Working draft · not stableSpecifications include planned features, not blanket implementation or proof claims.
SPEC-EV001 · 0.1.0-draft.5

Canonical Cairn text, rendered directly. Historical status statements below are preserved; consult current implementation scope and the status ledger separately. A shorter reading guide is on the guide.

On this page

Conformance and evidence records

Purpose

This accepted specification records Noble draft contracts, not completed implementation. Original requirement IDs, explanatory prose, examples, and open decisions remain authoritative. Scenario clauses declare designs; the conformance ledger records execution and evidence.

Requirements

Document: SPEC-EV001
Revision: 0.1.0-draft.5
Status: Record schema and document checks only

Independent dimensions

Requirement: EV-LIVE-01

r[EV-LIVE-01]

EV-LIVE-01. Live-reload designs MUST remain implementation: absent, execution: not-run, proof: open, trust: unassessed, with no test/evidence claim until the distinct opt-in binary emitter, selected VM and transactional reload actually run. Later receipts MUST bind both prior and proposed generation's exact source/import bytes, resolved definition and program-value identities, emitter/ABI/VM revision, validated module bytes, checked ordered stack/effect/capability interfaces, host policy, bounded execution and publication/refusal observation. An old proof or old source-bound receipt MUST NOT be inherited by a new body merely because its name, interface, source path or Wasm module format matches. Historical Core/M4/MC1/MC2/DXM1 evidence remains immutable and says nothing about live execution.

For guest self-edit, the receipt MUST separately identify the host-selected name, independently checked originating definition identity, grant's expected source generation, invocation-pinned generation observation, queued candidate's bounded byte snapshot and post-return snapshot comparison/independently checked recipe/interface, pending versus final commit/refusal, actual prior guest request prefix and stack outcome, derived-source origin/digest, and unchanged selected file bytes/hash. A denied import, trapped invocation and post-return candidate refusal are different outcomes; none is evidence that an unexecuted candidate body ran or that a completed invocation was rolled back.

Scenario: LIVE-06 for EV-LIVE-01

  • GIVEN the Live-Wasm-Draft profile and every field of input in LIVE-06
  • WHEN the runtime procedure for case LIVE-06 runs against those inputs
  • THEN the observations match every field of expected in case LIVE-06

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: LIVE-07 for EV-LIVE-01

  • GIVEN the Live-Wasm-Draft profile and every field of input in LIVE-07
  • WHEN the admission procedure for case LIVE-07 runs against those inputs
  • THEN the observations match every field of expected in case LIVE-07

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: LIVE-08 for EV-LIVE-01

  • GIVEN the Live-Wasm-Draft profile and every field of input in LIVE-08
  • WHEN the admission procedure for case LIVE-08 runs against those inputs
  • THEN the observations match every field of expected in case LIVE-08

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: LIVE-09 for EV-LIVE-01

  • GIVEN the Live-Wasm-Draft profile and every field of input in LIVE-09
  • WHEN the runtime procedure for case LIVE-09 runs against those inputs
  • THEN the observations match every field of expected in case LIVE-09

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: LIVE-10 for EV-LIVE-01

  • GIVEN the Live-Wasm-Draft profile and every field of input in LIVE-10
  • WHEN the runtime procedure for case LIVE-10 runs against those inputs
  • THEN the observations match every field of expected in case LIVE-10

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Requirement: EV-STATE-01

r[EV-STATE-01]

EV-STATE-01. Every scenario MUST carry independent implementation, execution, proof, and trust fields. One successful dimension MUST NOT imply success in another.

FieldAllowed values
implementationabsent, partial, implemented, unsupported
executionnot-run, passed, failed, timeout, unsupported
proofopen, accepted, failed, not-applicable
trustunassessed, explicit

Test design remains open for this requirement. No scenario or execution evidence is supplied.

Requirement: EV-STATE-02

r[EV-STATE-02]

EV-STATE-02. Execution or proof results other than unexecuted/open states MUST identify their evidence. Evidence binds a subject, revision, configuration, claim, assumptions, and result. implemented alone is not an executed result.

