# wai-det-jpegai — the JPEG AI exact decoder at the float level The module of record of WA²ISM™'s float level ([`extensions/waism.md`](../extensions/waism.md), `waism/1-det`): one WebAssembly module, with the interface of waism §5 and nothing else, that decodes `wai.image.jpegai` through `wai-rs`'s exact decoder (`jpegai_decode`). That is the decode a native sink runs when it holds the capability's integer weight bundle: the integer entropy decode and scale path, the integer networks through the integer operator set (`jpegai-int/1`), then dequantisation, latent refinement, the post-filters and colour conversion in IEEE-754 float, in the fixed order of the source. The bundle is not in the module: it arrives as the capability's parameter set. | capability | payload | set | output | |---|---|---|---| | `wai.image.jpegai` | the codestream | the file sequence of `norm.json` and `weights/*.json` | RGB8, rows top to bottom | A refusal is the capability's own code, with `wai_run` returning 1: `set_incomplete`, `model_mismatch`, `stream_exhausted`, `field_range`, or `configuration_undetermined` for a configuration whose result the JPEG AI reference software leaves to the machine it runs on (residual variance scaling without GRFS, a signalled colour matrix; SPEC §5, the `wai.image.jpegai` row; [decisions](../decisions.md) D74). A codestream outside the configurations the decoder implements ([determinism-tiers](../extensions/determinism-tiers.md) §5) has no code in that row: the module traps, and the sink refuses the decode through it (`module_trap`). ## What holds it - **Held at the float level, refused at the integer level.** Floating point is in the module, so both of the profile's checkers ([`../wai-module-check`](../wai-module-check)), the Rust checker and `check-module.mjs`, give it `held` at the float level and `refused R3` at the integer level, the same verdict from each. No import, one memory with a declared maximum (8,192 pages; a decode of the reference codestream reaches 3,964), the six exports of §5. Every allocation traps where memory cannot grow, so a failed `memory.grow` inside `wai_run` ends in a trap and never in an output (R13); `wai_alloc` alone returns 0. - **The probe first.** At the float level a sink states a tier for the output only on an engine set to the deterministic profile's NaN rule, shown by the probe of waism §9.3 run in the same process before the module's first run there (§6, check 10). The probe is [`../wai-module-check/planted/probe.wasm`](../wai-module-check/planted/probe.wasm), and the bytes it must give are the `bytes` line of [`probe.txt`](../wai-module-check/planted/probe.txt) beside it: the positive canonical NaN in every lane of the twelve cases. `planted/build.py` writes both, and both checkers hold the probe at the float level. - **The module corpus.** [`corpus.json`](corpus.json), written by [`corpus.py`](corpus.py): the reference codestream (`wai-rs/tests/vectors/jpegai/test256.bin`) over the pinned `jpegai/v1` bundle, to the SHA-256 of the RGB8 the native exact decoder gives (`wai-rs/tests/jpegai_cross_platform.rs`, `PINNED`'s `rgb`), and three refusal fixtures: the bundle without `weights/icci.json` (`set_incomplete`), and with `norm.json` or `weights/icci.json` not JSON (`model_mismatch`); and the configuration vectors (`wai-rs/tests/vectors/jpegai/configs/`, `jpegai/configs/`): each one in scope to the RGB8 digest its manifest pins (the native exact decoder's), the two whose configuration the reference leaves to the machine to `configuration_undetermined`, and each the decoder does not implement to the sink's `module_trap`. The bundle's files are named under `external`, with their SHA-256 from `tools/jpegai-e2e-v1.sha256`; `tools/wai_jpegai_fetch_e2e_inputs.sh` fetches them and checks each. The module contains `memory.grow`, so §9.2 also gives it grow-refusal cases; `corpus.json` does not hold them, and no run here makes a `memory.grow` fail. - **The engines.** [`host/`](host) runs the probe on each engine, then every case on an instance of its own, as waism §5 and §6 checks 10 to 13 state, holds every output to the corpus, and prints a **run digest**: the SHA-256 of one line a case, its name and its outcome. Each engine is set as waism §7 sets one at the float level: `threads-off`, `relaxed-off`, fuel, the memory limited to its declared maximum, a fixed stack, and `nan-canonical`. | engine | how its NaN rule is set | |---|---| | `wasmtime@50.0.1/cranelift` | `Config::cranelift_nan_canonicalization(true)` | | `wasmtime@50.0.1/winch` | the same setting, which Winch honours | | `wasmtime@50.0.1/pulley` | Cranelift compiling to Pulley, wasmtime's portable bytecode, run by its interpreter, with the same setting | | `wasmi@2.0.0` | the crate feature `deterministic`, which canonicalises every NaN a floating-point instruction gives | ## F1 and F2 - **F1: the probe tells the engines apart.** On wasmtime with Cranelift, with Winch and with Pulley, the probe gives its bytes with the NaN rule set and other bytes with it unset; on Wasmi built with its rule, its bytes. `host --f1` runs it on all seven and holds them to that. - **F2: the decode is the native exact decoder's on every engine the probe holds.