Verification and Trust Model
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-V001
Revision: 0.1.0-draft.5
Project: noble
Status: Canonical working verification contract; no implementation or completed proof
Depends on: SPEC-0001 and SPEC-S001 at 0.1.0-draft.5
1. Purpose and authority
Across this specification family, MUST and SHALL are mandatory, MUST NOT and SHALL NOT are prohibitions, SHOULD is a recommendation permitting a documented exception, and MAY is permission. Requirement identifiers are revisioned; normative obligation wording is not a claim that the obligation has been fulfilled.
Noble SHALL have a small, mechanized semantic core and support optional compositional verification of individual programs. Verification is a development and assurance commitment, not a fourth expression form or a requirement that every caller supply a handwritten proof.
This document specifies the claims to establish and the boundaries those claims must preserve. VERIFICATION-TOOLCHAIN.md requires an Aeneas-first Rust implementation, mandatory across the semantic kernel, with Lean 4 reference definitions and refinement proofs. Verus is available only for reviewed non-kernel exceptions. PROGRAM-CONTRACTS.md specifies typed contracts, first-class evidence companions, and optional proof-required admission. SPEC-0001.md integrates these requirements. SOURCES.md records the inherited text and unavailable historical citation records.
Requirement: V-LIVE-01
r[V-LIVE-01]
V-LIVE-01. Live-Wasm-Draft acceptance MUST exercise actual source-to-independent-kernel-to-in-process-binary-to-selected-VM execution and source/module reloads, not only spec parsing, a synthetic evaluator or an old external assembler path. Positive observations MUST include next-new-top-level-call freshness, transitively rebuilt dependent definitions, retained old Program/active-frame behavior, unchanged typed stack, source diagnostics, denied host effects and bounded generation retirement. Negative controls MUST include invalid/incompatible types and effects, missing/unstable import snapshot, malformed or feature-unsupported Wasm, start/active initialization attempts, table/memory/retained-code exhaustion, host policy revocation and trap after an effectful prefix. Actual selected VM revision, binary emitter/verifier source, source/import/ABI hashes, limits, stages, requests, retired owners and open correspondence assumptions MUST be recorded. Strict interpretation and semantic preservation cannot be inferred from V8 bytecode execution or green document checks; any actual interpreter engine requires its own pinned conformance and separately evaluated trust/verification boundary. No earlier M4/MC1/MC2/DXM1 evidence or proof attaches automatically to a new generation.
Guest self-edit acceptance additionally MUST invoke a real Wasm guest that constructs a checked runtime-captured pure Program, observes its pinned source generation through a declared effect and enqueues a proposal, then show host-only post-return independent admission, identity-bound grant and generation CAS, and next-call freshness without mutating a retained old Program or selected source file. Controls MUST distinguish static pre-invocation refusal, actual callback denial of a saved old owner after rearm, recoverable post-return candidate refusal with preserved completed invocation stack/request prefix, and trap after enqueue with cancelled candidate and terminal policy. A bounded self-edit refusal for dependents does not prove the separate transitive dependent rebuild of LIVE-02.
Scenario: LIVE-04 for V-LIVE-01
- GIVEN the
Live-Wasm-Draftprofile and every field ofinputin LIVE-04 - WHEN the
admissionprocedure for caseLIVE-04runs against those inputs - THEN the observations match every field of
expectedin caseLIVE-04
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Scenario: LIVE-08 for V-LIVE-01
- GIVEN the
Live-Wasm-Draftprofile and every field ofinputin LIVE-08 - WHEN the
admissionprocedure for caseLIVE-08runs against those inputs - THEN the observations match every field of
expectedin caseLIVE-08
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Scenario: LIVE-09 for V-LIVE-01
- GIVEN the
Live-Wasm-Draftprofile and every field ofinputin LIVE-09 - WHEN the
runtimeprocedure for caseLIVE-09runs against those inputs - THEN the observations match every field of
expectedin caseLIVE-09
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Scenario: LIVE-10 for V-LIVE-01
- GIVEN the
Live-Wasm-Draftprofile and every field ofinputin LIVE-10 - WHEN the
runtimeprocedure for caseLIVE-10runs against those inputs - THEN the observations match every field of
expectedin caseLIVE-10
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Requirement: V-STATUS-01
r[V-STATUS-01]
V-STATUS-01. A specification requirement, a mathematical theorem statement, an implemented checker, an executed test, and an accepted proof MUST be reported as distinct things. The obligations in this package are all open; the accompanying fixtures are expectations, not executed Noble tests.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-AUTH-01
r[V-AUTH-01]
V-AUTH-01. The reviewed Lean 4 formalization SHALL be the reference mathematical model for verification claims. Its exact source revision and supporting definitions MUST be identified. A mismatch between that model and the language specification is a specification defect to resolve explicitly; neither a successful proof nor an informal paragraph silently overrides the other.
The existing draft already requires stack/type safety, conservative effects, resource accounting, immutable recipes, and explicit preparation. It explicitly does not establish a complete checking algorithm or soundness proof. These are the source-derived starting obligations, not completed results. [B1 §§5.4, 8–9, 11, 16; references in SOURCES.md]
Test design remains open for this requirement. No scenario or execution evidence is supplied.
2. Preserve the small language
Requirement: V-CORE-01
r[V-CORE-01]
V-CORE-01. Ordinary Program<S,T,e>, quote, compose, run, and reflect MUST retain their existing semantics. Ordinary execution MUST NOT require Lean, Aeneas, Verus, an SMT solver, or proof search to be installed or invoked. An admission service MAY check evidence separately under an explicit host policy.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-CORE-02
r[V-CORE-02]
V-CORE-02.
The optional contract profile MUST NOT introduce dependent Program parameters, general refinement inference, implicit behavioral coercions, or mandatory termination for ordinary programs. Typed contract and intrinsic Noble-source proof declarations are separate from the three executable core expression forms. The selected Intrinsic-Proofs-Draft permits predicative proof-level Type0/value Pi and dependent Eq only in nonexecuting declarations; it does not create runtime dependent values or an unbounded dependent core checker. Ordinary preparation/execution MUST NOT evaluate Lean proof terms or solver directives. Explicit finite rule replay follows SPEC-V002 without a general theorem prover. The core checker MUST NOT depend on arbitrary application contracts or proofs to establish stack, eligibility or effect validity.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-CORE-03
r[V-CORE-03]
V-CORE-03.
