Exact

Exact Union RegExp

Repository files (latest commit first)

Filename	Latest commit message	Latest commit date
Transaction Science e0285b049b wai.neural.int_hyper: leaves, a parameter set of WAI's own, trained on generated pictures and held to one decode natively and in WebAssembly ... The set: tools/wai_int_hyper_train.py at 40db6ee57 (the commit that brought it) trained a mean-scale hyperprior, 48 synthesis and 64 latent channels, from seed 20261110 on the dead-leaves pictures tools/wai_int_synth_real.py generates from that seed, streams 1 to 4096: 30,000 steps of 16, Adam at 1e-3 then 1e-4 for the last quarter, lambda 0.0483, on the CPU at 8 threads with deterministic algorithms (Python 3.11.17, torch 2.14.0+cpu, numpy 1.26.4, compressai 1.2.8; the Framework build host, Linux x86_64). Exported by export_model and coded on stream 0 (128 x 128). Listing digest (the set's identity): ab796bb19cb202115400ea2ce16724617c3da035a2f4429d6266d419ab705ac0. Its record, wai-rs/tests/vectors/int_hyper/leaves_set.json, reads Own: Apache-2.0, a start from a seed, and training and coded pictures generated by a tool below wai/ with no third-party source; the tools' blob ids are their blobs at 40db6ee57. Reproducibility: two runs from the same seeds, side by side on that host, wrote the same bytes in every file (float model 2ce350db..., all three members, the bitstream, the decode and the record), and logged the same loss at every logged step. 300-step runs at 8 and at 4 threads from the same seeds agree at step 0 and differ by step 299 (float models 1820d9ec... and 699bfbea...): the thread count is part of the recipe, and the record states it. Where its decode is held, each to the one digest in int_hyper/sha256.json (3ea9b4c3026e7c81683f3845f8256b1f0e1e8268284d81f3f16f290a5dde1374): - int-payload-conformance gains `leaves`, a WIH1 payload (SHA-256 048c7c0d...), decoded natively by wai_int_payload_vectors verify; int_payload_verify.py parses and converts it independently. - wai-int-module/corpus.json, 121 cases: both modules of record gave every case its outcome on wasmtime 50.0.1 (Cranelift, Winch) and Wasmi 2.0.0, run digest f245b5f0... (Framework). The browser crate's Node check names the payload too (CI). - byte-exact-conformance: wih1_trained_set_is_golden, by the sink's decode and through the lowering under the reference and the fast kernels, on every leg of the matrix. - ops-conformance certifies it (index and two lowered graphs): acc64, first acc32 failure at h_s's input range, 43 bits; int_hyper_ops holds the set and the vector to them. verifiers/int-hyper-set.sh runs `check --own` and `self-test`: the unchanged record and a copy started from a set that reads Own read Own, and 11 planted faults are caught (one weight moved by one; a third-party source in the training record, beside generated pictures, or in the start's record; a start from a checkpoint that is not WAI's; each of three seeds removed; the coded picture among the training pictures; a decode digest changed; the decode of another picture with its digest restated). The trainer's check of a start from a set now reads a start that does not read Own as `not own`, rather than failing. Run through rb on the Framework (Linux x86_64), each passed: wai_int_payload_vectors verify (release and debug), int_payload_verify.py, check --own, self-test, corpus.py --check, wai-verify vectors verify, wai_ops_check.py, byte-exact-conformance (release), int_payload (debug), int_hyper_ops (debug and release), corpus_provenance, the module host on all 121 cases. Decode-side plants, each caught, with the committed case passing first: a WebAssembly decode of the set with one g_s weight moved (0 of 1 cases the corpus's), and the same drift in the committed file under the native corpus verify, the record check and the byte-exact golden. Co-Authored-By: Claude Opus 5.5 <[email protected]>		2026-11-10 13:41:55 -05:00
..
wai.audio.int_codec	wai-ops: load step 8, the bounds pass and the certificate (D2b)	2026-10-10 09:26:27 -04:00
wai.image.jpegai	wai: gate 5's falsification found the synthetic networks refuse every decode	2026-10-18 01:43:32 -04:00
wai.neural.int_hyper	wai.neural.int_hyper: leaves, a parameter set of WAI's own, trained on generated pictures and held to one decode natively and in WebAssembly	2026-11-10 13:41:55 -05:00
wai.neural.int_mlicpp	wai-ops: load step 8, the bounds pass and the certificate (D2b)	2026-10-10 09:26:27 -04:00
wai.neural.int_synth	wai/neural-corpus-licence, converted to the layout of the fix: no shared append point	2026-11-05 01:45:34 -05:00
wai.splat.int_codec	wai-ops: load step 8, the bounds pass and the certificate (D2b)	2026-10-10 09:26:27 -04:00
wai.video.int_hyper	wai-ops: load step 8, the bounds pass and the certificate (D2b)	2026-10-10 09:26:27 -04:00
wai.world.model	wai-ops: load step 8, the bounds pass and the certificate (D2b)	2026-10-10 09:26:27 -04:00
external.json	wai: D2c M7b gate 1: the networks' lowering and certificate, reproduced from the text	2026-10-18 01:02:55 -04:00
falsification.json	wai-ops: load step 8, the bounds pass and the certificate (D2b)	2026-10-10 09:26:27 -04:00
index.json	wai.neural.int_hyper: leaves, a parameter set of WAI's own, trained on generated pictures and held to one decode natively and in WebAssembly	2026-11-10 13:41:55 -05:00