Rendered from standard/VSTD-Graph-1.md at build time without changing its status. The repository source controls if this presentation differs.

Verifier Standard (VSTD)-Graph-1 — Recorded Lineage#

Acronyms: application programming interface (API); conjunctive normal form (CNF); Davis-Putnam-Logemann-Loveland (DPLL); operating system (OS); Boolean satisfiability problem (SAT); Secure Hash Algorithm 256-bit (SHA-256); satisfiability modulo theories (SMT); Software Package Data Exchange (SPDX); uniform resource identifier (URI).

Reader aid: concept glossary and primary precedents.

Numbered profile: VSTD-Graph-1; required closure coordinate: Recorded Lineage (see LADDER.md) Receipt serialization: schema_version = "VSTD-DATA-0.1" — frozen; see WIRE_IDENTIFIERS.md Status: Project Specification with Implemented Reference Subset Maintainer: TimeLordRaps Date: 2026-08-21

1. Purpose & Core Thesis#

Dataset and training provenance is the foundational substrate of computational verifiability: data sits directly upstream of training runs, checkpoints, fine-tuned adapters, evaluations, model behavior, downstream software products, licensing, and attribution.

VSTD-Graph-1 establishes a content-addressed Hypergraph Specification for capturing recorded and evidenced lineage of datasets, neural weights, and computational outputs within a declared observation boundary. It does not infer unobserved history or prove that the recorded graph is complete in the real world. Transformations are first-class N-ary Hyperedges, which represent many-to-many merges, sharding, and multi-input processing without flattening those relationships into ambiguous binary links.

This document defines the first numbered profile of the Graph axis. VSTD-Graph-2.md through VSTD-Graph-5.md apply progressively stronger object and transformation-edge requirements to the same closed collection. LADDER.md defines the computed candidate Graph profile and its ceiling certificate; the compatibility API verifier.data.graph_level.graph_level implements that computation.

2. The Provenance Hypergraph Abstraction#

A Dataset Provenance Hypergraph is a 6-tuple:

2.1 Artifact Nodes (\mathcal{A})#

Represents any discrete, inspectable data object or model state:

• artifact_id: Unique identifier (e.g. art:sha256:...).
• artifact_type: RAW_SOURCE_FILE, CORPUS, SHARD, DATASET_SPLIT, TOKENIZED_CORPUS, CHECKPOINT, ADAPTER, MODEL, EVALUATION_REPORT, SUBMISSION_ARTIFACT.
• Content-Addressable Cryptographic Digests:
  • content_digest: a declared SHA-256 over raw payload bytes. It becomes a verified byte-identity statement only when a named mechanism actually hashes accessible bytes and binds the observation as evidence.
  • metadata_digest: a declared SHA-256 over explicitly normalized metadata.
  • provenance_digest: a declared SHA-256 over an explicitly canonicalized ancestor subgraph.
• byte_size, record_count, mime_type, storage_uris.
• status: VALID, CHALLENGED, STALE, SUPERSEDED, REVOKED, UNKNOWN.

2.2 Transformation Hyperedges (\mathcal{T})#

Represents a declared N-ary transformation relationship consuming inputs and producing outputs. The edge records ancestry; it does not by itself establish causal influence:

• transformation_id: Unique process identifier.
• transformation_type: COLLECTION, EXTRACTION, FILTERING, DEDUPLICATION, NORMALIZATION, AUGMENTATION, SYNTHETIC_GENERATION, TOKENIZATION, TRAINING, FINE_TUNING, DISTILLATION, QUANTIZATION, EVALUATION.
• inputs: List of input artifact references with role bindings (e.g. TRAINING_SPLIT, BASE_WEIGHTS, CONFIG).
• outputs: List of produced artifact references with role bindings (e.g. CHECKPOINT_WEIGHTS, METRICS_LOG).
• software_provenance: Git repository, commit SHA, branch, clean/dirty state, script path, execution command.
• parameters: Exact hyperparameter dictionary, filter criteria, or random seeds.
• execution_environment: Python runtime, host OS, hardware acceleration class, timestamp.

The frozen VSTD-DATA-0.1 serialization defines artifact and transformation identifiers inside separate collections; historical readers therefore retain a payload in which one string occurs once in each collection. Direct new construction, evidence-bound Graph establishment, and VSTD-GRAPH-ASSURANCE-1 require the two sets to be globally disjoint because current evidence maps and the assurance overlay's subject_id do not carry an artifact/transformation kind. A historical overlap is readable and reproducible as recorded lineage but is inadmissible to those stricter current mechanisms. This compatibility rule does not let duplicates within either collection replace recorded evidence.

2.3 Contributor Nodes (\mathcal{C})#

• contributor_id, name, contributor_type (INDIVIDUAL, ORGANIZATION, MODEL_GENERATOR, AUTOMATED_SYSTEM), uri.

2.4 Rights & Licensing Nodes (\mathcal{R})#

• rights_id, license_spdx (e.g. CC-BY-NC-4.0, MIT, Apache-2.0), commercial_allowed, attribution_required.