** The module of record decodes the corpus's vector and each configuration vector in scope to the native exact decoder's pinned RGB8 (`dbc494a5…` for `test256`), gives each refusal fixture its code and each configuration the reference leaves to the machine `configuration_undetermined`, and traps (the sink's `module_trap`) on each configuration the decoder does not implement, on the four engines above, on two instruction sets. F1 was measured with the earlier module of record, on Linux x86_64 and Linux arm64, by `host --f1`; the probe and the engines are unchanged since, and on 2026-11-12 F1 held again on Linux x86_64 and on macOS arm64: | host | F1: the probe's bytes, the NaN rule set | F1: other bytes, the rule unset | |---|---|---| | Linux x86_64 | Cranelift, Winch, Pulley, Wasmi | Cranelift and Winch in all twelve cases, Pulley in eleven | | Linux arm64 | Cranelift, Winch, Pulley, Wasmi | Cranelift and Winch in nine cases, Pulley in eight | With the rule unset, a case on a NaN operand (1, 2 and 6 to 12) keeps that NaN's sign, quieted, on both instruction sets, except `f32.nearest` (case 7) on Pulley, which gives the canonical NaN; a NaN made from numbers (cases 3 to 5, and lane 3 of case 11) is negative on x86_64 and the canonical one on arm64, whose hardware makes that NaN. F2, 2026-11-12, the module of record over the corpus's 63 cases on all four engines, on both instruction sets (`host --engines E --external J corpus.json record/wai_det_jpegai.wasm`, the probe first on each engine in the same process): | host | engine | cases the corpus's | time | |---|---|---|---| | Linux x86_64, 32 threads | Cranelift | 63 of 63 | 173 s | | Linux x86_64 | Winch | 63 of 63 | 355 s | | macOS arm64, 28 cores | Cranelift | 63 of 63 | 166 s | | macOS arm64 | Winch | 63 of 63 | 429 s | | Linux x86_64 | Wasmi | 63 of 63 | 1,082 s | | Linux x86_64 | Pulley | 63 of 63 | 3,135 s | | macOS arm64 | Wasmi | 63 of 63 | 992 s | | macOS arm64 | Pulley | 63 of 63 | 4,023 s | Every run gave one run digest, `85141976194249c9a3c81c252cb7100fad8ce5df69bfa2c5c6373bd2f027b141`, and an instance reached at most 3,964 pages of the 8,192 the module declares. A configuration vector's decode takes the interpreters minutes where the compilers take seconds (a 192 × 192 picture with eICCI off: 4 s on Cranelift, 8 s on Winch, 27 s on Wasmi and 68 s on Pulley, compilation included, Linux x86_64), so CI runs every case on Cranelift and Winch and the `jpegai/v1` cases on Pulley and Wasmi (`.github/workflows/wai-det-jpegai.yml`). The earlier module of record (666,425 bytes, `53d54e58…`), run past the probe (`--no-probe`) on the three engines with the rule unset, gave every case of the corpus as it then was the corpus's on both hosts: no NaN its decode of that corpus made reached its output. A sink does not rest on that. At the float level the output holds a tier only on an engine the probe holds (§6, check 10), and then the tier §2 gives that level. ## The module of record [`record/`](record) holds one build of this crate for `wasm32-unknown-unknown`, named by its SHA-256 in [`record/manifest.json`](record/manifest.json): `wai_det_jpegai.wasm`, 673,165 bytes, `e13ebb9640dec9d576df61cb834c6e4c5fdd8c271c60a7b4dc343fa6987fa2bb`, built on Linux x86_64 from main at `f218af5ac`. The module is the same source on every host, and its bytes are not ([`../wai-int-module`](../wai-int-module) says why), and they move with any change to the decoder's crates: a build of the same decoder one main commit earlier was 673,157 bytes (`acc6c819…`) and gave the same run digest. So `record/` is named by its SHA-256, and a build is held to the corpus, not to `record/`'s bytes. ## The planted faults Each check is shown failing on the fault it exists for. - **A probe with a NaN-dependent branch fails F1.** `probe-nan-branch.wasm` is the probe with a branch after each operation on whether its result is a NaN, writing the canonical NaN itself. It meets R1 to R12, and both checkers hold it at the float level, but it gives the probe's bytes on every engine, the ones with the NaN rule unset among them: `host --f1 --probe planted/probe-nan-branch.wasm` reports that it accepted an engine whose NaN rule is not set, and fails. - **A decoder with a NaN-dependent branch departs from the pinned output where the probe refuses the engine.** The feature `planted-nan-branch` (`build.sh --planted`) adds to the decode a branch on the sign of a NaN made at run time from a negative signalling NaN, which flips one bit of the output when the sign is set. The deterministic profile makes that NaN the positive canonical NaN, so on the four engines above the planted module still decodes to the pinned RGB8. On the same wasmtime engines with the NaN rule unset, run past check 10 (`--no-probe`), its output is another; with check 10 in place, the probe refuses those engines (`module_probe_mismatch`) and the module never runs there. ## Building and running ``` bash build.sh [--out DIR] [--planted] # the module, held by both checkers bash ../tools/wai_jpegai_fetch_e2e_inputs.sh J # the jpegai/v1 bundle, each file checked (cd host && cargo build --locked --release) host/target/release/wai-det-jpegai-host --f1 host/target/release/wai-det-jpegai-host --f1 --probe ../wai-module-check/planted/probe-nan-branch.wasm host/target/release/wai-det-jpegai-host --external J corpus.json record/wai_det_jpegai.wasm python3 -I corpus.py --check # corpus.json is what corpus.py writes ``` `build.sh` names no path of the machine that runs it in the module: the checkout, the Cargo home and the toolchain are remapped. CI: `.github/wai-conformance/verifiers/module-check.sh` holds `record/` to its manifest, the module of record at the float level and refused R3 at the integer level by both checkers, and the corpus to `corpus.py`; `.github/workflows/wai-det-jpegai.yml` runs the build, F1, the engines and the planted faults on Linux x86_64 and Linux arm64.