A new core construct or semantic change MUST identify its effect on the typing rules, operational model, ownership invariant, recipe observations, and existing proofs. Before stable-core inclusion, its relevant obligations MUST be discharged or the feature MUST remain explicitly outside that stable subset. A tool limitation is not permission to weaken the language silently.
Typed contract declarations and first-class companion APIs are selected deliverables under SPEC-V002. The implemented bounded MC1/MC2 grammars and representations retain their separately scoped evidence; wider forms remain explicit entry gates. The optional Intrinsic-Proofs-Draft freezes only its versioned, nonexecuting Noble module proof-declaration grammar under K-PROOF-01. Existing mathematical or harness examples outside that profile do not become accepted source merely by resemblance. # remains the comment prefix and the three ordinary executable expression forms are unchanged.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-CLAIM-01
r[V-CLAIM-01]
V-CLAIM-01. Every assurance claim MUST state its target, scope, semantic revision, assumptions, evidence class, and unresolved correspondence boundaries. “Formally verified Noble” MUST NOT be used as an unqualified substitute for those fields.
A kernel proof about an Aeneas-generated Lean function is evidence about that function. Its application to Rust also relies on the extraction path and external models. A Verus result belongs to the Verus verification chain. Neither automatically proves that a Noble-to-Wasm compiler preserves semantics. These distinctions are normative Noble policy; the selected tools' documented roles motivate them. [R1–R7]
Test design remains open for this requirement. No scenario or execution evidence is supplied.
4. Formal model requirements
4.1 Domains and explicit assumptions
Requirement: V-MODEL-01
r[V-MODEL-01]
V-MODEL-01. The Lean model MUST represent resolved bodies and references, value and stack types, instantiated first-class program interfaces, rank-1 named schemes, effect bounds, eligibility judgments, values and captures, recipe normalization, and an immutable definition environment. Recursive references MUST retain owner scope.
The initial mechanization MAY cover a smaller bootstrap subset, but MUST enumerate omissions. A result for nonrecursive arithmetic quotations MUST NOT be presented as a result for the whole language, lists, recursion, host operations, or resources.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-MODEL-02
r[V-MODEL-02]
V-MODEL-02. The model MUST separate language evaluation, host responses, resource ownership, and execution-profile failures. Host contracts SHALL be explicit parameters with well-formedness and preservation conditions. Their satisfaction by a concrete adapter is a separate implementation obligation.
A useful configuration refinement is:
Configuration = (pending code, data stack, continuation frames,
host state, ownership accounting, event history)
This is proof notation, not a frozen runtime representation. Continuation frames include values temporarily held by dip. Pure programs need not carry a concrete host-state object in their runtime representation.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
4.2 Steps, waiting, failure, and divergence
Requirement: V-MODEL-03
r[V-MODEL-03]
V-MODEL-03. The model MUST distinguish normal return, permitted abnormal termination, host waiting where supported, divergence, and forbidden language-level stuck states. Ordinary Result error variants are normal returns. An implementation MUST NOT conceal a wrong operand type, stack underflow, or consumed-handle use by labeling it an allowed trap.
The ideal core semantics may abstract from allocation and execution budgets. The relation to a bounded execution profile must then state how allocation failure, cancellation, and fuel exhaustion are represented. Exhausting fuel in a test evaluator is not evidence that the source program diverges.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-MODEL-04
r[V-MODEL-04]
V-MODEL-04. Safety and effect claims MUST cover finite prefixes of executions, including prefixes of diverging, waiting, and abnormally terminating computations. A theorem only about successful returns is insufficient for either claim.
A closed, well-typed configuration must either be at an allowed outcome, be waiting at a permitted host boundary, or admit a modeled step. This is progress relative to the stated host model, not a host-availability or termination guarantee.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
4.3 Host effects and authority
Requirement: V-EFFECT-01
r[V-EFFECT-01]
V-EFFECT-01. Effect soundness SHALL establish that each guest-requested HostOpId in every modeled execution prefix is in the instantiated effect bound. The model MUST account for requests that are denied or fail, not only successful responses. Logical request labels need not become publicly stored audit logs.
Operation internals are governed by the host contract; an effect set is not an inventory of every OS action inside native code. Runtime housekeeping and abort cleanup MUST be classified separately from guest-requested operations rather than incorrectly added to quotation construction's latent effects.
Scenario: CORE-17 for V-EFFECT-01
- GIVEN the
Core-Bootstrapprofile and every field ofinputin CORE-17 - WHEN the
runtimeprocedure for caseCORE-17runs against those inputs - THEN the observations match every field of
expectedin caseCORE-17
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Requirement: V-EFFECT-02
r[V-EFFECT-02]
V-EFFECT-02. Construction and inspection MUST issue none of the guest host requests latent in the programs they manipulate. Invocation and higher-order combinators MUST propagate their called programs' bounds without changing event order.
Effect sets summarize possible operation identities. They do not prove event counts, temporal protocols, success, authorization, or information-flow properties.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
4.4 Ownership across the whole machine
Requirement: V-RESOURCE-01
r[V-RESOURCE-01]
V-RESOURCE-01. Ownership invariants MUST range over the entire live configuration, including aggregate payloads, hidden dip values, invocation/session transfers, and pending host-boundary state. The model MUST track multiplicity of ownership obligations; a set that silently collapses duplicate handles is insufficient.
A normal host transition may create, transfer, replace, or retire obligations according to its contract. Conservation therefore means preservation modulo those declared transitions, not preservation of an unchanging number of handles. Proofs MUST distinguish a guest handle's obligation from physical exclusivity of the underlying external object.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-RESOURCE-02
r[V-RESOURCE-02]
V-RESOURCE-02. The formal eligibility relation MUST exclude resource-bearing values recursively from Data and Capture, including all payload alternatives allowed by an instantiated schema. Program interfaces may mention resources without making the program's environment resource-owning. Captured environments MUST actually satisfy their own eligibility invariant.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-RESOURCE-03
r[V-RESOURCE-03]
V-RESOURCE-03. Normal-path preservation and abnormal cleanup SHALL have separate claims. A guest typing proof does not prove host cleanup. The cleanup claim MUST state invocation boundaries, locally observable retirement, repeat-retirement behavior, and prevention of stale-handle resurrection. It MUST NOT assert remote exactly-once cleanup or cleanup after total machine failure without a separately modeled protocol.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
4.5 Arithmetic and observation
Requirement: V-NUM-01
r[V-NUM-01]
V-NUM-01. The model and every implementation bridge MUST use the baseline's wrapping I64 semantics for +, -, and *. Mathematical integers in Lean or Verus MUST be connected to these operations by an explicit modulo-2^64 representation relation. Out-of-range literals remain rejections, not wrapped literals. [B1 §4.4]
Scenario: CORE-14 for V-NUM-01
- GIVEN the
Core-Bootstrapprofile and every field ofinputin CORE-14 - WHEN the
runtimeprocedure for caseCORE-14runs against those inputs - THEN the observations match every field of
expectedin caseCORE-14
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Requirement: V-OBS-01
r[V-OBS-01]
V-OBS-01. Proofs MUST distinguish execution-only equivalence from recipe equality and contextual equivalence in a language with reflect. Program-valued results and programs nested inside returned data retain their observable recipes. Equal scalar answers alone do not justify substituting program values in all contexts.
