determinism-tiers: a tier covers what the sink emits, not the decode alone #81
Loading…
Reference in a new issue
No description provided.
Delete branch "wai/determinism-tier-covers-emitted-medium"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
decode-equivalenceclaims "the whole decode to the final medium reproduces byte-for-byte", and §3 makes it the precondition for a receipt carrying a reproducible decoded-medium hash and for a hard provenance binding over the rendered content.But the tier was defined over the decode stage only. The file has no occurrence of
post-decode,post-filter,SEI, orhint— verified directly, not inferred.What walks through the gap
Two mechanisms in force today, both arriving in the same pipe as the codestream:
nnpfc/nnpfaSEI activate a neural post-filter after decoding, changing output pixels.So a capability can hold
decode-equivalencehonestly, emit a receipt carrying a "reproducible decoded-medium hash", and have that hash describe something the viewer never saw. The hash verifies. It attests the wrong artifact.This is the same defect shape as the proof ladder fixed earlier in #65: a claim that outruns what was actually verified. There, an absent solver escalated into a stand-in that proved anything; the fix was that it yields
Unproved. Here, a tier claim is void if the sink alters the medium after decode.The fix
Adds §2 "What a tier covers" with three rules:
decode-equivalence, and a receipt over its output MUST NOT carry a reproducible decoded-medium hash.Plus a §3 cross-reference so the rule is visible where tiers are declared.
Rule 2 is the one with teeth for the energy work: reporting joules while concealing the degradation that bought them states a cost without its consideration.
No wire format change — vocabulary and conformance framing only, as the extension already is. Surfaced by the landscape survey in #80 and verified directly against this file.
🤖 Generated with Claude Code
`decode-equivalence` claims "the whole decode to the final medium reproduces byte-for-byte", and §3 makes it the precondition for a receipt carrying a reproducible decoded-medium hash and for a hard provenance binding over the RENDERED content. But the tier was defined over the decode stage only, and said nothing about processing applied after a conformant decode. Two mechanisms in force today walk through that gap, both arriving in the same pipe as the codestream: * ITU-T H.274 (V4, 01/2026) `nnpfc`/`nnpfa` SEI activate a neural post-filter AFTER decoding, changing output pixels. * ISO/IEC 23001-11 decoding-operation-reduction metadata exists precisely to make a decoder spend fewer joules by doing less work — deliberate degradation, invisible to a conformance check that stops at the decoder. So a capability could hold `decode-equivalence` honestly, emit a receipt with a "reproducible decoded-medium hash", and have that hash describe something the viewer never saw. The hash verifies. It attests the wrong artifact. This is the same defect shape the proof ladder had: a claim that outruns what was actually verified. There the fix was that an absent solver yields Unproved; here it is that a tier claim is void if the sink alters the medium after decode. Adds §2 "What a tier covers" with three rules — evaluate the tier over the EMITTED medium; an object honouring a reduction hint MUST NOT claim `decode-equivalence` nor carry a reproducible decoded-medium hash; a registration SHOULD state whether its payload can carry post-decode directives at all — plus a §3 cross-reference so the rule is visible where tiers are declared. The second rule is the one with teeth for the energy work: reporting joules while concealing the degradation that bought them states a cost without its consideration. No wire format change; vocabulary and conformance framing only, as the extension already is. Found by the landscape survey in #80, then verified directly against this file — it has no occurrence of post-decode, post-filter, SEI, or hint. Co-Authored-By: Claude Opus 5 <[email protected]>03a1db0ebcf5cfabb8c3