A failed test and an open proof can coexist with implemented code. A document-validator pass belongs to the document check lane, not to scenario execution.

Each evidence record MUST include a nonempty claim string that states the observed or proved claim. Non-review scenario results and component execution results MUST cite test evidence with the same execution outcome. Only scenarios whose kind is review can use review evidence for their execution field. Other evidence can accompany a result but cannot replace its required evidence class. Document validation checks these fields, not the truth of the claim.

The proof-obligation ledger MUST use the same evidence rules as scenario and component records. Each accepted or failed obligation MUST cite proof evidence with its own obligation ID and the same result. Review or test evidence alone MUST NOT establish a proof result. All supplied records MUST retain the required bindings, including records attached to open obligations.

SPEC-V002 adds claim outcomes: proved, disproved, unknown, timeout, unsupported, error, and not-run. These describe propositions and proof attempts, not the independent status fields in this document. A rejected proof does not establish a false proposition. Fixture expectations that name proved are not accepted-proof evidence.

Test design remains open for this requirement. No scenario or execution evidence is supplied.

Requirement: EV-STATE-03

r[EV-STATE-03]

EV-STATE-03. A greenfield snapshot MUST NOT report Noble runtime execution or accepted proofs without concrete implementation and evidence records. The current snapshot records no such results.

Scenario: WORKER-11 for EV-STATE-03

  • GIVEN the Worker-Design profile and every field of input in WORKER-11
  • WHEN the review procedure for case WORKER-11 runs against those inputs
  • THEN the observations match every field of expected in case WORKER-11

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario shape

Each file names schema_version, the specification revision, its origin, and a cases array. Each case has:

  1. A stable id, profile, and kind.
  2. Nonempty requirement references from the current family.
  3. An input object containing source or a named structured harness with parameters.
  4. An expected object naming the observation stage, outcome, and discriminating observations.
  5. A state object and an evidence array.

Requirement: EV-CASE-01

r[EV-CASE-01]

EV-CASE-01. Requirement references MUST resolve against the current revision. Unsupported or incomplete harness designs remain visible and unexecuted.

Kinds are static, runtime, admission, adapter, identity, and review. A release-policy review is not a runtime language test.

Contract scenarios use Contracts-Draft for the selected optional profile. Their logical predicates, symbolic artifact identities, and companion operations are harness inputs, not frozen Noble source syntax. Proof, runtime, and admission observations require their own executed evidence.

Backend-Experiment, Decoder-Experiment, Implementation-Policy, and Contract-Design label test-design lanes, not new language conformance profiles. A compile-fail design in a review record is not an executed Rust test.

I64 values in JSON use decimal strings inside typed records. This prevents a JavaScript JSON parser from losing integer precision. Expected stacks are bottom to top.

The calculator scenarios use Calculator-Design and AI-Authoring-Design as application/test lanes. Their expression strings are calculator input, not Noble source. Exact numerators, denominators, and conversion values use decimal strings. These records do not select portable Noble encodings. Calculator validation errors occur inside an accepted application and are not Noble static-rejection results.

Test design remains open for this requirement. No scenario or execution evidence is supplied.

Requirement: EV-CASE-02

r[EV-CASE-02]

EV-CASE-02. A static or admission rejection MUST observe zero candidate-body host requests and zero protected operations. Compiler-service effects, where applicable, are recorded separately.

Scenario: S-CASE-05 for EV-CASE-02

  • GIVEN the Core-Bootstrap profile and every field of input in S-CASE-05
  • WHEN the static procedure for case S-CASE-05 runs against those inputs
  • THEN the observations match every field of expected in case S-CASE-05

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Requirement: EV-CASE-03

r[EV-CASE-03]

EV-CASE-03.