2.5 Policy & Formal Constraints (\mathcal{P})#

• Machine-checkable Boolean admission rules. The current reference subset evaluates bounded CNF with its minimal DPLL implementation; general SMT is not implemented.

2.6 Conflict Records (\mathcal{X})#

• conflict_id, subject_id, and predicate identify the disputed coordinate.
• competing_values retains at least two incompatible values.
• evidence_refs retains at least two evidence records rather than selecting a winner.

A conflict record does not mutate the frozen artifact-status vocabulary. It makes the subject inadmissible to a clean candidate Graph profile. The VSTD-Graph-1 receipt has no conflict-resolution transition and remains immutable. The separate non-receipt VSTD-GRAPH-ASSURANCE-1 overlay can record additive, mechanism-checked resolution while retaining the competing evidence. A selected status is projected into that overlay's current view; resolving any other predicate does not by itself establish a clean admissibility effect. No general non-status admissibility-effect mechanism is implemented in the current reference runtime, so such a conflict remains blocking.

3. Provenance Completeness Dimensions#

VSTD-Graph-1 rejects treating a monolithic score as proof. The reference subset reports six descriptive dimensions plus a disclosed weighted summary:

1. Source-declaration coverage (C_{\text{src}}): Share of root artifacts with a non-empty storage URI or source_repository declaration [0.0, 1.0].
2. Transformation-declaration coverage (C_{\text{trans}}): Share of hyperedges with a recorded commit identifier or script path [0.0, 1.0].
3. Content-digest declaration coverage (C_{\text{integ}}): Share of artifacts with a syntactically valid 64-hex-character digest [0.0, 1.0]. This metric does not by itself show that the referenced physical bytes were rehashed.
4. License-metadata coverage (C_{\text{lic}}): Share of root artifacts linked to an explicit rights record [0.0, 1.0]. It is not a legal-validity score.
5. Contributor Coverage (C_{\text{contrib}}): Share of artifacts attributed to identified agents [0.0, 1.0].
6. Downstream Lineage Depth (C_{\text{lineage}}): Integer topological depth from root sources to reachable outputs.

The current weighted summary is 0.25*C_src + 0.25*C_trans + 0.25*C_integ + 0.15*C_lic + 0.10*C_contrib. It is a coverage summary, not a probability, trust score, or verification verdict.

4. Epistemic Incompleteness & Fail-Closed Law#

• The UNKNOWN Principle: If an artifact's status is omitted, or its upstream origin or transformation is not evidenced, the applicable state remains UNKNOWN or the applicable coverage dimension remains incomplete. It never silently becomes observed real-world truth.
• The CONFLICTED Principle: Incompatible retained evidence remains an explicit conflict record. It is neither averaged nor collapsed into UNKNOWN, VALID, or a scalar confidence value.
• Fail-Closed Policy Admission: A policy passes only the Boolean condition it actually encodes. For example, "no ancestor is marked REVOKED" does not establish that every ancestor is VALID; a clean-ancestor policy must explicitly require VALID and reject UNKNOWN, CHALLENGED, STALE, and SUPERSEDED.

5. Challenge & Revocation Blast Radius#

When an upstream source S is marked REVOKED (e.g. due to copyright claim, data poisoning, or corruption):

1. The hypergraph query engine computes the forward reachability closure: $\text{BlastRadius}(S) = \{ a \in \mathcal{A} \mid S \rightsquigarrow a \}$
2. An integrating lifecycle controller can use that returned set to create additive CHALLENGED or REVOKED records. The reference query does not silently mutate historical artifact nodes.

6. Threat Model & Explicit Non-Guarantees#

What the implemented reference subset can establish#

• Receipt integrity: Detects changes to stable fields bound by the receipt's canonical digest.
• Recorded graph structure: Checks references, acyclicity, reachability, and the declared coverage metrics of the stored hypergraph.
• Declared lineage queries: Computes ancestors, descendants, and forward blast radius over recorded edges.
• Bounded policy evaluation: Evaluates the recorded CNF condition over its declared graph-to-variable mapping. This does not prove that the mapping captured every real-world fact.
• Byte identity when separately observed: A named adapter that rehashes accessible bytes can establish whether those bytes match a recorded digest at that observation time. Receipt validation alone does not access unbundled upstream files.

What VSTD-Graph-1 Does NOT Guarantee#

• Real-World Ground Truth: A hash proves byte identity; it does not prove the data is empirically accurate.
• Legal Copyright Validity: A declared SPDX license string records claimed provenance; it is not a judicial copyright ruling.
• Authenticity of declarations: A digest binds bytes or fields; it does not prove that a claimed origin, contributor, execution, or license declaration is authentic.
• Complete real-world lineage: Missing instrumentation, hidden inputs, pre-observation contamination, and out-of-band transformations remain outside the graph unless separately evidenced.
• Automatic physical-file checking: A stored VSTD-Graph receipt validates its own stable content. It flags a physical-file mismatch only when an adapter supplies and rehashes that file.
• Translation completeness: SAT success establishes the encoded formula, not the completeness or correctness of the translation from policy prose or the external world into that formula.