For example, calling [ 1 1 + ] and [ 2 ] gives the same integer in the ideal pure semantics, but reflecting those program values distinguishes them. A backend may optimize executable instructions while preserving the specified recipe. An explicit source transformation may produce a different program identity. [B1 §11.2]
Test design remains open for this requirement. No scenario or execution evidence is supplied.
5. Checking architecture and proof obligations
5.1 Separate elaboration from acceptance
The selected architecture is:
untrusted source / syntax / imported description
-> parsing, resolution, inference, elaboration
-> explicit resolved candidate and witnesses
-> small acceptance checker
-> checked representation
-> backend and loader
Requirement: V-CHECK-01
r[V-CHECK-01]
V-CHECK-01. Acceptance MUST validate the candidate from an untrusted representation. A proof whose checker input already has the desired typing invariant by construction does not establish validation of hostile input. Declared signatures, type witnesses, schemas, effects, and dependency identities MUST NOT be accepted solely because the producer supplies them.
Scenario: CORE-16 for V-CHECK-01
- GIVEN the
Core-Bootstrapprofile and every field ofinputin CORE-16 - WHEN the
staticprocedure for caseCORE-16runs against those inputs - THEN the observations match every field of
expectedin caseCORE-16
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Requirement: V-CHECK-02
r[V-CHECK-02]
V-CHECK-02. Inference can be complex and initially unproved during staged development. It remains an Aeneas target under VT-SCOPE-01, not a permanent verification exclusion. Acceptance MUST independently validate the relevant elaboration evidence against the declarative rules. Acceptance MUST include reference scope, scheme instantiation, stack order, eligibility, conservative effects, and resource obligations. The exact witness encoding and complete inference algorithm remain separate open deliverables.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-CHECK-03
r[V-CHECK-03]
V-CHECK-03. The acceptance theorem SHALL have the following shape, with explicit representation and environment well-formedness premises:
check(candidate, expected_interface, environment) = Accepted(checked)
implies
WellFormed(checked) and TypingDerivation(checked, expected_interface)
Correct extraction of the Rust function and its correspondence to check are additional claims. Well-formedness of an externally supplied environment must itself be established or appear as a disclosed trusted boundary.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-CHECK-04
r[V-CHECK-04]
V-CHECK-04. Rejection, unsupported input, resource exhaustion, internal failure, and successful acceptance MUST be distinguishable internally and in assurance reports. A checker MUST NOT accept on timeout. Soundness does not imply completeness, and a diagnostic MUST NOT claim a proof of semantic impossibility without an applicable theorem.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-CHECK-05
r[V-CHECK-05]
V-CHECK-05. A usable-checker claim MUST include positive acceptance coverage and a termination argument for the selected finite checking problem under the ideal model. Rejecting every candidate cannot satisfy the advertised supported-language coverage. Cyclic type equations, recursive schema checks, and effect constraints must not be hidden sources of divergence.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-CHECK-06
r[V-CHECK-06]
V-CHECK-06. Validation MUST finish before the submission body runs. prepare performs its own documented service effects but MUST NOT execute candidate bodies, tactics, or arbitrary user macros as an undocumented checking step. Successful checking does not prove an optional behavioral contract.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
5.2 Required theorem families
The machine-readable backlog is verification/obligations.json. This ledger is newly transcribed from the PO and SO tables, not recovered from an older package. Each obligation remains open. Identifiers name theorem obligations, not implemented Lean declarations.
| Obligation | Claim to establish | Initial evidence route |
|---|---|---|
| PO-01 | Well-formed syntax, references, types, substitutions, and eligibility relations | Lean 4 |
| PO-02 | Preservation of stack and value typing across modeled steps | Lean 4 |
| PO-03 | Progress relative to host contracts; exclusion of forbidden stuck states | Lean 4 |
| PO-04 | Effect soundness for every finite execution prefix | Lean 4 |
| PO-05 | Whole-configuration ownership preservation on normal transitions | Lean 4 |
| PO-06 | Recursive Data/Capture soundness and resource-free program environments | Lean 4 |
| PO-07 | quote/compose/run execution laws and construction-effect separation | Lean 4 |
| PO-08 | Recipe normalization, quotation boundaries, and reflection consistency | Lean 4 |
| PO-09 | Acceptance-checker soundness against the declarative judgments | Lean 4 |
| PO-10 | Checker termination and supported-fragment acceptance coverage | Lean 4 plus executed fixtures |
| PO-11 | Actual Rust acceptance checker refines the Lean checker contract | Charon/Aeneas to Lean 4 |
| PO-12 | Rust arithmetic and recipe representation refine their logical models | Charon/Aeneas to Lean 4 |
| PO-13 | Handle-table access and retirement preserve the deterministic state contract | Charon/Aeneas to Lean 4 |
| PO-14 | Host adapter, wrapper, and cross-boundary contract correspondence | Aeneas-refined decisions plus explicit external-boundary evidence |
| PO-15 | Behavioral contract sequencing and builder-family rules are sound | Lean 4 |
| PO-16 | Evidence admission binds the intended claim, subject, and assumptions | Lean 4 / Rust refinement, after envelope design |
| PO-17 | Selected backend lowering preserves execution and recipe observations | Aeneas-extracted lowering plus Lean/Wasm preservation proof or translation validation |
| PO-18 | Artifact loading binds executed Wasm to checked program and recipe | Aeneas-refined loading decisions plus concrete loader/build correspondence |
| PO-19 | Contract IR export and obligation generation preserve the intended Noble subject and proposition | Lean semantics plus Aeneas refinement of the exporter/checker |
| PO-20 | Companion construction, finite proof-rule replay, and proof erasure preserve certified claims and ordinary program observations | Lean rules plus Aeneas refinement of companion operations |
| PO-21 | Applicability checks establish the exact precondition before certified invocation | Lean guard correspondence plus Aeneas refinement and boundary tests |
Typing rules for recursive named definitions and generalization are part of the eventual PO-01/02/09 coverage. They are not discharged by a theorem for only monomorphic, nonrecursive source. Parser correctness, canonical binary encoding, ABI details, and full schema checking remain separately scoped until their specifications exist.
