Research Data Contract v0.1

Cycle 1 uses Markdown + YAML frontmatter.

Object namespaces

Q-, C-, A-, TST-, R-, D-, SRC-, RP-.

MVP object types

Question, Claim, Assumption, ResearchPackage, Test, Result, Decision, Source.

Invariants

• IDs unique repository-wide.
• Referenced IDs must exist.
• Status must belong to object type.
• Superseded objects remain stored.
• Public claims require scope and does_not_imply.
• Result statuses are VALID/INVALIDATED; SUPPORTED/FALSIFIED belong to claims.
• No automatic transitive edge inference.
• Invalidated evidence cannot silently remain sole support for a supported current claim.

Projection

Shared content code parses canonical records into:

1. stable/public projection for autorite.org;
2. live/operational projection for autorite.net. Neither website writes canonical research records at runtime.

Implementation profile

This profile completes serialization rules without adding scientific object types, epistemic statuses or relations. Validate living records only; baseline and historical packages are source documents, not duplicate graph records.

Common fields

Every graph record requires schema_version: "0.1", id, type, title, created_at, updated_at, source_refs and relations, followed by a non-empty Markdown body. Dates use ISO 8601; updated_at cannot precede created_at. source_refs contains repository-relative file paths with optional headings; relations contains {relation, target} entries referencing graph IDs.

IDs use the listed namespace followed by an uppercase alphanumeric suffix. Preserve existing labels such as Q0 and RP002A in aliases; use Q-0 and RP-002A as graph IDs. Existing DECISION-0001/0002 documents may be referenced through source_refs by D-0001/0002 records. Aliases must be unique and cannot shadow IDs. Filenames and titles do not determine identity.

Arrays must be explicit, including empty arrays. Required strings cannot be empty; unknown values must be disclosed in the body and cannot satisfy an execution gate. Interpretations that do not fit empirical/formal Claim remain typed editorial prose linked to records. Do not invent a third claim kind.

status is required only for types with a baseline vocabulary; it is forbidden on Assumption, Test and Source. Their absence of status is not evidence of validity. Workflow metadata must not substitute for epistemic status. Unknown keys fail validation so spelling errors cannot silently alter semantics.

Relations

An edge reads source relation target. tests: Test → Claim/Question; supports and challenges: Result/Claim → Claim; depends_on: any graph record → any graph record; derived_from: any graph record → any graph record; supersedes: same type → same type; informs and inspired_by: any graph record → any graph record. These are permitted serialization endpoints, not assertions that an edge is scientifically justified. Every edge requires an explanation in the record body. No self-edges, duplicate edges or cycles in supersedes chains.

Non-edge reference fields also require existing records of the indicated type. question_refs targets Questions; assumption_refs targets Assumptions; package_ref targets ResearchPackage; test_ref targets Test. File references must exist inside the repository and cannot traverse outside it. A Result's Test must belong to the same ResearchPackage. source_refs establishes provenance without generating graph edges.

SUPPORTED/PROVED Claims require a non-empty justification in the body and explicit supporting records or proof artifact references. A validator checks presence and references, not scientific truth. INVALIDATED Results require invalidation_reason and invalidation_decision_ref; retain their original observations and artifacts. A claim losing its only valid support must be flagged for explicit review and must block publication of the current supported projection until resolved. Never promote or demote a claim automatically.

Cycles and runs

Cycles are governance manifests under cycles/cycle-NN/, using the lifecycle already specified by baseline, not a ninth graph object type. Required manifest fields: schema_version, cycle_number, title, status, package_refs, obligations, close_decision_ref, snapshot_ref, publication_refs. The last three may be null/empty before their lifecycle gates. cycle_ref is a manifest path, not a graph ID.

Run metadata and review/publication artifacts are repository files referenced by records, not new graph types. Each run has a unique repository-relative directory, an immutable configuration/seed/environment manifest, and checksums for outputs. Corrections create a new run directory. Invalidation is recorded in a Decision/Result record rather than editing original run artifacts.

Validation boundary

Validate schema, unique IDs/aliases, all graph/file references, status compatibility, relation endpoints, supersession chains, Result/Test package agreement, claim limitations/support, run artifact checksums, and Cycle lifecycle gates. Both site projections consume the same validated records. Scientific review remains explicit and is never inferred from a successful build.