Rendered from docs/ARCHITECTURE.md at build time without changing its status. The repository source controls if this presentation differs.

Verifier Standard (VSTD) conformance architecture#

Acronyms: application programming interface (API); Boolean satisfiability problem (SAT); command-line interface (CLI); identifier (ID); JavaScript Object Notation (JSON); reduced instruction set computer (RISC); Supply Chain Integrity, Transparency, and Trust (SCITT); Verifier Standard (VSTD); zero-identity/zero-knowledge (ZIZK).

Status: implementation and ownership map; normative meaning remains in standard/

TRUST is mechanism-earned forward artifact support; ROT is typed, time-indexed degradation of current admissibility; and RUST is inverse-TRUST diagnostic traversal toward recorded ancestors. They are formal semantic names, not acronyms or actor ratings.

Use this order when two surfaces appear to disagree:

1. normative numbered-profile document;
2. serialized receipt identifier (schema_version) and profile discriminator;
3. published JSON Schema;
4. typed model and validator;
5. conformance tests;
6. generated reference and examples.

A lower item cannot silently redefine a higher item. A passing schema check establishes shape only; a passing validator establishes only its named implemented checks.

Numbered-profile ownership#

Compatibility VSTD-4 and Graph paths still compute candidates from caller-supplied references or ratings and return conformance_status = NOT_ESTABLISHED. Separate evidence-bound paths resolve exact bytes, pin and rerun registered mechanisms, enforce bounds, and recheck certificates before they can report ESTABLISHED. A mechanism result is limited to its proposition, evidence, trust roots, implementation digest, and bounds; the repository claims no external witness or independent implementation.

Artifact control is an orthogonal mechanism beneath the axes. It can preserve and close an artifact used in any numbered profile, but its successful verification establishes only exact-byte integrity, structural closure, and any separately supplied external anchor. It cannot supply a numbered-profile result, semantic correctness, encryption, trusted time, actor trust, or continuous temporal mediation. A thaw sidecar is additive lineage metadata, not parent authentication: sidecar-only agreement remains NOT_ESTABLISHED. Current clean or dirty status requires an actual supplied, cleanly sealed parent whose artifact, content, freeze, seal, kind, and media-type coordinates match. Even that comparison does not authenticate the historical copy operation. Creation paths preserve the caller-supplied final entry through classification: freeze refuses symbolic-link sources, while bundle, descendant, and sidecar creation require an absent lexical destination. These checks do not claim universal race-free filesystem security against concurrent privileged replacement. Authoritative members inside a bundle are stricter: the freeze manifest, payload, seals container, and seal envelopes must have their required ordinary lexical types and cannot borrow bytes through symbolic links or supported reparse-point aliases. That internal closure rule does not reject an outer parent-bundle or explicit thaw-record read alias, whose resolved bytes and bindings are verified. Ordinary hard links remain byte-and-path objects rather than claims of exclusive inode ownership.

Verification complex and profile satisfaction#

VSTD is organized by named closure coordinates, not interchangeable layers or scalar assurance levels. A numbered profile is a cumulative requirement formula over those coordinates. The two axes use different coordinate sets even when their profile numbers match:

Evidence enters through a named mechanism, establishes or fails to establish exact coordinate facts, and is then evaluated against the selected profile formula. A complete assessment preserves ESTABLISHED, REFUTED, UNKNOWN, CONFLICTED, and NOT_ESTABLISHED rather than collapsing absent evidence into false or a satisfiable encoding into conformance. A satisfying assignment over unvalidated caller assertions is only a candidate. Satisfaction suffices only when every satisfying fact is itself bound to evidence by the mechanism that earned it.

“Closure” is always qualified. VSTD-2 surface closure, Graph provenance closure, refutability closure, and artifact-seal structural closure are distinct propositions. “Profile” is likewise qualified as a numbered, receipt, application, or geometry profile when context does not make the category unique. The exact compatibility names and exceptions are normative in standard/LADDER.md.

Operational traversal and recursive Graph materialization#

This non-serialized implementation view does not redefine the numbered profiles:

native computation
  -> VSTD-1 execution and claim capture
  -> profiler or domain adapter
  -> VSTD-2 verification-surface normalization
  -> VSTD-3 substrate accountability
  -> VSTD-4 portable refutation
  -> VSTD-5 independently evidenced witness corroboration
  -> VSTD-Graph collection assessment
  -> content-addressed result artifact
  -> later bounded verification loop

VSTD-2 is the semantic target for adjacent adapters, not the adapter implementation itself. Geometry profiles constrain reusable selections of VSTD-2 geometry; they are connected only by explicit shared coordinates, seams, mappings, and evidence-bearing transformations. The current VSTD-2 receipt has no geometry-profile or profile-composition field, so this relationship is conceptual rather than a new serialized contract. See VSTD-2 section 8.1.