Worker-contract coverage
The worker amendment extends existing obligations rather than declaring a verified substrate:
| Contract | Required obligation coverage |
|---|---|
| Typed dispatch and effect widening | PO-01/02/04/07/09 and actual checker refinement under PO-11 |
| Round trips and package closure | PO-08/09/12/17/18, including encoding and artifact correspondence |
| Bounded candidate admission | PO-09/10/11 and hostile-boundary checks under SO-05/06 |
| Async ownership and cancellation races | PO-05/13/14 and synchronization correspondence under SO-09 |
| Bounded execution and outcome records | PO-03/04/14/17 and safety-relevant failures under SO-07 |
Worker conformance supplies test designs, not proofs of these claims. Scalar bootstrap evidence cannot close schema, package, async, or complete execution-profile coverage. An ideal semantic theorem does not establish a concrete interruption bound or eventual native retirement.
Octet-adoption coverage
H-AUTH-01–04 extend PO-05/06/13/14 and SO-02/04/06 for witness ownership, constructor opacity, and protected admission. H-RECEIPT-01/02 extend PO-14/16/18 for observed-outcome and artifact/claim correspondence. A proved policy function alone does not establish current credentials, revocation facts, or external success.
DX-TYPE-03/04, DX-PROTOCOL-03, and DX-MODULE-04 extend PO-01/02/04/09 and the corresponding Rust refinement obligations. Explicit transition coverage does not prove liveness. Resolved effect bounds for higher-order programs do not require enumeration of every runtime program value.
VT-OCTET-01–03 and EV-BIND-01/EV-TIER-01 add separate policy and evidence-admission obligations. Their results cannot close these semantic proof obligations by themselves. No obligation is accepted through the adoption test designs.
6. Evidence classes and trust closure
Requirement: V-EVIDENCE-01
r[V-EVIDENCE-01]
V-EVIDENCE-01. Reports MUST label evidence as one of: lean-kernel, aeneas-lean, verus, translation-validation, test, review, or assumption. These are not a numerical ranking and are not interchangeable. A translation-validation result MUST name its validator and that validator's own assurance.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-EVIDENCE-02
r[V-EVIDENCE-02]
V-EVIDENCE-02. A claim's trust description MUST include the transitive assumptions of its supporting claims and bridges. No translation, external model, host contract, or build step becomes proved merely by being placed between two proved components.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-EVIDENCE-03
r[V-EVIDENCE-03]
V-EVIDENCE-03. “Lean-kernel checked” SHALL mean that the exact theorem and proof have been accepted under the selected Lean policy. It MUST NOT mean that Lean verified the correctness of Charon, Aeneas, Verus, rustc, LLVM, the Wasm engine, or the hardware. Additional compiler/native-evaluation trust must be disclosed if permitted.
Lean's validation guidance distinguishes proof validity from theorem meaning and recommends axiom inspection and stronger checking for untrusted submissions. This workstream adopts the concrete policies in IMPL-V001 rather than treating a green editor indicator as the complete assurance argument. [R4]
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-EVIDENCE-04
r[V-EVIDENCE-04]
V-EVIDENCE-04. A proof result MUST bind exact relevant source and dependency revisions, feature flags, target architecture, semantic model, statement, and toolchain. Changed inputs invalidate the prior applicability decision until revalidated. Unchanged code may acquire new proofs without changing its language identity.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
7. Release gates and staged delivery
These stages group assurance deliverables, not time estimates or mandatory serial dependencies. Pure V4 contract work can follow V2 without the V3 resource extension. ROADMAP.md defines the MC1/MC2 dependency edges.
| Stage | Deliverable | Permitted claim after actual completion |
|---|---|---|
| V0 | This specification, boundary register, fixture expectations | Verification design specified |
| V1 | Lean model and metatheory for enumerated bootstrap quotation subset | Named theorems for that model/subset |
| V2 | Explicit acceptance checker; Aeneas extraction and refinement; compatible pinned tools | Rust-checker properties relative to disclosed extraction/build assumptions |
| V3 | Resource extension, Aeneas-refined handle transitions, checked wrappers and external-boundary evidence | Scoped resource assurance with explicit host assumptions |
| V4 | Typed contracts, first-class companions, builder-family proofs, admission and applicability checks | Exact behavioral claims for accepted subjects under the named policy |
| V5 | Selected Wasm lowering and loading correspondence | Only the specifically connected compilation/execution claims |
Requirement: V-GATE-01
r[V-GATE-01]
V-GATE-01. A stable core specification MUST have reviewed, Lean-kernel-checked metatheory for the entire subset it labels stable, covering PO-01 through PO-10 as applicable. A proof plan alone is insufficient. Exploratory drafts and implementations MAY exist earlier but MUST retain their draft and coverage labels.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-GATE-02
r[V-GATE-02]
V-GATE-02. An implementation advertised as having a verified acceptance checker MUST additionally satisfy PO-11 for the actual supported implementation and publish its translation/model assumptions and configuration. Finishing V1 alone cannot justify that label.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-GATE-03
r[V-GATE-03]
V-GATE-03. A claim of source-to-Wasm behavioral preservation or end-to-end assurance MUST identify evidence covering the backend, loading, runtime adapters, and remaining execution trust. Otherwise those boundaries MUST remain visibly unproved or assumed. Stable language semantics do not by themselves mean a verified compiler.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-GATE-04
r[V-GATE-04]
V-GATE-04. Any release that advertises verification SHALL publish a claim ledger, feature coverage, transitive assumption register, pinned toolchain, and executed validation results. Unproved obligations and unsupported features MUST remain visible. CI MUST reject unexplained growth of the trusted boundary or weakened claims.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-GATE-05
r[V-GATE-05]
V-GATE-05. A verified-kernel claim MUST cover every applicable kernel function under VT-SCOPE-02/05, not only the acceptance checker. A whole-project claim additionally requires explicit coverage of all Noble-owned production components and their correspondence boundaries. Extraction success, test success, and an approved exception MUST NOT substitute for a required refinement proof. Narrow experimental results MUST retain their exact subset labels.