Authority denial requires a declared effect that includes the request. A denied request counts toward the effect trace. Denial MUST NOT substitute for static effect rejection. To promote S-CASE-06, a new immutable source-bound prepromotion acceptance MUST show independent preparation/compilation of the host-selected complete WIT/world/export-source recipe and exact component-byte correspondence before preopening the file or establishing Directory/READ authority. It MUST show the matched actual "main.rs" fs.read Noble component invoking production Wasmtime once with the canonical live host-owned logical Directory, read_right:false, stage:authorization, outcome:denied, guest_requests:1, actual bounded callback request_trace:["fs.read"], protected_operations:0 and no protected file-content reads. A separately READ-granted run of the same compiled guest bytes MUST show exact bounded bytes read from the vetted host-preopened regular File and owner retirement, with hostile path/owner/rights refusals before any protected content read. A mismatched source/WIT recipe MUST refuse with status 2 at admission before file open/owner creation, and MUST NOT masquerade as the canonical one-request denial. An oversized granted file MUST instead be labeled a bounded postauthorization failure and MUST NOT be misreported as zero-read authorization denial. Source, pinned tools, WIT/guest/CLI and fixture identity, command outputs and assumptions MUST be bound; structural/document validation, static rejection, a synthetic callback, and a prior fixture receipt are not substitutes. Only after passing prepromotion acceptance MAY native deltas sync and canonical state/evidence/views change; postpromotion documents/Cairn checks are separate. Proof including owner law, Octet and Aeneas remains open.

Scenario: S-CASE-06 for EV-CASE-03

  • GIVEN the Resources-Draft profile and every field of input in S-CASE-06
  • WHEN the runtime procedure for case S-CASE-06 runs against those inputs
  • THEN the observations match every field of expected in case S-CASE-06

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: WI-08 for EV-CASE-03

  • GIVEN the Component-Sync-Bootstrap profile and every field of input in WI-08
  • WHEN the runtime procedure for case WI-08 runs against those inputs
  • THEN the observations match every field of expected in case WI-08

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: S-CASE-06 prepromotion runtime acceptance

  • GIVEN the exact unpromoted canonical S-CASE-06 record, source-bound compiled guest, host-preopened file and independent READ-or-none grants
  • WHEN the denied and granted invocations and hostile controls execute against the production host before any case promotion
  • THEN the denial matches every canonical expected field after one real request and zero protected reads; the granted control returns only exact vetted bounded bytes, owner obligations settle, and distinct prepromotion and postpromotion receipts remain bound to their respective source states

Requirement: EV-CASE-04

r[EV-CASE-04]

EV-CASE-04. Runtime builder cases MUST obtain operands after compilation and disable the Noble preparation service during execution. Optimized literal-only examples do not satisfy them.

Scenario: CORE-03 for EV-CASE-04

  • GIVEN the Core-Bootstrap profile and every field of input in CORE-03
  • WHEN the runtime procedure for case CORE-03 runs against those inputs
  • THEN the observations match every field of expected in case CORE-03

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Requirement: EV-CASE-05

r[EV-CASE-05]

EV-CASE-05. A matrix case MUST report every input combination and its expected outcome. Missing or unsupported elements MUST NOT count as passed coverage.

The language-workflow scenarios use design-lane labels, not new conformance profiles. Property-runner and documentation-harness tests retain their own subject and outcomes. A successful harness control does not mark its embedded Noble example as executed or proved.

The worker scenarios use Worker-Design, not a new language or concurrency profile. WORKER-CONFORMANCE.md separates pure worker requests, compiler-service effects, shell actions, and native retirement. Cancellation acknowledgement does not establish external completion. Round-trip expectations retain the canonical encoding gate.

Harness names and symbolic operation IDs describe future test drivers. They do not freeze source annotation syntax, canonical identifier bytes, or an executable test implementation. The validator checks record structure, not language behavior.

Test design remains open for this requirement. No scenario or execution evidence is supplied.

Requirement index

requirements.json is a derived index of every numbered requirement in the registered normative documents. It records document ownership, planned evidence routes, scenario links, and open test-design gaps.

Requirement: EV-TRACE-01

r[EV-TRACE-01]

EV-TRACE-01. Every numbered requirement MUST have a declared evidence route. A requirement with no concrete scenario remains test-design-open, not covered or passed.

Native requirement headings and their matching bracketed markers own the index. Legacy labels support historical preservation checks only. Fenced examples MUST NOT create requirements, scenario targets, or preservation obligations.

