docs: make benchmark validation reproducible

README-only: explicit train split, exact seven-case validation matrix pinned to immutable fixture/verifier commits, and fail-closed scope/limitations. No fixture or license changes.

README.md

@@ -20,7 +20,9 @@ task_categories:
   - other
 configs:
   - config_name: default
-    data_files: "bench.jsonl"
+    data_files:
+      - split: train
+        path: bench.jsonl
 ---
 
 <div align="center">
@@ -37,18 +39,23 @@ configs:
 
 A small **benchmark corpus of governance decision receipts** for the open
 [`governed-receipt-spec`](https://github.com/szl-holdings/governed-receipt-spec).
-Every receipt here has a known expected outcome under the spec's **dependency-free
-offline verifier**: the `valid/` receipts must **PASS**, the `invalid/` receipts
-must **FAIL**. Use it to test a verifier implementation, to see what a real SZL
-receipt looks like, or to check your own receipts against the same rules.
+`bench.jsonl` declares one expected outcome for each case under the spec's
+**dependency-free offline verifier**. With the fixture bytes at dataset commit
+[`2887f9a`](https://huggingface.co/datasets/SZLHOLDINGS/governed-receipts-bench/commit/2887f9af1137bef789edf2e9eaad9177a1d6d27e)
+and verifier commit
+[`007106b`](https://github.com/szl-holdings/governed-receipt-spec/commit/007106bcb0212138345e19eb5efba8bb69327d57),
+the five `valid/` rows return **PASS** and the two `invalid/` rows return
+**FAIL**. These are conformance labels for the checks that verifier performs,
+not proof that every recorded statement is true or that signer identity was
+cryptographically established.
 
 Built and maintained by [SZL Holdings](https://a-11-oy.com). Apache-2.0.
 
-> **Honesty first.** The `valid/` receipts are **real records** drawn factually
-> from SZL's public receipt datasets — nothing is fabricated. The `invalid/`
-> receipts are **clearly-labeled tampered derivatives** of a real receipt,
-> included only so the verifier's failure paths are testable. No metric here is
-> invented; where a value was not measured (e.g. energy joules) it is honestly
+> **Fixture provenance.** The manifest identifies the public datasets from which
+> the `valid/` files are declared to originate. This benchmark checks the bytes
+> supplied here; it does not independently prove those provenance statements or
+> the real-world truth of payload claims. The `invalid/` files are explicitly
+> labeled mutations. Where energy was not measured, the fixture records
 > `null` / `UNAVAILABLE`.
 
 ---
@@ -58,12 +65,14 @@ Built and maintained by [SZL Holdings](https://a-11-oy.com). Apache-2.0.
 A **governed inference receipt** is the small, replayable, hash-chained record a
 governed AI runtime emits for one governed action (e.g. an inference): what it
 **decided**, the **Λ** governance-floor status, whether **energy** was actually
-measured, and a signed **DSSE** envelope that lets anyone re-check it offline.
+measured, and a **DSSE-style** envelope that lets anyone re-check supported structural,
+hash, schema, and chain relations offline.
 
 A receipt is explicitly **NOT** a zero-knowledge proof and **NOT** a proof of
-computation — it is a signed, tamper-evident audit record of a governance
-*decision* and its bound content hashes. That is the "receipt tier" of trust
-(cheaper than TEE or zkML; deployable today). See the spec for the full framing.
+computation. It is an audit-record format for a governance decision and any
+byte-binding or chain relations that are actually present and checked. Some
+fixtures here are unsigned or lack a checked content-hash field, so their trust
+strength is limited to the checks recorded below. See the spec for the full framing.
 
 ---
 
@@ -71,24 +80,64 @@ computation — it is a signed, tamper-evident audit record of a governance
 
 ```
 bench.jsonl          # one row per case: file, category, expected_result, source, note
-valid/               # real receipts — MUST verify PASS
+valid/               # manifest expectation: PASS under the pinned verifier
   a11oy-khipu-chain.json      # 5-receipt signed hash chain (seq 0->4)
   a11oy-khipu-genesis.json    # single signed genesis receipt (seq 0, prev = 64 zeros)
   lake-inference-receipt.json # decision=allow + measured-or-null energy in the clear
   readiness-audit-receipt.json# unsigned envelope, content bound by payloadSha256
   daily-activity-receipt.json # HMAC-stub daily activity receipt
-invalid/             # LABELED tampered derivatives — MUST verify FAIL
+invalid/             # labeled mutations; manifest expectation: FAIL
   tampered-payload.json       # a payload byte flipped -> DSSE PAE content-hash mismatch
   broken-chain.json           # prev rewritten -> hash-chain link breaks
 ```
 