Scenario: VERIFY-01 for V-GATE-05
- GIVEN the
Implementation-Policyprofile and every field ofinputin VERIFY-01 - WHEN the
reviewprocedure for caseVERIFY-01runs against those inputs - THEN the observations match every field of
expectedin caseVERIFY-01
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Scenario: VERIFY-05 for V-GATE-05
- GIVEN the
Implementation-Policyprofile and every field ofinputin VERIFY-05 - WHEN the
reviewprocedure for caseVERIFY-05runs against those inputs - THEN the observations match every field of
expectedin caseVERIFY-05
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
7A. Safety extension obligations
SPEC-S001 defines the mandatory Noble-safe claim. The following obligations extend, rather than replace, PO-01 through PO-21.
| Obligation | Claim to establish | Initial evidence route |
|---|---|---|
| SO-01 | Accepted core executions have no forbidden stuck state and no undefined-behavior semantic outcome | Lean 4 |
| SO-02 | Representation/eligibility rules exclude resource or authority forgery through ordinary data | Lean 4 |
| SO-03 | Effect soundness holds for every finite prefix, including denied/failed host requests | Lean 4 |
| SO-04 | Whole-configuration ownership is preserved modulo declared host transitions | Lean 4 |
| SO-05 | Syntax, untrusted metadata, and artifacts cannot become callable without validation/preparation | Lean 4 + checker refinement |
| SO-06 | Concrete public host/resource wrappers establish all verified internal preconditions from hostile representations | Aeneas-refined validation plus concrete wrapper evidence |
| SO-07 | Selected Noble-to-Wasm lowering preserves the claimed safety-relevant behavior/failure relation | Backend proof or translation validation |
| SO-08 | Standard concurrency semantics preserve isolation, scope-owned conversational state, protocol typing, and runtime capability checks | Lean 4, after concurrency profile freeze |
| SO-09 | Concrete concurrency runtime state machinery refines SO-08 under documented synchronization assumptions | Aeneas-refined sequential transitions plus separate synchronization evidence |
Requirement: V-SAFE-01
r[V-SAFE-01]
V-SAFE-01. A stable Noble-safe core claim MUST include completed applicable SO obligations in addition to the existing stable-core gates. A theorem about ideal source semantics alone does not establish safety of hostile host boundaries or emitted executable artifacts.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-SAFE-02
r[V-SAFE-02]
V-SAFE-02. A guest-language safety proof MUST model all permitted outcomes as defined outcomes. Wrong operand types, stack underflow, consumed-handle use, forged live authority, and arbitrary representation reinterpretation are forbidden stuck states or rejected boundary inputs, not allowed traps used to hide unsoundness.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-SAFE-03
r[V-SAFE-03]
V-SAFE-03. Under the standard profile, Noble participants MUST have no direct shared mutable Noble memory. This is the standard-concurrency race-freedom invariant. It does not remove the obligation to prove or validate synchronization in the Rust runtime that implements the abstract dataspace/facet model.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: V-SAFE-04
r[V-SAFE-04]
V-SAFE-04. Future advanced type-system features may join a stable-safe subset only when their declarative rules, acceptance checks, implementation correspondence, and affected safety theorems are included in the release evidence.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
8. Explicit non-goals and open work
This revision does not prove every application, make host availability decidable, verify Wasmtime or rustc, establish constant-time execution, supply cryptographic proof compression, or introduce a new general-purpose prover. Those properties require separate models and evidence where desired.
The full checker algorithm, Lean mechanization, Rust components, compatible tools, boundary models, evidence transport, and Wasm correspondence remain open.
CORE-BOOTSTRAP.md scopes the first finite checker. ROADMAP.md places extraction feasibility beside the checker and Wasm experiments. EVIDENCE.md prevents document validation from changing runtime or proof status.
Requirement: V-INTRINSIC-01
r[V-INTRINSIC-01]
V-INTRINSIC-01.
The reviewed Lean 4 reference semantics MUST define the exact PureTyCode/El interpretation, predicative logical binder/equality rules, typed source proof calculus, captured program subject and partial-correctness propositions, finite rule expansion and proof erasure for Intrinsic-Proofs-Draft. The first independent checker target for Noble-source proof declarations MUST be the pinned strict Lean consumer, checking a restricted translated closed term against the independently regenerated exact MC1 obligation with inspected transitive assumptions; it MUST NOT accept producer Lean declarations, unchecked axioms, native proof evaluation, tactics or a Bend2 proof route. exec_lit_iff, exec_add_iff, pc_exact, pc_sequence, exportedClaim_unaryI64, exportedClaim_consequence, holds_true, holds_eq and List.append_assoc MUST be named only when their exact reviewed theorem declarations, premises and model/standard-library revisions are available. Both pc_sequence implications MUST be proved from checked exact-result hypotheses; symbolic append-associativity is not definitionally equal by fiat. The conversion of an inner PC into exact exported MC1 typed/pre/post Holds predicates MUST independently prove typing, output existential and expression evaluation; it MUST NOT assume PC and exportedClaim are interchangeable. A proof about wrapping I64 normal returns MUST NOT be reported as termination, runtime Wasm correspondence, host authority or a theorem for arbitrary integers.
The source proof typing/translation-soundness theorem, proof-kernel consistency/normalization and decidability/coverage, independently checked type-code/domain fidelity, exact MC1 source/contract/subject correspondence, proof erasure preservation, and actual Rust parser/checker/translator refinement and backend/artifact correspondence are separate open obligations. Lean kernel acceptance of generated output alone MUST NOT close them. Reports MUST identify exact source declaration/version, accepted subject/recipe, expected claim, assumptions, proof-library/model/consumer revisions, evidence class and unresolved boundaries.
For the selected named-call revision, the independently expected NamedV2.exportedNamedClaim MUST include StackTyped for before, empty typed params and after, pre before [] params, preservation of every bottom-stack tail, named Exec of the accepted twice definition BODY and Holds₂ for the entire original MC1 Term vocabulary. A strict Lean example with step [ 1 + ] specialized at two distinct uses and twice [ d.step d.step ] can prove the wrapping x+2 proposition, with only propext, Classical.choice and Quot.sound as axioms. That is a model/fixture result, NOT a theorem about the authentic submitted one-call root, exact source occurrences, no-orphan graph, owner policy, Rust implementation or backend. Independent host source and graph admission MUST establish those premises before any named source proof can be accepted; NamedV1.Satisfies or NamedV1.ExactClosure alone cannot establish them. No implementation/evidence promotion follows from this specification or model fixture.