The index uses these planned routes:

FamiliesPlanned evidence
K, P, Bsemantic tests, reference-model proof, review
BEexperiment tests, correspondence evidence, review
W, H, S, WI, RAboundary tests, correspondence evidence, review
C, EVvalidation/release-policy tests and review
DXsemantic tests, policy tests, and review
CALCapplication tests, benchmark-policy tests, and review
V, VT, VCproof or scoped assumption review, with implementation tests where relevant

Routes do not prove completion. Exact theorem links and execution observations belong in evidence records as implementation proceeds.

The Aeneas-first scope rules in VT-SCOPE-01 through VT-SCOPE-05 additionally require a production-source inventory. The requirement index does not replace function-level extraction and refinement coverage. Parser, backend, resource, runtime, and CLI work must remain visible even before implementation.

Test design remains open for this requirement. No scenario or execution evidence is supplied.

Requirement: EV-TRACE-02

r[EV-TRACE-02]

EV-TRACE-02. Regenerating the index MUST NOT manufacture test results or overwrite observed evidence. The index contains only derived planning metadata. Scenario states and the obligation ledger remain separate files.

Test design remains open for this requirement. No scenario or execution evidence is supplied.

Bound evidence and imported tool results

Requirement: EV-BIND-01

r[EV-BIND-01]

EV-BIND-01. A reusable evidence record MUST bind its subject, claim, source scope, semantic dependencies, policy, toolchain, target/features, command, inputs, assumptions, and result. Digests and versioned references can identify those inputs. Missing relevant bindings or unknown coverage MUST block reuse for a required claim.

Consumers select the expected subject, claim, policy, and trust context independently of producer-supplied evidence. A changed binding requires renewed applicability validation. Old evidence can remain valid for its historical subject without satisfying the current gate. A producer cannot establish trust by replacing both an artifact and its unkeyed digest.

Scenario: CONTRACT-16 for EV-BIND-01

  • GIVEN the Contracts-Draft profile and every field of input in CONTRACT-16
  • WHEN the admission procedure for case CONTRACT-16 runs against those inputs
  • THEN the observations match every field of expected in case CONTRACT-16

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: DX-13 for EV-BIND-01

  • GIVEN the Property-Test-Design profile and every field of input in DX-13
  • WHEN the review procedure for case DX-13 runs against those inputs
  • THEN the observations match every field of expected in case DX-13

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: OCTET-09 for EV-BIND-01

  • GIVEN the Octet-Adoption-Design profile and every field of input in OCTET-09
  • WHEN the review procedure for case OCTET-09 runs against those inputs
  • THEN the observations match every field of expected in case OCTET-09

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Requirement: EV-TIER-01

r[EV-TIER-01]

EV-TIER-01. Imported Octet results MUST preserve their lint, policy, transcript, integrity, or formal-proof role without promotion. Formal-proof claims require the existing verifier-backed evidence and exact subject correspondence. Fixture-scoped extraction, source-shape conformance, and artifact verification MUST NOT establish whole-kernel correctness or release eligibility.

The record MUST retain each required phase outcome and each incomplete scope. An omitted or failed phase cannot become an overall acceptance through a successful sibling phase. A type-valid configuration is not executed policy enforcement. These rules extend existing evidence bindings without making every receipt a proof or changing ordinary program admission.

Octet adoption cases contain explicit stale-binding, tier-promotion, and incomplete-coverage controls. These are unexecuted designs. The document validator does not invoke Octet or a proof tool.

Scenario: OCTET-09 for EV-TIER-01

  • GIVEN the Octet-Adoption-Design profile and every field of input in OCTET-09
  • WHEN the review procedure for case OCTET-09 runs against those inputs
  • THEN the observations match every field of expected in case OCTET-09

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: OCTET-10 for EV-TIER-01

  • GIVEN the Octet-Adoption-Design profile and every field of input in OCTET-10
  • WHEN the review procedure for case OCTET-10 runs against those inputs
  • THEN the observations match every field of expected in case OCTET-10

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Release claims

Requirement: EV-GATE-01

r[EV-GATE-01]

EV-GATE-01. A release claim MUST identify its supported subset and all applicable gates. Passing document validation cannot satisfy compiler, resource, component, or safety conformance.

