. All signed rescaling uses exact nearest ties to even.

Each normalization applies decision 0050's integer mean, centered-square mean, epsilon-one inverse-square-root and gamma/beta formulas to its full declared input domain. The first domain is -129..130 and the second is derived from the frozen feed-forward bounds. For this two-feature graph only, exact finite tables lower the pair to two Q8 outputs. The graph constrains the pair index, feature offset and returned values. The tables are generated by integer comparisons and checked entry by entry against a separate 80-digit Decimal implementation. This does not reuse the four-field normalization proof at the larger domain.

IntegerGraph validates the existing parameter and topology contracts once and executes their nine operators on flattened row-major tuples. It checks exact root sets, types, shapes, ranges, Boolean masks and lookup indices. Multiplication, scale alignment and every partial sum must fit signed int32 before rescaling; no overflow is hidden by later cancellation or scaling. The runtime rejects resource excess instead of truncating. It bounds total declared tensor elements to 262,144 and work per chunk to 2,097,152. execute_encoder additionally bounds constant bytes to 4 MiB, aggregate work to 33,554,432 and trace elements to 8,388,608. Returned traces are private and parsing them does not authenticate computation. This executor is not an OS sandbox or proof verifier.

Acceptance requires every tensor in frozen empty, boundary, Unicode, NUL and multi-chunk vectors to match independent rational/Decimal formulas; all four embedding classes and masks to preserve padding isolation; exact near-maximum and maximum input coverage; strict-threshold tests; and rejection of changed bindings, malformed inputs, overflow and oversized graphs. Installed-wheel checks must execute the same caller-pinned small vectors. Keep initial failures.

The standing approval permits one foreground Mac campaign, four threads, 7,200 seconds, 8 GiB worker RSS, at most 2 GiB new artifacts and zero spend. The checkout initially consumed 90,720,592 KiB and had 524,486,598,656 free disk bytes. No cloud job, model download, private data, new dependency or reviewer contact is needed. Freeze final code, fixture, environment and commands before the retained campaign; repeat into a fresh directory and compare substantive results.

This graph is a synthetic pipeline candidate. Larger widths/contexts, learned weights, float/PTQ drift, quality, circuit lowering, whole-graph proof costs, operational enrollment, independent verifier/security review and hardware attestation remain separate gates. The finite normalization tables are practical only at this deliberately small width. No prior statistical result is reused, and owner review is required before any new profile reuses prior results.