A reusable reviewed two-call rule MUST be parameterized by independently accepted actual graph/row and source premises and instantiated at the regenerated exact claim by the strict consumer. The synthetic example with fixed slots 7/8 MUST NOT be reused as authority over actual source. The rule, source-origin authentication and generated exact claim are joint admission prerequisites, not substitutes for each other; no host result is claimed here.
Reuse existing policy, not a new language backend: under VT-ROLE-01 and VT-SCOPE-01, actual production Rust module parsing/resolution, source-term checker, typed Noble→Lean proof lowering and admission decisions follow the pinned Charon → Aeneas → Lean implementation-refinement route with source coverage and explicit external models. Charon/Aeneas consume Rust representations; they do not parse Noble source, accept Noble proof terms or provide the independent exact application-theorem kernel check. No new extraction/refinement is claimed by this spec change. Eurydice (Rust→C), Scylla (C→Rust) and Bend2 are not selected proof dependencies.
Scenario: CONTRACT-18 for V-INTRINSIC-01
- GIVEN the
Intrinsic-Proofs-Draftprofile and every field ofinputin CONTRACT-18 - WHEN the
admissionprocedure for caseCONTRACT-18runs against those inputs - THEN the observations match every field of
expectedin caseCONTRACT-18
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Scenario: CONTRACT-22 for V-INTRINSIC-01
- GIVEN the
Intrinsic-Proofs-Draftprofile and every field ofinputin CONTRACT-22 - WHEN the
runtimeprocedure for caseCONTRACT-22runs against those inputs - THEN the observations match every field of
expectedin caseCONTRACT-22
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Scenario: CONTRACT-24 for V-INTRINSIC-01
- GIVEN the
Intrinsic-Proofs-Draftprofile and every field ofinputin CONTRACT-24 - WHEN the
reviewprocedure for caseCONTRACT-24runs against those inputs - THEN the observations match every field of
expectedin caseCONTRACT-24
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Scenario: CONTRACT-26 for V-INTRINSIC-01
- GIVEN the
Intrinsic-Proofs-Draftprofile and every field ofinputin CONTRACT-26 - WHEN the
admissionprocedure for caseCONTRACT-26runs against those inputs - THEN the observations match every field of
expectedin caseCONTRACT-26
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Requirement: V-INTRINSIC-02
r[V-INTRINSIC-02]
V-INTRINSIC-02.
An implementation acceptance plan MUST exercise actual .noble module source (not a standalone .noble-contract proof simulation) through parse/resolve, independent source-term/type acceptance, restricted Lean lowering, independent strict exact-claim check and optional MC2 evidence admission with distinct outcomes. It MUST include positive Pi(A:Type0).Pi(x:El(A)).Eq(A,x,x) with checked pure type instantiations, equality introduction/elimination, and genuine reviewed-rule composition of accepted [ 1 + ] including signed wrapping boundary. Hostile controls MUST include resource/Program/Syntax type-code smuggling, Type0:Type0, partial values, false propositions, missing or forged bridge/join, stale/foreign theorem/model, altered subject/recipe/capture, vacuous or weakened candidate-authored law against owner-frozen policy, mismatched proof/artifact, holes, cycles, oversized terms and verifier exhaustion. Failed proof-required routes MUST start zero candidate/host-protected operations. Case implementation/execution states MUST follow their distinct exact source-bound receipts; document validation or successful Lean compilation alone does not promote proof, implementation or execution status.
The additive CONTRACT-26 design MUST separately test revision-2 grammar, two authenticated occurrences of the same imported lexical step with fresh per-use specialization, complete source-bound graph and typed instantiation, independently generated x+2 wrapping claim and exact strict checker result. Wrong two-call body, slot, owner, partial d versus full d.step token byte span, wrong original word token despite nonempty ordered span, ordinal, incomplete Inst, orphan/unreachable rows, replay of a previously prepared @5 source-bound goal/graph/claim after alias/import-version rebinding to @6 despite identical token text, false/weakened predicates, cross-revision references, unsupported revision-2 colon proofs and quotation/reflection identity confusion MUST refuse on their own boundaries. Fresh independently authenticated @6 source and its new exact proof remain eligible; the hostile is stale @5 replay, not every version change. Proof erasure and ordinary runtime identity MUST be observed separately from Lean model checking. CONTRACT-26 has finite source-bound implemented/passed acceptance under the separately retained named-v2 receipt, with proof open; it does not discharge translation/Rust refinement, owner policy, artifact or host-release obligations.
Scenario: CONTRACT-17 for V-INTRINSIC-02
- GIVEN the
Intrinsic-Proofs-Draftprofile and every field ofinputin CONTRACT-17 - WHEN the
staticprocedure for caseCONTRACT-17runs against those inputs - THEN the observations match every field of
expectedin caseCONTRACT-17
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Scenario: CONTRACT-18 for V-INTRINSIC-02
- GIVEN the
Intrinsic-Proofs-Draftprofile and every field ofinputin CONTRACT-18 - WHEN the
admissionprocedure for caseCONTRACT-18runs against those inputs - THEN the observations match every field of
expectedin caseCONTRACT-18
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Scenario: CONTRACT-20 for V-INTRINSIC-02
- GIVEN the
Intrinsic-Proofs-Draftprofile and every field ofinputin CONTRACT-20 - WHEN the
admissionprocedure for caseCONTRACT-20runs against those inputs - THEN the observations match every field of
expectedin caseCONTRACT-20
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Scenario: CONTRACT-21 for V-INTRINSIC-02
- GIVEN the
Intrinsic-Proofs-Draftprofile and every field ofinputin CONTRACT-21 - WHEN the
admissionprocedure for caseCONTRACT-21runs against those inputs - THEN the observations match every field of
expectedin caseCONTRACT-21
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Scenario: CONTRACT-23 for V-INTRINSIC-02
- GIVEN the
Intrinsic-Proofs-Draftprofile and every field ofinputin CONTRACT-23 - WHEN the
admissionprocedure for caseCONTRACT-23runs against those inputs - THEN the observations match every field of
expectedin caseCONTRACT-23
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Scenario: CONTRACT-24 for V-INTRINSIC-02
- GIVEN the
Intrinsic-Proofs-Draftprofile and every field ofinputin CONTRACT-24 - WHEN the
reviewprocedure for caseCONTRACT-24runs against those inputs - THEN the observations match every field of
expectedin caseCONTRACT-24
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Scenario: CONTRACT-25 for V-INTRINSIC-02
- GIVEN the
Intrinsic-Proofs-Draftprofile and every field ofinputin CONTRACT-25 - WHEN the
staticprocedure for caseCONTRACT-25runs against those inputs - THEN the observations match every field of
expectedin caseCONTRACT-25
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Scenario: CONTRACT-26 for V-INTRINSIC-02
- GIVEN the
Intrinsic-Proofs-Draftprofile and every field ofinputin CONTRACT-26 - WHEN the
admissionprocedure for caseCONTRACT-26runs against those inputs - THEN the observations match every field of
expectedin caseCONTRACT-26
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Requirement: NGRAPH-DERIV-01
r[NGRAPH-DERIV-01]
NGRAPH-DERIV-01.