Proof obligations in verification/obligations.json remain open. They name expected claims, not completed Lean declarations.

Scenario: ADAPT-12 for EV-GATE-01

  • GIVEN the Implementation-Policy profile and every field of input in ADAPT-12
  • WHEN the review procedure for case ADAPT-12 runs against those inputs
  • THEN the observations match every field of expected in case ADAPT-12

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: WORKER-11 for EV-GATE-01

  • GIVEN the Worker-Design profile and every field of input in WORKER-11
  • WHEN the review procedure for case WORKER-11 runs against those inputs
  • THEN the observations match every field of expected in case WORKER-11

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Requirement: EV-GATE-02

r[EV-GATE-02]

EV-GATE-02. A proof accepted under a narrower subset or different configuration MUST NOT silently close a broader obligation. Trust assumptions and cross-tool boundaries remain explicit.

Scenario: ADAPT-13 for EV-GATE-02

  • GIVEN the Implementation-Policy profile and every field of input in ADAPT-13
  • WHEN the review procedure for case ADAPT-13 runs against those inputs
  • THEN the observations match every field of expected in case ADAPT-13

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Document checks

bun tools/check-specs.mjs
bun tools/check-specs.mjs --self-test

After intentional requirement or scenario changes:

bun tools/check-specs.mjs --refresh-ledger
bun tools/check-specs.mjs

The refresh command changes only derived requirement metadata. It does not change scenario states or proof status. Selected oracle checks reject known specification regressions; they do not implement the referenced harnesses.

Requirement: EV-LSLOT-01

r[EV-LSLOT-01]

EV-LSLOT-01.

Every LSLOT scenario MUST remain implementation: absent, execution: not-run, proof: open, trust: unassessed, evidence: [] until a separate implementation and actual execution/proof are independently recorded. Later evidence MUST bind the generic caller identity and exact selected target ProgramValueId/DefinitionId, captures, context, claim, assumptions, admitted artifact, pinned slot-registry map/global epoch (distinct from Live-Wasm-Draft namespace/source generation), slot incarnation/generation, publication or refusal CAS inputs/outcome, current host policy decisions, nested version selections and actual effect requests/outcomes. Replay evidence MUST also bind exact host-issued typed root inputs and every ordered host request/response/accounting event, and MUST refuse wrong epoch/incarnation/generation, substituted artifact/captures, context or trace order/accounting even with a frozen admitted map. A design gate, wrapper proof, historical LIVE receipt or compatible interface MUST NOT count as live-target execution, replay, behavioral proof or acceptance. Existing LIVE-01..10 and accepted milestone receipts MUST remain unchanged.

Scenario: LSLOT-02 for EV-LSLOT-01

  • GIVEN the Live-Slot-Design profile and every field of input in LSLOT-02
  • WHEN the runtime procedure for case LSLOT-02 runs against those inputs
  • THEN the observations match every field of expected in case LSLOT-02

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: LSLOT-04 for EV-LSLOT-01

  • GIVEN the Live-Slot-Design profile and every field of input in LSLOT-04
  • WHEN the admission procedure for case LSLOT-04 runs against those inputs
  • THEN the observations match every field of expected in case LSLOT-04

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: LSLOT-05 for EV-LSLOT-01

  • GIVEN the Live-Slot-Design profile and every field of input in LSLOT-05
  • WHEN the admission procedure for case LSLOT-05 runs against those inputs
  • THEN the observations match every field of expected in case LSLOT-05

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: LSLOT-07 for EV-LSLOT-01

  • GIVEN the Live-Slot-Design profile and every field of input in LSLOT-07
  • WHEN the runtime procedure for case LSLOT-07 runs against those inputs
  • THEN the observations match every field of expected in case LSLOT-07

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: LSLOT-08 for EV-LSLOT-01

  • GIVEN the Live-Slot-Design profile and every field of input in LSLOT-08
  • WHEN the runtime procedure for case LSLOT-08 runs against those inputs
  • THEN the observations match every field of expected in case LSLOT-08

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: LSLOT-09 for EV-LSLOT-01

  • GIVEN the Live-Slot-Design profile and every field of input in LSLOT-09
  • WHEN the static procedure for case LSLOT-09 runs against those inputs
  • THEN the observations match every field of expected in case LSLOT-09

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Requirement: NGRAPH-SNAP-01

r[NGRAPH-SNAP-01]

NGRAPH-SNAP-01.