The complete apparatus that constructs or assesses a Graph is a verifying process. In a later order it may become the subject of a new VSTD-2 surface, while an adjacent adapter maps its selected observable outputs into that geometry. Treating the whole apparatus as the adapter would erase the builder, mapping, verifier, and output seams that the next assessment must examine.

A materialized Graph result must retain or bind the source Graph receipt, target collection or induced subgraph, selection query, object and edge ratings with their evidence, selected surface, lifecycle and conflict state, Graph profile certificate, materialization mechanism, and declared information loss. It earns no strength from size, path count, repetition, storage, or agreement. Its result is capped by every applicable member, transformation, mapping, substrate, refutation, witness, and materialization obligation. The compatibility path establishes only its candidate computation. The evidence-bound path can establish a Graph profile only after rerunning every required rating mechanism across the complete closure.

Governing ZIZK architecture and mechanism ownership#

ZIZK artifact-first TRUST is a governing VSTD architecture, not a side experiment, numbered profile, scalar trust system, or actor reputation system. VSTD evaluates bounded validity propositions about computational processes represented by software and evidence-bearing artifacts; it does not determine whether an actor is good or bad. Its normative source is standard/LADDER.md section 1.1.

Zero identity means zero identity-derived verdict weight, not anonymity or absence of identifiers. Checked identity evidence may establish only its exact attribution, authorization, or separation proposition, adjacent to the process claim. Zero knowledge means zero unevidenced knowledge is presumed: absent a mechanism-earned result, the exact proposition remains UNKNOWN. When a witness must remain confidential, cryptographic zero knowledge may enclose this architectural rule by binding the exact program, predicate, public commitments, output, proof parameters, and verifier. The proof remains bearer- and artifact-bound; prover identity or reputation supplies no TRUST. This property applies only where a named proof system establishes it under explicit assumptions.

TRUST, ROT, and RUST are formal semantic names, not acronyms, serialized receipt values, actor ratings, scalar scores, or references to the Rust programming language:

• TRUST is mechanism-earned artifact support moving forward only across admissible bound transformations while every child discharges its new obligations;
• ROT is typed, time-indexed degradation of current admissibility without rewriting historical evidence; and
• RUST is the inverse-TRUST diagnostic mechanic moving backward from an observed descendant deviation through recorded contributing ancestry.

The three never cancel, reverse direction, or manufacture a clean signal from UNKNOWN or CONFLICTED inputs. Identity, popularity, repetition, age alone, topology, and propagation supply no assurance.

Maturity attaches to mechanisms beneath that architecture:

Serialized receipt dispatch#

Dispatch first by schema_version, then by the required profile discriminator. VSTD-1 claim receipts require receipt_kind = "claim_mechanics"; generic-run receipts require receipt_kind = "generic_computational_run". Unknown identifiers, absent discriminators, and mismatched shapes fail closed. The current reader does not infer a profile from retired pre-current-profile field arrangements.

Installed specification ownership#

Every standard/*.md file has a byte-identical installed resource under src/verifier/specifications/. Verifier descriptors use those resources when no source checkout is present. The installed-wheel gate runs outside the checkout and rejects an unavailable specification digest.

The artifact-control schema has a byte-identical installed copy under src/verifier/artifact_control/. GitHub Pages publishes it at its declared /schemas/ route alongside receipt schemas without reclassifying it as a receipt.

Generic-run validation contract#

vstd validate is an integrity/profile validator for the generic_computational_run profile. It enforces the strict receipt shape and recomputes the stable-payload digest. The dynamic path keys inside source_state.source_file_hashes, unconstrained recorded evaluator values, and additional declarations inside assessment_context.refutation_surface are explicit data or extension surfaces. The refutation surface remains open for caller-defined domain refutations, but every additional value remains a declaration until an applicable mechanism checks it. Unknown object properties outside those named surfaces fail closed.

Validation does not rehash referenced artifacts, rerun the command, resolve evidence references, or establish that recorded declarations are true. Those are separate mechanisms. validate, inspect, and reproduce honor --json for generic-run and VSTD-Graph receipts; the envelope reports command completion without upgrading the receipt's claim semantics.

Generic-run assessment context#

assessment_context is a required VSTD-1 generic-run container, not a VSTD-4 object. It participates in the canonical digest and retains the manifest-declared mechanism, resource-bound, commitment, and refutation coordinates without using a numbered-profile identifier as a generic container name.

Closure coordinates identify assessment questions; they are not containers for generic verification context. The neutral container must not generate profile-numbered binding structures. Nothing in assessment_context supplies a VSTD-4 result.

Five-As human traversal#

Status: non-serialized architecture guide. The five As are roles in a human traversal of existing VSTD records, not five new object types, numbered profiles, statuses, or a scalar assurance score. No receipt or schema format is defined here.