A numeric graph training contract MUST name a versioned derivative policy that is separate from the forward semantics. ngraph-deriv-real/0 MUST differentiate the intended real-valued arithmetic, ignoring FP32 rounding, and MUST give the ReLU and ReGLU gate a zero derivative at a zero gate, including -0. It MUST NOT be described as the derivative of the floating-point function. By default, admission MUST statically refuse a training contract in which any declared parameter reaches the loss through a hard boolean node or through the index operand of an index read. Straight-through, smooth-surrogate, zero-index and discrete-difference index rules MUST each be explicit per-node opt-ins and MUST be recorded in training provenance. Gradients with respect to selected candidates MUST be kept distinct from gradients with respect to the index. Reverse evaluation MUST return gradients only for declared parameters and MUST NOT mutate the checked Program, the graph or any parameter snapshot.
Scenario: NGRAPH-CASE-11 for NGRAPH-DERIV-01
- GIVEN the
Numeric-Graph-v1profile and every field ofinputin NGRAPH-CASE-11 - WHEN the
admissionprocedure for caseNGRAPH-CASE-11runs against those inputs - THEN the observations match every field of
expectedin caseNGRAPH-CASE-11
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-DERIV-01
- GIVEN the
Numeric-Graph-v1profile and every field ofinputin NGRAPH-CASE-14 - WHEN the
identityprocedure for caseNGRAPH-CASE-14runs against those inputs - THEN the observations match every field of
expectedin caseNGRAPH-CASE-14
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Requirement: NGRAPH-PROOF-01
r[NGRAPH-PROOF-01]
NGRAPH-PROOF-01.
Numeric graph assurance claims MUST separately state reference agreement, finite differential or test coverage, discharged theorem obligations and trusted compiler, runtime and toolchain boundaries. Extraction preservation, graph-checker soundness, compiled-evaluator refinement, FP32 reference-model correspondence, fact-transfer soundness and derivative-policy conformance MUST remain open obligations until separately proved or evidenced. Matching hashes, agreement between evaluators, finite differential tests and producer-supplied metadata MUST NOT be reported as discharging any of them.
For a chained graph-rewrite certificate, every transition MUST bind the exact logical identity of the graph immediately before and immediately after that transition. The independent checker MUST derive each intermediate graph by replaying the preceding accepted transition and MUST check the next transition's premise against that graph. A prune transition MUST derive its survivor mapping from the graph itself; the mapping MUST preserve creation order, operand order, ordered outputs and output multiplicity. Pruning MUST retain all output-, guard-, fact- and interface-reachable nodes, and a removed node MUST NOT reach any output or guard. The exact pruned candidate MUST pass the existing independent graph-admission checker before execution. A changed transition grammar MUST use a distinct certificate revision and MUST NOT change the meaning of a previously accepted revision. These finite model, test, and measurement results do not establish correspondence between Rust code and a Lean model.
Scenario: NGRAPH-CASE-15 for NGRAPH-PROOF-01
- GIVEN the
Numeric-Graph-v1profile and every field ofinputin NGRAPH-CASE-15 - WHEN the
reviewprocedure for caseNGRAPH-CASE-15runs against those inputs - THEN the observations match every field of
expectedin caseNGRAPH-CASE-15
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Scenario: NGRAPH-CASE-16 for NGRAPH-PROOF-01
- GIVEN the
Numeric-Graph-v1profile and every field ofinputin NGRAPH-CASE-16 - WHEN the
runtimeprocedure for caseNGRAPH-CASE-16runs against those inputs - THEN the observations match every field of
expectedin caseNGRAPH-CASE-16
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Scenario: NGRAPH-CASE-16 chained rewrite admission and execution controls
- GIVEN the
Numeric-Graph-v1profile and every field ofinputin NGRAPH-CASE-16 - WHEN the independent checker replays each certificate transition, derives its prune mapping, admits the resulting candidate, and executes only admitted candidates
- THEN every executed output vector MUST equal the source reference in order and multiplicity, every listed stale or malformed variant MUST be refused before candidate execution, and the unchanged legacy revisions MUST retain their prior semantics
Requirement: NGV-F32-01
r[NGV-F32-01]
NGV-F32-01.
A kernel-checked binary32 reference MUST select exactly one model with a documented trusted base, and MUST NOT depend on an opaque host floating-point primitive or an unaudited toolchain pin change.
The selected model MUST define raw 32-bit encodings with sign, 8-bit biased exponent and 23-bit fraction; classify infinity, NaN, both zero signs, subnormal and normal; interpret every finite encoding exactly as mantissa * 2^exponent; round to nearest with ties to even at one rounding per operation; retain gradual underflow at 2^-149; form an infinity at or above 2^128 with the operand sign; keep signed zeros distinct; and treat every NaN encoding as one observation class while every other class keeps its exact bytes.
Addition, subtraction and multiplication MUST follow the IEEE 754 rules for NaN propagation, infinities, signed zeros, overflow and underflow. Reassociation and implicit contraction MUST be observable, so a fold order and a single rounding per operation are normative.
Scenario: NGV-F32-CASE-01 for NGV-F32-01
- GIVEN the
Binary32-Reference-Modelprofile and every field ofinputin NGV-F32-CASE-01 - WHEN the
identityprocedure for caseNGV-F32-CASE-01runs against those inputs - THEN the observations match every field of
expectedin caseNGV-F32-CASE-01
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Requirement: NGV-F32-02
r[NGV-F32-02]
NGV-F32-02.
Reference expectations MUST be proposed by an independent external tool and MUST each be checked against the selected model by the Lean kernel. A proposal alone MUST NOT be evidence, and inability to check one MUST NOT be a pass.