An optimizer step MUST produce a new immutable parameter snapshot and MUST NOT modify a parent snapshot, graph, Program or capture. Snapshot identity MUST be domain-separated and MUST bind the graph logical identity, the ordered parameter layout and the exact little-endian parameter bytes. Training provenance MUST bind the parent snapshot, the source subject and closure identities, the profile semantic revision, the derivative policy and its per-node options, the fact bindings, the loss graph, the exact ordered training data digest, the seed and PRNG revision, the optimizer revision with exact hyperparameter bits, the step count and the tool revisions. Evaluating with a snapshot whose graph or layout identity does not match MUST be refused before evaluation. A hash identifies bytes; it MUST NOT be reported as a correctness claim. Capturing a snapshot into a new Program MUST create a new subject whose evidence is checked for applicability anew; proofs MUST NOT transfer automatically.

Scenario: NGRAPH-CASE-10 for NGRAPH-SNAP-01

  • GIVEN the Numeric-Graph-v1 profile and every field of input in NGRAPH-CASE-10
  • WHEN the runtime procedure for case NGRAPH-CASE-10 runs against those inputs
  • THEN the observations match every field of expected in case NGRAPH-CASE-10

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: NGRAPH-CASE-12 for NGRAPH-SNAP-01

  • GIVEN the Numeric-Graph-v1 profile and every field of input in NGRAPH-CASE-12
  • WHEN the admission procedure for case NGRAPH-CASE-12 runs against those inputs
  • THEN the observations match every field of expected in case NGRAPH-CASE-12

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: NGRAPH-CASE-13 for NGRAPH-SNAP-01

  • GIVEN the Numeric-Graph-v1 profile and every field of input in NGRAPH-CASE-13
  • WHEN the identity procedure for case NGRAPH-CASE-13 runs against those inputs
  • THEN the observations match every field of expected in case NGRAPH-CASE-13

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: NGRAPH-CASE-14 for NGRAPH-SNAP-01

  • GIVEN the Numeric-Graph-v1 profile and every field of input in NGRAPH-CASE-14
  • WHEN the identity procedure for case NGRAPH-CASE-14 runs against those inputs
  • THEN the observations match every field of expected in case NGRAPH-CASE-14

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Requirement: NGRAPH-SNAP-02

r[NGRAPH-SNAP-02]

NGRAPH-SNAP-02.

Contents. A snapshot file MUST begin with noble-ngraph-snapshot/0. It MUST record, as structured fields:

  • the ParamSnapshotId, the GraphLogicalId and the ordered parameter layout;
  • every parameter's exact binary32 encoding, in param-slot order;
  • the optimizer-state digest;
  • the complete NGRAPH-SNAP-01 training provenance, including the fact-binding digest of the model graph;
  • a domain-separated digest of that provenance record.

Loading. Loading a snapshot file MUST recompute the layout, the snapshot identity and the provenance digest from the recorded fields, and MUST refuse any mismatch with E-SNAPSHOT-INTEGRITY before use. The provenance digest MUST be the domain-separated identity encoding (a domain tag, then length-prefixed byte strings and fixed-width little-endian integers) of every record in file order except the schema line and the digest itself, so that any reader can recompute it. Records that training can never publish, a non-finite parameter or optimizer hyperparameters outside the optimizer's domain, MUST also be refused with E-SNAPSHOT-INTEGRITY, however consistent their digests.

Binding. Binding a snapshot to a graph MUST refuse:

  • a different GraphLogicalId, or a different fact-binding digest, with E-SNAPSHOT-STALE;
  • a different layout with E-SNAPSHOT-LAYOUT;
  • a snapshot other than a caller-pinned snapshot identity with E-SNAPSHOT-ID.

