df20c392d8
Adds the "carries whatever it can reconstruct, verifies whatever it
carries" surface on top of the existing neural/codec WAI. All builds
green against this base (wai-rs all-features tests + clippy, wai-web
wasm, wai-spz-gen, wai-relay).
Reconstruct half — each a spec + conformance corpus + in-tab browser
sink that re-derives the corpus hash, over one wai.det.fixed64 floor:
- interactive worlds replay-equivalence (3 sim registrants,
relay multiplayer, session receipts, glTF + real-SPZ assets)
- spatial audio mixdown-equivalence
- haptics sample-equivalence
- state-feeds replay-equivalence (non-physics sim)
- avatars / digital humans pose-equivalence (+ telepresence)
- volumetric / 4D splats frame-state-equivalence (WebGPU EWA render)
Verify half (lever 4):
- asset provenance — signed, energy-accounted lineage for opaque
captured content WAI does not reconstruct; tamper-evident in-tab.
Cross-cutting: per-class JWP receipt profiles; a native energy meter
(IOReport via macmon, no sudo) emitting receipts with real measured
joules. Codecs gated behind a `codecs` feature (default on) so the
world/wasm sinks build without the C codec chain.
Excludes the separately-evolving neural subsystem (semantic path, GLC
rANS, ONNX tooling) to avoid reverting newer origin work.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
89 lines
4.1 KiB
Markdown
89 lines
4.1 KiB
Markdown
# WAI Extension: Asset Provenance
|
|
|
|
> Status: Draft. Lever 4 — content-addressing + receipts + measured
|
|
> energy for content that cannot be reduced to instructions. WAI does not
|
|
> reconstruct the asset; it carries the opaque bytes and a signed,
|
|
> energy-accounted lineage. Conformance is **content-hash-equivalence**
|
|
> (recompute, compare) + signature validity. Reference impl: the
|
|
> `provenance` feature of `wai-rs`; tooling: `wai_prov`; live sink:
|
|
> wai.transaction.science/provenance.
|
|
Keywords **MUST**, **MUST NOT**, **SHOULD**, **MAY** are RFC 2119/8174.
|
|
|
|
## 1. Scope and model
|
|
|
|
A `wai.asset.provenance` object is **not a content capability** — it is a
|
|
JWP receipt profile *about* a carried asset, the way `jwp-receipts` is.
|
|
The asset itself declares an ordinary capability (`wai.image.jpeg`,
|
|
`wai.audio.flac`, …) and WAI carries it unchanged; the provenance receipt
|
|
adds three things the carried bytes do not have on their own:
|
|
|
|
1. **Tamper-evidence** — the asset is content-addressed
|
|
([`asset_content_hash`], BLAKE3 over the complete WAI envelope); a
|
|
verifier recomputes it and compares.
|
|
2. **Lineage** — an ordered chain of processing **steps**, each binding
|
|
its input and output content hashes, signed as one group.
|
|
3. **Energy accounting** — each step carries its measured/declared
|
|
microjoules; the receipt totals them.
|
|
|
|
This is the "verifies whatever it carries" half of the thesis, for the
|
|
content the reconstruct half cannot touch. WAI never decodes the asset —
|
|
a sink hands the bytes to whatever native element renders that media.
|
|
|
|
In scope: the step model, the `AssetReceipt` group, the verification
|
|
obligations. Out of scope: a transport; a capture device; the asset's own
|
|
codec (it is opaque here).
|
|
|
|
## 2. The step
|
|
|
|
A `ProvStep` is one transformation in the asset's lineage:
|
|
|
|
| field | meaning |
|
|
|---|---|
|
|
| `op` | operation, e.g. `"origin.synthesize"`, `"encode.jpeg"`, `"derive.thumbnail"` |
|
|
| `input_hash` | content hash of the input (all-zero for an originating step) |
|
|
| `output_hash` | content hash of the output |
|
|
| `work` | exact deterministic work (pixels, samples) |
|
|
| `joules_micro` | measured/carried microjoules (0 = unmetered) |
|
|
|
|
Step content hash:
|
|
`BLAKE3("wai:prov-step\x01" || op_len | op | input | output | work | joules)`.
|
|
|
|
## 3. The receipt (`AssetReceipt`)
|
|
|
|
A JWP group: the asset's content hash is the identity, the per-step
|
|
hashes are the Merkle leaves (JWP's exact `moq-jwp:merkle-{leaf,node}`
|
|
domains), and one Ed25519 signature covers a profile payload carrying the
|
|
asset hash, the declared `media` + `capability`, the Merkle root, the
|
|
step chain, and the total joules. It chains across edits via
|
|
`parent_receipt_hash`, like the worlds / audio / haptic / splat receipts.
|
|
|
|
## 4. Conformance
|
|
|
|
> A conforming verifier MUST: (a) recompute the asset's content hash and
|
|
> require it equal the receipt's `asset_hash`; (b) require the step chain
|
|
> be linked — each step's `input_hash` equals the previous step's
|
|
> `output_hash`, and the final `output_hash` equals `asset_hash`; (c)
|
|
> recompute the Merkle root from the step hashes; (d) verify the Ed25519
|
|
> signature. A failure at any point is a failure — the asset is reported
|
|
> unverified, never "probably fine".
|
|
|
|
There is no tolerance parameter: the content hash matches or it does not.
|
|
|
|
## 5. Energy is measured, never invented
|
|
|
|
`joules_micro` per step is **carried** from a meter, exactly as the other
|
|
receipts carry it: `wai_prov` measures the real CPU energy of each encode
|
|
(IOReport via macmon, no sudo) and signs the figure. An unmetered step
|
|
carries `0`. The receipt never derives energy it did not measure.
|
|
|
|
## Appendix
|
|
|
|
The reconstruct classes prove WAI can ship a world, a sound, a body, a
|
|
volume as instructions and have every sink rebuild them identically.
|
|
Provenance proves the complement: for the photograph that is just a
|
|
photograph, WAI still adds something a bare file cannot — a tamper-evident
|
|
identity, a signed lineage, and an honest energy ledger — without ever
|
|
pretending to understand the pixels. Carries whatever it can reconstruct;
|
|
verifies whatever it carries.
|
|
|
|
[`asset_content_hash`]: ../wai-rs/src/world/mod.rs
|