Research line: integer serving (M4)
This line is about running a finished model on an ordinary laptop or in a browser using only whole-number arithmetic, adds and table lookups: no floating point and no multiplication. In the records the engine reproduced its float models exactly on the old stack, and the multiply-free audits pass on the WASM serving symbols. What is missing is a trained model worth serving: the models tried scored 1.20 to 1.24 bits per byte against a 0.90 target, and no model that obeys all the current rules has been exported.
| Catalog id | m4-serving |
| Time span | 7 Aug to 18 Oct 2026 (6 records undated) |
| Milestone | M4, one served model under D11 and a command-line interface (12 of 42 records); 4 tagged M1 |
| Records | 42 |
| Outcomes | tooling 10 · negative 9 · positive 7 · inconclusive 6 · review 3 · ops 2 · process 2 · blocked 1 · diagnostic-stop 1 · plan 1 |
| State | The D11 engine is ACTIVE (D11 integer serving). M4 is held; the export gap of a lawful model is unmeasured |
Counts: archive/catalog.jsonl. D11 is the owner decision that every lab serves under one multiplier-free contract. Terms follow the glossary; ACTIVE, READY and PARKED are defined on the mechanism board.
flowchart LR A["Aug<br/>R4G1 serving profiles,<br/>skip-mix, source-free table"] --> B["8 Sep<br/>Alpha claim withdrawn"] B --> C["9 to 10 Oct<br/>Served vs float gap<br/>is the model, not the engine"] C --> D["10 to 11 Oct<br/>Softmax-free reads<br/>and icosian weights"] D --> E["18 Oct<br/>WebAssembly runtime,<br/>zero-multiply audit"]
The engine was settled early. The quality gap was never in the engine.
What it set out to do
Make fast, cheap, local inference real: one served model, integer-only (D11), reachable from a command line, with tokens per second and energy measured on the owner's laptop. The product is serving; panel scoring and training stay on the GPU.
What happened, in order
- August. A source-free integer n-gram table predicted 22.26% of held-out targets against 5.41% for unigram, and one exact tie-break added 0.95 points (table, tie-break). The skip-mix lane beat base by +28.45 per mille top-1 (record). A vector (SIMD) scan sped one path up 1.60 times against a 2 times bar (record).
- 8 Sep. The alpha publication was withdrawn: the scorecard had used constant or flag evidence (record).
- 9 to 10 Oct: the gap is the model. The published 0.877 float against 0.933 served compared different numbers of windows. On a matched retrain the pair is 0.8776 against 0.8868 bits per byte (protocol pin). The recorded criterion is still "not met and not claimable", and the matched float model sits at 0.924 on the attribution record's own table. At identical positions quantization cost 0.0041 and the engine cost -0.0000008 (attribution). The integer engine matched float exactly (maximum logit difference 0) while three served models scored 1.20, 1.24 and 1.24 against the 0.90 target (D11 serving).
- 10 to 11 Oct: lawful pieces. Replacing the served softmax read with rank-weight tables cost 0.017 to 0.26 bits per byte more across four arms, all missing a 0.01 bar (softmax-free read). Weights built from two icosian table entries and a 2-bit mantissa reached 1.5148 against 1.4944 for float quaternions and 1.5136 for dense of equal size (icosian mantissa). The integer engine refuses product-key memory entirely, so the speed case there is unmeasured (port).
- 18 Oct. An outside contributor's WebAssembly integer runtime was merged after hardening: 165 tests pass and 7 of 7 serving symbols show no multiply, divide or float. The vector (SIMD) kernels are not integrated and the demo is not runnable from a clean clone (record).
Key results
| Result | Kind | Source |
|---|---|---|
| Integer engine equals float: negative log-likelihood difference 0.0, maximum logit difference 0 | measurement | D11 serving |
| Served quality 1.20138, 1.24402 and 1.24233 bits per byte; target at most 0.90 | measurement | same |
| Speed fell from 63.41 tokens/s at 384 context to 30.43 at 4,096, so replacing the quadratic read buys at most about 1.7 times | measurement | context curve |
| 3-bit E8 quantization of SmolLM2 MLP weights was +4.89 nats worse than 4-bit round-to-nearest, tripping its kill rule | measurement | record |
| Zero-multiply audit: 44 native symbol ranges and 7 of 7 WebAssembly serving symbols | measurement | WASM runtime (7 of 7); project track (44 native ranges) |
What it taught
Check the instrument before blaming the engine (lesson 1 in what the project knows): most of the apparent served-versus-float gap came from comparing different numbers of windows. Measure the shipped form (lesson 2). Savings counted on paper are not tokens per second (lesson 13). Exact integer export and bit-exact serving work (lesson 17).
Current state
The zero-multiply and bit-exact properties are verified on the old stack and on the WASM serving symbols. The engine has not served any model that obeys the current rules, and the WASM record lists its SIMD kernels as unintegrated fixtures, so its contract is not yet exercised end to end. The 19.9M-parameter stack that missed the old M1 bar read every weight on every token, and two of the integer schemas serve softmax through exponent tables, which the current rules ban (docs/labs/repo-renormalization-2026-10-19/audit/synthesis.md, section 1). The next step is to export one discrete artifact and measure its export gap (board). Product serving, including tokens per second and energy on the laptop, is allowed; panel scoring of an exported artifact is still GPU research compute.
Go deeper: the generated catalog page lists all 42 records.