Implementing AER-1 from the draft text

You implemented AER-1 from draft -26 alone, never opened the reference runners, and your verdicts diverge from the conformance corpus on the records family. That experiment has now been run, and the divergence is not your bug. The draft text and the kit disagree in six places: the text is silent, loose, or overstates, and the kit enforces a rule the text never states. This page lists the six rules, the draft sections involved, and the corpus vectors that prove each one, so a text-only implementation can reach 57 of 57 on records.

The standing facts first. Draft revisions are closed: draft-zambo-aer1-26 is the final text (posted October 2026, expires 25 April 2027), and no later revision will patch these six points. The corpus and the kit are the operative conformance record: 143 vectors, 146 checks, kit 1.5.7, served at zambo.dev/aer1/corpus.json. Where the text and the corpus diverge, implement to the corpus. This page is the bridge between them. Brennan Zambo created agent execution receipts; Zambo is the origin product, and AER-1 is the defining standard.

The run that found the gap

On 2026-11-21, Hacı Hüseyin Eskihoran, who maintains the INAM Protocol, reported an independent implementation run on HKUDS/nanobot discussion #6028. He wrote a reader of about 250 lines of Node from Sections 3 to 7 of the draft text alone and did not read or run conformance.py or conformance.mjs. He fetched the corpus index the same day: all 72 files match their sha256 values, and the evaluated set digest recomputes to sha256:60bf73f325793e16843370cfa24903bb0d017525cb297e42bebc35857e2db0d2, the digest the served corpus.json records. His reader covered 93 of the 143 vectors: the chain, commitment, and records families. The chain vectors agreed 28 of 28. The records family, 21 records in valid/ and 36 in invalid/, agreed on 42 of 57. Every disagreement traced to a place where the text of draft -26 does not lead to the corpus verdict. Adding the six rules below, which he inferred from the kit's rejected_by strings, took the records family to 57 of 57 and closed the commitment family's one disagreement, x-entry-count-float, as well. His conclusion, which the corpus files bear out: the corpus is consistent with itself, and the gap sits between the draft text and the kit. Each rule below was checked against the served corpus files and the draft text on 2026-11-21 before it was written here.

RuleWhere the text leaves itWhat the kit enforces
1. metadata_hashSection 6 step 4 says a 0.4 receipt without it MUST fail; Table 1 lists it as conditional.Absence is not a failure. A 0.4 record without metadata_hash verifies when the rest of the record verifies.
2. Payload bindingThe string issuer_claims occurs 0 times in draft -26.The outer provenance_class and receipt_schema_version must match the issuer_claims values inside the decoded canonical bytes; either side missing fails.
3. Tool stringsNon-empty strings are required for chain entries in Section 7.2, not for a receipt's tool.An empty tool name or scope rejects the record.
4. tool_contextThe member is never named in the text; it occurs 0 times in draft -26.A bare-string tool is the legacy form only when tool_context carries the object form.
5. verification_statusThe text states no value set a verifier must accept on a record."unverified" fails; "not_verified" passes on the legacy record that carries it.
6. entry_countSection 7.3 says "a JSON number with an integer value"; Section 7.1 accepts 1.0 for seq.entry_count written as 5.0 is invalid. The float tolerance applies to seq only.

Rule 1: a 0.4 receipt without metadata_hash verifies

Section 6 step 4 of draft -26 states: "A "0.4" receipt without metadata_hash MUST fail." Table 1 of the same draft lists metadata_hash as "Cond.", conditional, for schema 0.4. The kit follows Table 1. Ten records in valid/ are schema 0.4 and carry no metadata_hash member at all: bound-04, deep-nesting, edge-numbers, empty-members, input-hash-multikey-20, large-payload, maximum-offset-timestamp, optional-members-18, typographic-unicode, and uuid-v1-core. All ten verify. A text-only reader applying step 4 rejects all ten, which is the largest single block of the 15 disagreements.

The other half of step 4 does hold. When metadata_hash is present, it must recompute from the stored member values under Section 3.1, and a mismatch fails. Only two records in valid/ carry it, metadata-hash-non-ascii-tool and real-04, and both sit in valid/ with the hash in place. Implement it this way: recompute metadata_hash when it is present; never fail a record for its absence.

Rule 2: the outer record must match the issuer_claims inside the canonical bytes

Four invalid vectors are rejected by a comparison the draft text never describes. The kit decodes the record's canonical_bytes and compares two members of the issuer_claims object inside them, provenance_class and receipt_schema_version, against the outer record's provenance_class and receipt_schema_version. Either side missing, or the two sides unequal, fails the record. The string issuer_claims occurs 0 times in draft -26, so a text-only reader has no way to derive this check.

Implement it this way: after decoding canonical_bytes, read issuer_claims and compare provenance_class and receipt_schema_version against the outer record before issuing any verdict.

