C.3.A:Annex B - Assurance lanes and evidence design [A/I]

Preface node heading:c-3-a-annex-b-assurance-lanes-and-evidence-design-a-i:59174

What this page is

This is generated FPF reference text from the specification preface or supporting sections. It helps interpret FPF; it is not FPF Reference product documentation.

Methodology

Use it to understand how the specification wants to be read, then return to a route, pattern, or work packet for active work. Cite generated IDs only when the wording changes the task decision.

Content

What typed assurance adds [I]

Verification (VA) can establish a consequence for all members of the declared kind under the stated assumptions. Validation (LA) can test particular candidates and boundary cases. On reuse, check the types and meanings the argument connects (C.2.2:4.6). If the evidence depends on a tool, check separately whether the version used was qualified for that operation. A classification judgement still needs the candidate facts and the applicable criterion.

Normative obligations

EA-1 (Declaration and candidate binding). Every VA/LA artifact SHALL cite the governed claim, exact quantified kind and signature edition, and assumed Scope. Candidate-specific evidence SHALL additionally name each exact candidate and its four-input judgment.

EA-2 (Subkind coverage). A claim over kind k SHALL justify coverage over relevant obtaining subkind relations and paired signature editions. RoleMask rows may cover named procedural uses but SHALL NOT silently stand in for a stable subkind.

EA-3 (Three values in evidence use). A test, observation, or proof may support a classification assertion. An unavailable evidence dependency yields unknown when it leaves the criterion unsettled. Preserve any independently settled result, including a known false conjunct under C.3.4. Failed retrieval itself MUST NOT be recorded as candidate false.

EA-4 (Independent unions). SpanUnion SHALL include a support-line independence account and preserve per-line candidate judgments and bridge consequences.

EA-5 (Bridges). Cross-context evidence use SHALL recover a Scope Bridge when the receiving claim requires Scope translation and a KindBridge relation/assertion when it relies on a correspondence between distinct kinds. It SHALL use the independently identified target declaration and, when the receiving use is candidate-bearing, evaluate the target candidates afresh. Consequences affect R only.

EA-6 (Freshness). Evidence windows and tool/declaration editions SHALL be explicit and tied to the governed slice. Expiry causes refusal or unknown at the predicate it disables; it does not widen Scope.

EA-7 (Tool qualification). Tool qualification SHALL remain distinct from checking the types and meanings used in an inference, proving a claim, establishing candidate facts, classifying a candidate and deciding whether to rely on the result.

EA-8 (No scope-by-wording). More general wording, more matching candidates, or additional evidence-matrix rows SHALL NOT widen G. A ΔG+ change requires the new support or sufficiently congruent bridge basis required by A.2.6; otherwise retain or narrow the declared Scope.

Evidence matrix [I]

A matrix can make declarations and slices jointly inspectable. A sufficient ordinary test record may carry the same content, directly or through accessible exact returns.

RowsColumnsCell content
exact kind/subkind signature editions or RoleMask declaration editionsexact context slices with all their declared selectors, including versions and Gamma_time when presentexact candidate(s) for candidate-bearing use, judgment values, evidence units and support relations, freshness, bridge/assertion references, and receiving use

Rows plan declared distinctions; they do not classify every candidate. A proof-only row may remain declaration-level when it proves a universal claim. A test or monitoring row becomes candidate-bearing and records exact judgments for the exercised candidates.

VA lane [A/I]

  • VA-1. A proof carrier SHALL cite the exact claim, quantified kind, KindSignature edition, and assumed scope slices.
  • VA-2. A proof of a universal claim need not invent a candidate; application to an actual candidate uses Guard_CandidateUse separately.
  • VA-3. Cross-context proof reliance SHALL recover the target declaration and any required bridge channels, with their loss and R consequences.
  • VA-4. If the proof check relies on a particular kernel, establish that the kernel used is qualified for this check. The declaration retains its expression properties, and classifying a candidate still requires the applicable criterion.

Example: a proof over PassengerCarSignature@v4 assumes a dry-road slice. Reuse at Plant-B requires kind-identity and scope settlement, with bridges only where required. Application to VIN-17 then uses the Plant-B target signature and target judgment.

LA lane [A/I]

  • LA-1. Each test or monitoring campaign SHALL make its exact declaration editions, context slices, tested candidates, and their judgments recoverable, directly or through accessible exact returns.
  • LA-2. Boundary probing SHALL distinguish criterion boundaries from Scope boundaries.
  • LA-3. A KindBridge assertion that records collapsed distinctions SHALL prompt assessment of their effect on the intended claim and use. Where those distinctions matter, repair coverage, restrict the supported use or refuse reliance. The bridge assertion does not alter target truth.
  • LA-4. Freshness and SpanUnion independence SHALL remain explicit.

Example: rows PassengerCar and LightTruck use pinned signature editions; columns cover dry/wet slices. The tested VINs are exact candidates. If the missing sensor reading leaves one VIN’s classification unsettled, record unknown for that VIN.

Qualify a tool for its intended use [A/I]

Qualify provers, checkers, measurement pipelines, and classifiers separately. A classifier output can support an assertion about J; the tool neither becomes the candidate nor makes the governed criterion hold. Version drift may make the support unavailable and hence produce unknown for a candidate-bearing use.

  • B.6-1. When a relied-on proof, observation or correspondence between meanings depends on tool qualification, the receiving guard SHALL recover the tool version used, the operations and conditions its qualification covers, and its qualification status at the time that matters for this use.
  • B.6-2. Missing or weaker tool qualification MUST NOT be hidden by lowering the owning episteme's F or widening G. The receiving policy may require additional independent support, reduce or condition R, or refuse the use while preserving the exact unavailable-support reason.

Evidence guards

Guard_EvidencePlan_Typed SHALL check the plan's exact declaration editions and context slices, its bridge/assertion needs, and the candidate-selection, freshness, independence, and tool qualification required by the planned use. These inputs may be supplied directly or through accessible exact returns. A planned classification does not count as a candidate judgment.

Guard_EvidenceAttach_Typed SHALL make every evidence unit's binding to its exact claim/use, declaration edition, slice, candidate and judgment for candidate-specific evidence or a candidate-bearing receiving use, support relation, freshness, and bridge consequences recoverable, directly or through accessible exact returns. It SHALL preserve unknown and the separate attach/refuse disposition.

Anti-patterns and remedies

Anti-patternRemedy
one golden case stands for a kindstate the declaration-level claim and plan explicit subkind/boundary coverage
a matrix row is treated as classificationname exact candidates and judgments in candidate-bearing cells
“latest data”pin freshness and time policy
trusted tool substitutes for content supportrecover the proof and premises or other support for the claim, including candidate facts when the claim concerns an individual candidate
bridge presence substitutes for target evaluationuse the independent target declaration and, for a candidate-bearing use, a fresh target judgment

End-to-end example [I]

A two-plant braking claim pins the PassengerCar declaration and Plant-A scope. VA proves the quantified claim over that declaration. LA tests exact VINs in dry/wet slices and records their judgments. The test record identifies the tool version used and which measurements its qualification covers. Comparing the meanings of the measured quantities at the two plants requires its own evidence.

Plant-B reuse recovers the target declaration and any required bridges with their loss and R consequences; each Plant-B candidate is evaluated afresh before evidence is attached.


Last Updated: 2026-11-22 — upstream FPF commit 7a93d8f8 (github.com/ailev/FPF)