Trust as code
Verify an answer once. Every re-ask after that is a read, not a re-roll.
THE BRIEF
An answer becomes an asset: stored with its provenance, checked independently, then recalled byte for byte. Disagreement goes to human review.
0wav is an asset-side determinism layer for multimodal AI: speech recognition, diarization, prosody, and encoder features, stored so that a verified answer is locked byte-exact with its provenance.
The verifier is never the model's own confidence. Five transcription engines from four vendor families run the same audio; where all five agree, the word locks. Where they argue, it goes to a human, and the ruling is hash-bound to the exact queue entry.
ON THE RECORD
PORTFOLIO SNAPSHOT / JUL 2026Figures and source references supplied with this portfolio. Scope and limitations accompany each result.
Read speech, certified
≥99%of accepted words correct, pre-registered
9 of the 12 misses were archaic spellings in the gold transcript.
Reported measurement · asr_gold_gate_alpha001_2026-07-12.json — 12/1,952, CP-95 upper bound 0.009941
Financial QA, certified
≥92.2%of accepted answers correct at 90% confidence
Effective n ≈ the unique underlying questions. The certificate covers this suite's query mix.
Reported measurement · metaharness_descent2_pooled_2026-07-18.json — n=2,271 pooled, 21.6% coverage
Byte-exact replay
20 / 20identical SHA across live runs
Reported measurement · T3a replay lane
Batch-shape drift
63.9%of cells changed answer based on who shared the server
Stated as a loss on purpose. It is the industry default this work replaces.
Reported measurement · inference_batch_card.json — B ∈ {2,4,8}
Receipt chain
×36 intactevery iteration re-verifies from its own stored bytes
Reported measurement · loop-contract trail, ADR-043
Energy per re-ask
~100,000×less than regenerating the answer
Declared constants, not a metered measurement. Shown with CO₂e beside cost.
Declared assumption · harness constants — 1.0 Wh fresh pass vs 0.00001 Wh recall read
Handing the model the whole document tree scores 9.6. Deterministic structure alone beats it by 43 points. The model is used at the last inch only — one bounded call per query — and it adds the final 7.4.
CARD · multi_doc_financebench_hybrid_llm_2026-07-06.json
Five stenographers from four different companies transcribe the same tape. Where all five write the same word it matched human gold 998 times in 1,000 — a 216× separation from the words they argue about. Those go to a human, and the ruling is hash-bound to the exact queue entry.
CARD · asr_gold_gate_2026-07-11 · 1 miss in 479, and it was the gold's archaic spelling
Ask a cloud model the same question twice and you can get different answers — not because it learned anything, but because of which other requests shared the batch. Nearly two-thirds of outputs shifted. A locked answer cannot be touched by that.
CARD · inference_batch_card.json (T3b) · T3a replay 20/20