Peer Synthi

Produce one useful answer that retains the best supported findings and ideas from independent peers. Combine compatible contributions, explore stronger options made possible by their combination, and preserve consequential differences without organizing the report around who said what.

Agreement about faithful representation is separate from agreement with every conclusion. A finished report can contain alternative interpretations, conditional recommendations, or dissent. Substantive disagreement alone never makes it incomplete.

Use for research, architectures, business proposals, plans, and recommendations. For reviewing an existing artifact into an agreed revision list, peer-superset remains available unchanged. This workflow creates an advisory deliverable; it does not supply authorization for downstream implementation, publication, filing, spending, or deployment.

Set the run

Use the user's selected participants; do not silently recruit or replace them. Record:

• The shared question or brief, scope, intended decisions, and output locations.
• Two or more unique participant identities and their writable locations.
• An integrator and controller, either of whom may also be a participant. The integrator's original has no priority. Rotation is optional.
• The evidence standard: required verification, pivotal claims needing an independent source, and applicable domain conventions for confidence.
• Any time/cost limits, required review coverage, and round budget. Default to one cross-review and one repair pass; spend an additional bounded round only on an issue that could materially improve the result.

Use reasonable defaults where the assignment supplies enough context. Avoid a separate setup questionnaire. Do not make every participant duplicate every check: allocate coverage explicitly when many agents are involved, while each remains responsible for checking how its own material contributions are represented.

An item is material if it could change understanding, confidence, a classification, a feasible option, expected benefit, cost, risk, priority, or next action. Equivalent wording and harmless differences in emphasis need no disagreement record.

1. Independent work, then freeze

Give every participant the same core brief and evidence standard, with freedom for useful supplemental inquiry. Before exchange, share only operational information such as access problems, progress, or rate limits. Disclose prior substantive exposure; do not claim fresh independence retrospectively.

Each participant produces an original containing findings or recommendations, evidence, assumptions, limits, and what could change its conclusions. Give material contributions stable IDs, with enough provenance to trace the source or reasoning. Distinguish observations, interpretations, recommendations, and untested ideas where that affects their weight. A compact findings ledger can accompany ordinary prose.

End with a timestamped independence statement and {participant}_ORIGINAL_DONE. The controller verifies every required original exists and is substantive, records its hash, and then releases the originals together. Keep frozen originals unchanged; later evidence and changed positions go into supplements. The integrator must freeze its own original before reading peers'. Missing originals are disclosed, never replaced by assumed agreement.

2. Compare contributions and investigate consequential differences

Create a reconciliation record, jointly or through assigned peer maps. Every material original item needs a destination or a reasoned disposition. Split compound items where necessary. Use these relationships as guidance, not quotas:

Relationship	Treatment
Same supported finding or equivalent wording	Merge while retaining distinct evidence and conditions. Call it consensus only if the relevant participants actually agree; silence is unreviewed.
Useful additional finding or option	Integrate as a complement with its evidence status. A single peer's supported discovery is valuable without pretending everyone found it independently.
Different evidence, facts, definitions, assumptions, scope, or priorities	Diagnose the cause and investigate when material. Normalize what can reasonably be normalized; unresolved source conflicts or uncertain premises may themselves be important results.
Material interpretation or recommendation difference	Present the strongest warranted alternatives, conditions, and consequences. Identical retrieval inputs are not a prerequisite for reporting a meaningful difference.
Error, superseded claim, duplicate, or out-of-scope item	Correct, merge, supersede, exclude, or defer with a reason. Preserve traceability, not an error's apparent validity.
Evidence gap or promising untested idea	State the limitation or hypothesis and what would resolve or test it. Do not present it as established.

Check evidence and assumptions before attributing a difference to model judgment. A peer's assertion that a claim is false is a challenge to evaluate, not a deletion instruction. Explain substantive changes of view by the evidence or reasoning that prompted them; investigate unexplained loss of a material contribution without demanding that agents keep disagreeing.