 `bench.jsonl` is the machine-readable manifest (the dataset viewer renders it):
 each row carries `file`, `category` (`valid`/`invalid`), `expected_result`
-(`PASS`/`FAIL`), the `source_dataset` it was drawn from, `n_receipts`, and a `note`.
+(`PASS`/`FAIL`), the manifest-declared `source_dataset`, `n_receipts`, and a
+`note`. The manifest has **7 unique paths**, **5 PASS** labels, **2 FAIL** labels,
+and **19 receipt objects** across the referenced files.
+
+## Load the manifest directly
+
+The `default` configuration has one explicit `train` split and needs no custom
+dataset loading code:
+
+```python
+from collections import Counter
+from datasets import load_dataset
+
+bench = load_dataset(
+    "SZLHOLDINGS/governed-receipts-bench",
+    name="default",
+    split="train",
+)
+assert len(bench) == 7
+assert Counter(bench["expected_result"]) == {"PASS": 5, "FAIL": 2}
+```
+
+Loading the manifest does not itself verify the referenced fixture files. Run
+the pinned verifier below for that step.
+
+## Exact case matrix
+
+The following results were reproduced against the pinned dataset and verifier
+commits linked above. The manifest and verifier do **not** expose stable,
+machine-readable failure codes, so none are invented here; the last column
+records the verifier's supported check or prose failure reason.
+
+| File | Objects | Expected | Reproduced | Supported result / reason |
+| --- | ---: | ---: | ---: | --- |
+| `valid/a11oy-khipu-chain.json` | 5 | PASS | PASS | 5 PAE hashes and 5 schema checks pass; chain intact across 5 receipts |
+| `valid/a11oy-khipu-genesis.json` | 1 | PASS | PASS | PAE hash and schema pass; one-receipt genesis chain intact |
+| `valid/lake-inference-receipt.json` | 1 | PASS | PASS | PAE hash and schema pass; one-receipt chain intact |
+| `valid/readiness-audit-receipt.json` | 1 | PASS | PASS | `payloadSha256` passes; envelope is unsigned with 0 signatures; schema skipped; chain not applicable |
+| `valid/daily-activity-receipt.json` | 1 | PASS | PASS | Structural envelope check only; unsigned with 1 signature entry; no content-hash field; schema skipped; chain not applicable |
+| `invalid/tampered-payload.json` | 5 | FAIL | FAIL | Receipt 2 payload is invalid JSON and its recomputed PAE SHA-256 mismatches `_pae_sha256`; chain then also fails |
+| `invalid/broken-chain.json` | 5 | FAIL | FAIL | Receipt at `seq=3` has `prev=deadbeef...`, which does not equal the `seq=2` digest |
 
 ---
 
-## Provenance of the `valid/` receipts (real data)
+## Manifest-declared provenance of the `valid/` fixtures
 
 | File | Source dataset | Shows |
 | --- | --- | --- |
@@ -98,13 +147,15 @@ each row carries `file`, `category` (`valid`/`invalid`), `expected_result`
 | `valid/readiness-audit-receipt.json` | [`SZLHOLDINGS/readiness-runs`](https://huggingface.co/datasets/SZLHOLDINGS/readiness-runs) | unsigned envelope with `payloadSha256` |
 | `valid/daily-activity-receipt.json` | `SZLHOLDINGS/szl-evidence` | HMAC-stub daily activity receipt |
 
-The two `invalid/` files are byte-level tampers of `a11oy-khipu-chain.json`,
-included purely to exercise the verifier's failure detection — they are **not**
-real records.
+The two `invalid/` files are byte-level mutations of
+`a11oy-khipu-chain.json`, included purely to exercise two failure paths. The
+`tampered-payload` file differs by one byte. The `broken-chain` file rewrites the
+`seq=3` `prev` value and recomputes the PAE-bound identifiers, leaving the chain
+relation invalid.
 
 ---
 
-## How to verify (one command, zero dependencies)
+## Verify with the standard-library reference verifier
 
 The verifier is Python 3 standard library only — no third-party packages, no
 network.
@@ -115,8 +166,9 @@ git clone https://huggingface.co/datasets/SZLHOLDINGS/governed-receipts-bench
 # 2. get the spec + verifier
 git clone https://github.com/szl-holdings/governed-receipt-spec
 cd governed-receipt-spec
+git checkout 007106bcb0212138345e19eb5efba8bb69327d57
 
-# 3. the real receipts PASS ...
+# 3. the manifest-labeled valid fixtures PASS ...
 python verify.py ../governed-receipts-bench/valid/*.json          # -> OVERALL: PASS
 
 # 4. ... and the tampered ones FAIL
@@ -127,11 +179,30 @@ python verify.py ../governed-receipts-bench/invalid/broken-chain.json      # ->
 Or paste any receipt into the live verifier Space:
 **[`SZLHOLDINGS/governed-receipt-verifier`](https://huggingface.co/spaces/SZLHOLDINGS/governed-receipt-verifier)**.
 
-For each receipt the verifier: (1) validates the decoded decision object against
-the JSON Schema, (2) recomputes the content hash (`sha256(DSSE PAE) == _pae_sha256`
-for signed receipts, or `sha256(payload) == payloadSha256` for readiness
-receipts), (3) checks the `prev`-hash chain (`prev == previous.digest`, contiguous
-`seq`, genesis is 64 zeros), and (4) structurally checks the DSSE envelope.
+At the pinned verifier commit, inference decision objects are checked against
+the JSON Schema; non-inference objects are skipped. The verifier recomputes a
+content hash only when `_pae_sha256` or `payloadSha256` is present, checks
+`prev`/`digest` and contiguous `seq` relations for chainable records, and checks
+envelope structure. It does **not** cryptographically verify signature bytes,
+public-key ownership, or signer identity, and it does not re-derive the runtime's
+internal `digest`.
+
+## Fail-closed interpretation and limitations
+
+- Treat **PASS** as success only for checks the verifier actually ran. `SKIP` or
+  `n/a` is not evidence that the omitted property holds.
+- Treat malformed input, an unknown schema/payload type/algorithm/key, a missing
+  required verifier resource, or a verifier exception as **FAIL** or
+  **UNAVAILABLE**; never as PASS. Integrations may need stricter policy than the
+  reference verifier.
+- A structurally present signature entry is not a cryptographically verified
+  signature. In particular, the daily-activity fixture has no independently
+  checked content hash at this verifier revision.
+- A receipt can bind bytes without proving that payload claims are factually
+  correct, that computation occurred, or that a decision was safe.
+- The two negative fixtures cover only a PAE/payload mutation and a broken chain
+  link. They are not an exhaustive adversarial suite, a security certification,
+  or a comparison of verifier implementations.
 
 ---