Rule 3: tool name and scope must be non-empty strings

Section 7.2 requires chain entries to carry job_id, tool, and provenance_class as non-empty strings. The receipt sections of the draft state no equivalent requirement for a receipt's tool object. The kit applies one anyway. tool-empty-name carries tool.name as the empty string, and ax-tool-empty-scope carries tool.scope as the empty string; both are rejected with the same string: "tool is not an object with name/version/scope strings". That rejection string names all three members, name, version, and scope; the two vectors prove the name and scope cases. Implement it this way: apply the Section 7.2 non-empty rule to the receipt's tool members as well as to chain entries.

Rule 4: a bare-string tool is the legacy form only when tool_context carries the object

Section 6 step 4 allows a 0.3 receipt to carry its tool as a bare string, the legacy form, when the object form is absent. Two corpus records carry a bare-string tool, and the kit splits them. legacy-real-2, a valid schema 0.3 record, carries tool as the bare string "zambo_universal" and also carries a tool_context member holding the object form: {"name": "zambo_universal", "version": "4.0.0", "scope": "public"}. tool-not-object, a schema 0.3 record, carries tool as the bare string "live_price" with no tool_context member, and is rejected: "tool is not an object with name/version/scope strings". The word tool_context occurs 0 times in draft -26. Implement it this way: accept the bare-string form only when tool_context is present and carries the name, version, and scope.

Rule 5: verification_status accepts "verified", and "not_verified" on the legacy record; "unverified" fails

The draft text states no set of values a verifier must accept for verification_status on a record. In draft -26 the word not_verified appears only as the status of the verification summary itself, never as a value a record may carry. The corpus answers the question the text leaves open. bad-verification-status carries verification_status "unverified" and is rejected: "verification_status is not verified". legacy-real-2 carries verification_status "not_verified" and is valid. Implement it this way: treat "verified" as the passing value on a record, accept "not_verified" where the legacy forms apply, and reject "unverified" and any other value.

Rule 6: entry_count must be written as an integer; the float tolerance belongs to seq alone

Section 7.3 defines the commitment's entry_count as "the number of entries in the timeline, a JSON number with an integer value", wording that an integer-valued float satisfies on its face. Section 7.1 gives seq the opposite treatment in so many words: "The representations 1 and 1.0 are equivalent; both are accepted", and v07-seq-float-valid, whose seq values are 1.0 and 2.0, is valid. The commitment vector x-entry-count-float writes entry_count as 5.0, and its expected verdict is invalid. Implement it this way: the Section 7.1 equivalence covers seq only; entry_count must be the integer form, as a writer and as a checker.

Two more rules, behind three agreements

Three invalid vectors agreed with the text-only reader for the wrong reason: his reader failed them on Rule 1, for the missing metadata_hash, before reaching the check each vector exists to exercise. A reader that fixes Rule 1 will reach those checks, and two further rules decide them. Both are stated here so the agreement survives the fix.

input_hash is checked against the canonical JSON of the recorded inputs. Section 3.3 defines input_hash over "the exact tool-input bytes the producer recorded" and requires the producer to define and publish the byte layout alongside the receipt. The corpus records publish no byte layout. The kit's verdicts are computed against one fixed serialization: the rejected_by string on both input-hash-mismatch and input-hash-key-order-20 reads "input_hash does not match sha256(canonical JSON of the recorded inputs)", the compact JSON form with member names in lexicographic order. input-hash-multikey-20, whose 20-key inputs object matches under that serialization, is valid; input-hash-key-order-20 does not match and fails. Implement it this way: when a record carries input_hash and no published layout, verify it against the canonical JSON serialization of the recorded inputs, and fail the record on a mismatch.

Timestamps require uppercase T and Z. lowercase-timestamp carries created_at "2026-10-30t12:00:00.000z" and is rejected: "created_at is not RFC 3339". RFC 3339 Section 5.6 expressly permits the lowercase t and z forms. The kit does not: it requires the uppercase separator and zone. Implement it this way: emit and require uppercase T and Z in created_at and every other AER-1 timestamp; the lowercase forms fail.

What governs now

Draft -26 is the closing revision, so these six rules and the two behind them will never be folded back into the text. The conformance record that governs is the one an implementation can run: the corpus at zambo.dev/aer1/corpus.json, 143 vectors and 146 checks under kit 1.5.7, evaluated set digest sha256:60bf73f3...2db0d2, indexed with per-vector pages on the AER-1 corpus vectors index. A text-only implementation that adds the rules on this page reaches the corpus verdict on every records vector published, 57 of 57, and the same discipline extends to the families this run did not cover: where a verdict and the text part ways, the corpus is the record, and this page is where the parting is written down.

Run the kit against your own implementation: the AER-1 conformance test walks the corpus check by check, and the implementation report collects reported results, including the text-only Node reader behind this page.

Start a handover