The retained family MUST cover halfway ties in both directions, cancellation, both zero signs, the smallest and largest subnormals, the normal/subnormal boundary, gradual underflow, overflow in both signs, infinities and NaNs. Native or external tools MAY propose expected bits, but no native_decide, decide +native or native #eval MAY replace kernel checking.
Scenario: NGV-F32-CASE-02 for NGV-F32-02
- GIVEN the
Binary32-Reference-Modelprofile and every field ofinputin NGV-F32-CASE-02 - WHEN the
identityprocedure for caseNGV-F32-CASE-02runs against those inputs - THEN the observations match every field of
expectedin caseNGV-F32-CASE-02
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Requirement: NGV-F32-03
r[NGV-F32-03]
NGV-F32-03.
A receipt MUST identify the model, the Lean and Mathlib pins, the proposer, the checker and the clean source revision, and MUST list every proved theorem with its printed statement and axioms. Stale evidence MUST be refused rather than reinterpreted.
Separate claims MUST be maintained for the proved model, tested implementations, trusted correspondence and open obligations. A model-level receipt MUST NOT be reported as implementation or correspondence evidence, and no universal target claim MAY come from one host.
Scenario: NGV-F32-CASE-03 for NGV-F32-03
- GIVEN the
Binary32-Reference-Modelprofile and every field ofinputin NGV-F32-CASE-03 - WHEN the
reviewprocedure for caseNGV-F32-CASE-03runs against those inputs - THEN the observations match every field of
expectedin caseNGV-F32-CASE-03
This is a scenario design, not an execution result. The case's state and evidence fields record its status.
Requirement: FXORACLE-SEM-01
r[FXORACLE-SEM-01]
FXORACLE-SEM-01.
An opt-in finite-effects/v1 reference semantics MUST interpret the existing finite effect-constraint grammar without changing production acceptance, using an independently represented finite universe, exact set union, explicit equations and rigid binders. For a concrete system it MUST define the assignment A(const S) = S, A(union a b) = A(a) ∪ A(b) and A(a) = A(b) for every equation, MUST interpret the solver's result as the unique greatest assignment, and MUST determine satisfiability by bounded exhaustive enumeration. For the rigid lane it MUST separate a rigid valuation from a hole assignment, MUST establish rigid consistency first, MUST report goals-valid only when every enumerated rigid valuation and equation-satisfying hole assignment satisfies every rigid goal, and MUST NOT report an inconsistent system as a vacuously valid implication. Cyclic or otherwise unrepresentable systems MUST be reported unsupported-reference with no semantic verdict, and a production refusal of a semantically valid goal MUST be characterized as a conservative refusal unless the case is in the explicitly declared symbolically exact subfamily.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: FXORACLE-DIFF-01
r[FXORACLE-DIFF-01]
FXORACLE-DIFF-01.
A finite-model differential over the effect solver MUST compare the complete final bound table of every case against an independent reference representation that uses neither production narrowing, symbolic propagation nor the EffSet helpers, and MUST drive production only through its public effect constructors. The outcome matrix MUST distinguish agreement to accept, agreement to refuse, conservative refusal, wrong assignment and unsound acceptance as separate classes; a class with any wrong assignment or unsound acceptance MUST fail the lane rather than being reported as agreement. Exhaustion under a low work budget MUST be reported as exhaustion, not as a verdict. The differential is finite evidence and MUST NOT be reported as a solver correctness theorem.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: FXORACLE-CORPUS-01
r[FXORACLE-CORPUS-01]
FXORACLE-CORPUS-01.
A deterministic bounded corpus for the finite-model differential MUST declare its caps on universe size, holes, nodes, equations, rigid binders and goals, MUST record a reproducible seed schedule, and MUST retain one receipt naming the source revision, toolchain, bounds, seeds, outcome counts, conservative refusals, minimized discrepancies and elapsed work. Cases outside the declared caps MUST be reported unsupported or exhausted rather than silently widened.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: FXORACLE-CTRL-01
r[FXORACLE-CTRL-01]
FXORACLE-CTRL-01.
Every declared test-local faulty variant of the finite-model oracle MUST be caught by a named deterministic case, and the lane MUST fail when a declared variant escapes its expected class. A variant must never be reported as production evidence, and the controls MUST remain test-local: no production behaviour is selected by them.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Requirement: VC-LIVE-01
r[VC-LIVE-01]
VC-LIVE-01.
A kernel-checked live-dispatch reference MUST define bounded roots, ordered entry maps, single-slot publications and an ordered guest trace with separately identifiable control events.
It MUST prove that a root's live selections come from its pinned map, that a single-slot publication changes only its own slot, that execution equals a frozen-entry-map reference for the same request sequence, and that two separate single-slot publications are not one atomic replacement. It MUST state that authorized execution alone carries the event-count conclusion, and MUST keep snapshot consistency distinct from coordinated-version coherence.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Scenario: VC-LIVE-CASE-01 pinning and mixed maps
- GIVEN a root pinned between two separate single-slot publications
- WHEN it selects the slot both publications touched
- THEN the selection is the second publication's generation, while a root pinned before either publication keeps the first, so the pair is mixed rather than atomically replaced
Requirement: VC-LIVE-02
r[VC-LIVE-02]
VC-LIVE-02.
Receipts MUST bind the live-dispatch model to a clean source commit with the inventoried theorem statements and an audited axiom set.
The reference MUST NOT promote canonical live-slot status, and MUST NOT claim runtime correspondence from kernel checking alone.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Scenario: VC-LIVE-CASE-02 receipt scope
- GIVEN the retained kernel receipt
- WHEN its scope statement is read
- THEN it names the clean commit, the theorem count, the live-dispatch theorems, the axiom allow-list and the absence of any runtime-correspondence claim
Requirement: VC-LIVE-03
r[VC-LIVE-03]
VC-LIVE-03.
Implementation comparison MUST bind an existing supported observation or trace rather than inventing a host API, and MUST include a broken per-dispatch-current-map control that exposes the duplicate or omitted event.
An event-count conclusion MUST NOT apply when a denied dispatch or effect stopped execution before the event, and committed prefixes MUST NOT be presented as rollback.
Test design remains open for this requirement. No scenario or execution evidence is supplied.
Scenario: VC-LIVE-CASE-03 relocation witness
- GIVEN two interface-compatible versions where the later one moves a distinguished event from the caller into the callee
- WHEN the broken per-dispatch-current-map control runs
- THEN the mixed selection duplicates or omits the event, and the coherent all-old or all-new entry maps produce exactly one event under successful authorized execution