ASSURANCE_0 -> ATTRIBUTION -> ASSIGNMENT -> ASSESSMENT -> ASSURANCE_1 reads as follows:

First-hand and second-hand describe provenance, not strength. A first-hand self-observation may be weak; a second-hand certificate may be strongly bound to a narrow proposition. EvidenceClassification records how evidence entered a profile, but its name, source, or placement never substitutes for the profile's verification mechanism.

The smallest operational loop is:

1. select one proposition and retain every applicable input state, limitation, conflict, unknown, trust root, and freshness bound;
2. bind each input to that proposition through an inspectable attribution, preserving transformations and declared information loss;
3. record Assignment only to the specificity evidenced, leaving absent coordinates UNKNOWN;
4. run the named assessment mechanism under its specification and bounds; and
5. record the output as a new evidence artifact and transformation, without changing any input record. A later loop may consume that output only as lineage-preserving input to another explicitly identified assessment.

No semantic strength is gained by storage location, field name, repetition, graph multiplicity, actor reputation, or propagation. Every increase in assurance must identify the verification mechanism that earned it.

This is the human forward traversal of the same topology VSTD-Graph stores for machines. TRUST is bounded, mechanism-earned support across an admissible recorded transformation; the child still discharges its new obligations. ROT is typed, time-indexed degradation of current admissibility while historical evidence remains immutable. RUST is inverse-TRUST diagnostic reachability from a downstream deviation toward recorded ancestors. Together they describe memetic causal-provenance and lifecycle behavior over one development graph. They do not establish actor standing, moral character, responsibility, or automatic ancestor falsification. The reference runtime emits bounded TRUST, ROT, and RUST events only after their exact mechanisms run. Causal localization and artifact-relative BLAME or GUILT require additional exact propositions. BLAME establishes bounded responsibility or material contribution. GUILT is not directionally opposite: it composes three separately bound passing evaluations for responsibility, applicability of an exact scoped obligation, and violation of that same obligation relative to the same localized deviation. The final evaluation binds all three component digests. One compound mechanism may emit those three evaluations in one invocation, but one opaque combined result cannot substitute for them. Neither becomes actor reputation, moral character, automatic legal liability, or social scoring; missing GUILT does not establish innocence or exoneration.

Recursive-amplification falsification outcomes#

Separation and Graph boundaries#

The historical independent_audit field name does not prove independence. Its independence_basis records actor, implementation, and runtime separation. Repeated or matching results are artifact agreement, not evidence that separate actors performed the runs; absent separation evidence is NOT_DEMONSTRATED. Serialized status words and evidence-reference strings cannot self-promote that result. The generic-run compatibility path treats supplied assertions as no stronger than DECLARED, rejects receipts that label them EVIDENCED, and never derives EVIDENCED. The distinct VSTD-5 reference path can establish only its exact declarant/witness separation propositions after all seven seams are rerun by registered mechanisms. Typed binding, identity, separation, and corroboration errors keep computed_independence fail-closed; no error-message text is interpreted as a semantic category. Witness identities and independence assertions serialize as separate ordered arrays, so duplicates, orphan assertions, and missing cardinality survive receipt build and replay rather than disappearing inside a keyed map. Permissive assessment remains separate from portable-record admission: the builder raises unless the result inhabits the strict VSTD-5 schema and embeds every verdict-material evidence byte, and the rechecker enforces the same zero-dependency shape gate before replay. It also compares every redundant VSTD-4 entry coordinate and treats corroboration_class as part of the mechanism-checked expected proposition. The admitted evidence-bound VSTD-4 result retains its exact claim_id, and VSTD-5 requires the bundle to use that identifier; a shared binding or certificate digest cannot turn a neighboring identifier into an alias. Thus relabeling any of these coordinates cannot retain an established result. This path does not upgrade the legacy generic-run fields.

Graph conflict records retain incompatible values and their evidence references without adding a scalar score or changing the frozen artifact-status vocabulary. A conflict makes the subject inadmissible to a clean candidate Graph profile. The frozen VSTD-DATA-0.1 reader retains its historical separate artifact/transformation namespaces. Direct new construction, evidence-bound Graph establishment, and the assurance overlay require global cross-kind disjointness, so an untyped subject_id cannot ambiguously name both kinds of Graph object.

Recursive current-state audit#

Historical receipt bytes and their recorded PASS remain unchanged. A later current-state question is a new assessment over the retained graph and applicable lifecycle records:

The forward blast-radius query remains discovery only. AssuranceLedger is the distinct binding mechanism for explicit edge-local TRUST, ROT, RUST, challenge projection, and conflict declaration/resolution. recheck_assurance_log reconstructs the historical Graph, rehashes every embedded evidence item, reruns the exact registered mechanisms, reproduces the event chain, and compares the derived current view. The ledger never infers an unrecorded edge, converts topology into assurance, or treats RUST as causality. Domain-specific transfer and localization propositions still require their registered mechanisms and may return UNKNOWN.