Every binding refusal MUST occur before any graph node is evaluated.

Publication. Training MUST publish its result only as a new file at a path that did not exist. The write MUST never truncate, replace or modify an existing file, including the parent snapshot. A training call that stops MUST NOT leave a snapshot or partial file at the requested path. Once the new file exists with its complete bytes, the call MUST report the publication as done, even if cleaning up a staging name fails.

Replay.

  • Replaying a snapshot's provenance MUST first require that the supplied model, loss graph, training data and parent match the recorded identities; otherwise E-SNAPSHOT-STALE.
  • Replay MUST admit the recorded training contract exactly as training admits a new one, so a recorded derivative opt-in that training would refuse MUST be refused with the code training gives it.
  • On the same pinned platform, replay MUST then reproduce the parameter bytes, the optimizer-state digest and the snapshot identity exactly; otherwise E-SNAPSHOT-REPLAY.

Non-claim. A snapshot file identifies bytes. A matching identity or a successful replay MUST NOT be reported as evidence that the trained parameters, the optimizer or the evaluators are correct.

Scenario: NGRAPH-CASE-10 for NGRAPH-SNAP-02

  • GIVEN the Numeric-Graph-v1 profile and every field of input in NGRAPH-CASE-10
  • WHEN the runtime procedure for case NGRAPH-CASE-10 runs against those inputs
  • THEN the observations match every field of expected in case NGRAPH-CASE-10

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: NGRAPH-CASE-11 for NGRAPH-SNAP-02

  • GIVEN the Numeric-Graph-v1 profile and every field of input in NGRAPH-CASE-11
  • WHEN the admission procedure for case NGRAPH-CASE-11 runs against those inputs
  • THEN the observations match every field of expected in case NGRAPH-CASE-11

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: NGRAPH-CASE-12 for NGRAPH-SNAP-02

  • GIVEN the Numeric-Graph-v1 profile and every field of input in NGRAPH-CASE-12
  • WHEN the admission procedure for case NGRAPH-CASE-12 runs against those inputs
  • THEN the observations match every field of expected in case NGRAPH-CASE-12

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: NGRAPH-CASE-13 for NGRAPH-SNAP-02

  • GIVEN the Numeric-Graph-v1 profile and every field of input in NGRAPH-CASE-13
  • WHEN the identity procedure for case NGRAPH-CASE-13 runs against those inputs
  • THEN the observations match every field of expected in case NGRAPH-CASE-13

This is a scenario design, not an execution result. The case's state and evidence fields record its status.

Scenario: NGRAPH-CASE-10 immutable training publication

  • GIVEN the Numeric-Graph-v1 profile and the model, loss, data, seed and follow-up runs specified by NGRAPH-CASE-10
  • WHEN noble ngraph train publishes snapshots, retrains to a second path, trains from the first snapshot and targets an existing path
  • THEN the recorded identities MUST match, repeated training MUST be byte-identical, and the parent, the graph and existing paths MUST stay byte-unchanged

Scenario: NGRAPH-CASE-12 snapshot integrity and binding refusals

  • GIVEN every tampered, foreign, permuted, changed-fact, pinned-identity and schema variant specified by NGRAPH-CASE-12
  • WHEN noble ngraph eval and noble ngraph verify load and bind each snapshot
  • THEN each MUST be refused with its recorded code before any graph node is evaluated

Scenario: NGRAPH-CASE-13 provenance replay

  • GIVEN the seed, parent, changed-input and forged replay variants specified by NGRAPH-CASE-13
  • WHEN noble ngraph verify --replay rebuilds and reruns each training contract
  • THEN matching inputs MUST reproduce the snapshot exactly and every other variant MUST be refused with its recorded code

Scenario: NGRAPH-CASE-11 training contract refusals

  • GIVEN every derivative, sub-profile, non-finite, budget, parent, guard, loss-interface and usage variant specified by NGRAPH-CASE-11
  • WHEN noble ngraph train runs each variant
  • THEN each MUST report its recorded code and exit status and MUST NOT create an output or partial file, except the recorded opt-in success