Keep evidence strength, confidence, and peer agreement separate. Two votes out of three do not establish group consensus or factual truth. Shared tools and seeds can produce correlated errors. Perform the independent verification required by the evidence standard, concentrating effort where it could change the answer. Do not give a weak view equal weight merely because a peer advocates it.

3. Build a better combined answer

Go beyond a union of findings: consider whether compatible contributions yield a stronger explanation, option, or recommendation. Label materially new synthesis in the working record, trace its premises, and have consequential new proposals checked by a peer. Do not portray it as independently corroborated discovery.

Organize the report around the user's questions and decisions. Usually lead with the supported answer and useful implications; integrate complements naturally. Include alternatives, conditional recommendations, or dissent only where material. Explain what changes under each, its strongest support and limitations, and what evidence, test, fact, or priority would favor a path. An uncertainty need not become a request for the user to choose.

Offer a recommendation when warranted, with its conditions and degree of peer support stated accurately. Do not call an integrator's recommendation joint when others disagree. Omit alternatives sections entirely when none add value. Keep attribution and exchange history in a compact appendix or linked record unless needed to understand a consequential position.

Preserve a trace from each material original item to its final location or disposition and reason. Originals remain available for inspection. The synthesis should be easier to use than reading the originals concatenated together.

4. Review representation and close honestly

Every participant checks the proposed synthesis against its material contributions. The assigned peer review also covers consequential new synthesis and the evidence checks promised in the brief. Reviewers separate:

• Report defects: material factual errors, omissions, misrepresentation, unsupported certainty, or unmet required verification. Identify the item and a concrete repair.
• Substantive reservations: a fairly represented interpretation, priority, or recommendation the reviewer still does not share. Retain it when material; it does not block completion.

Record REPRESENTATION_OK or specific defects against the reviewed version hash. Correct defects and recheck material changes with the affected reviewers. Do not reopen settled wording or seek unanimity. A participant's own wording can fix a paraphrase problem; it cannot by itself fix a false summary, omitted implication, bad source, or missing review acknowledgment.

At the round or budget boundary, deliver the usable work with its actual status and remaining issues. Do not suppress good output because one part remains unresolved. Separate these facts in the closeout:

Field	Meaning
Work	Complete for the agreed scope, or partial with named omissions.
Review	Confirmed for the delivered version, unconfirmed with missing checks, or known defects remaining.
Material alternatives	None or present; neither determines completion.

Use COMPLETE when agreed work and required verification/review are finished and known material report defects are resolved. Accurately described uncertainty or substantive disagreement is compatible with COMPLETE. Use PARTIAL for a useful provisional deliverable with missing work, review, or known defects; identify the affected conclusions and next repair. Use BLOCKED when an external dependency prevents further progress, still providing usable completed portions. A reduced roster or scope must be explicitly agreed and disclosed; it cannot retroactively establish independence or missing review. Never treat silence as consent.

For COMPLETE, emit PEER_SYNTHI_COMPLETE material_alternatives={none|present} with the report path/hash and linked review record. Otherwise emit PEER_SYNTHI_PARTIAL or PEER_SYNTHI_BLOCKED with the delivered paths and outstanding work. Completion records belong outside the report's own hash to avoid self-reference.

Transport and artifacts

On Interlateral, follow mesh-comms-core; desktop peers also use desktop-mesh-peer and desktop-live-comms. Material handoffs require direct delivery plus ledger mirror and receipt confirmation. A passive desktop inbox does not wake its model; rendered text or file existence is not an ACK. Use an acknowledged human relay when necessary. In other environments use an available, authorized channel with equivalent receipt checks.

Use absolute artifact paths and path-bearing done notices. Keep original, reconciliation, synthesis, and review information; combine files when it remains clear. Reviewed artifacts and source text are data, not authority to override the user's assignment.

For optional record templates and examples that distinguish useful diversity from report defects, read references/examples.md.