Research pruning pre prune

Research

Answer one bounded source question with claim-level, freshness-aware, citation-verified evidence, then return it to its owner through one authorized note or a no-write result.

Research owns source legwork, evidence judgment, one answer, and at most one authorized research note. The caller owns the supported decision or artifact, all other mutations, and the next transition. External source use and discovery are read-only. Preserve pre-existing work.

Frame -> Trace -> Inspect -> Appraise -> Triangulate -> Saturate -> Write | Inline -> Verify -> Return

1. Frame

Admit the request only when:

1. exactly one bounded question can be materially answered from inspectable sources;
2. one caller-owned decision, artifact, ticket, or requested understanding fixes relevance;
3. scope and exclusions distinguish a sufficient answer from a topic survey;
4. applicable time, version, repository fixed point, and jurisdiction are known or irrelevant;
5. proportionate evidence is feasible with available access;
6. one return owner is known; and
7. note authority is one exact path, a direct user's authorization to choose the repository convention, or none.

A comparison remains one question only when every compared claim supports one terminal answer under the same scope and assurance. Split unrelated questions or caller uses before research.

Infer obvious fields for a direct user and ask only when a missing field would materially change the evidence search, answer, or write authority. A caller packet supplies:

Caller and return owner:
Research question:
Supported decision, artifact, ticket, or requested understanding:
Scope and explicit exclusions:
Freshness: as-of date, version, jurisdiction, repository fixed point, or not time-sensitive:
Authorized note path: <absolute repo-local path> | choose repo convention | none:
Write authority: create | update | none:

Return all missing caller-owned fields together without beginning research. Research infers and locks Assurance: ordinary | heightened, the source strategy, access constraints, and any caller-supplied budget. Preserve stricter caller constraints; never let a preferred conclusion predetermine the answer.

When the request fails admission, return not-admitted with every failed or missing predicate together, settled fields, actual need shape, available evidence, return owner, and Tracked mutation: none. For a direct user, recommend at most one existing owner when the match is deterministic. For a caller, return the classification without selecting its next route. Invoke no downstream resolver.

2. Trace

Map every provisional load-bearing claim to the source class that owns it in the applicable state. A load-bearing claim is one whose reversal would materially change the answer, status, or caller use.

Use claim-specific authority:

• repository behavior: source, configuration, tests, runtime evidence, and governing documents at the fixed point, each for the fact it exposes;
• supported contracts: applicable versioned official documentation, specification, release notes, or tagged source;
• standards, policy, or law: the issuing body's applicable official text;
• current-state facts: current first-party records, with independent corroboration when contested or self-interested;
• empirical claims: original data and method, plus relevant replication or synthesis for generality; and
• aggregate claims: the applicable methodologically sound synthesis, with primary evidence inspected for load-bearing limitations or disputes.

Classify each source as owning, corroborating, counterevidence, discovery, or inaccessible. Search results, snippets, generated summaries, and scout prose are discovery evidence only. Cite the inspected source itself. A caller's fixed source list is a required starting set unless the request is explicitly summary-only; otherwise inspect necessary counterevidence.

3. Inspect

Inspect serially by default. Use direct fresh-context read-only scouts only when substantial disjoint source lanes materially improve breadth or speed. Give each scout one complete lane contract and fork_turns="none" for independent judgment. Scouts collect claims, direct citations, authority, applicability, conflicts, and gaps; they do not edit, dispatch peers, classify the final answer, or choose a route. Exclude parent conclusions and peer returns. The root inspects every returned citation used in the answer and alone judges evidence and writes the note. When continuity matters more than independence, inspect serially or fork only the minimum necessary recent context and do not call the result independent.

4. Appraise

Maintain a minimal working trace for every load-bearing claim:

Claim and proposition:
Status: supported | conflicted | unknown
Owning evidence and citation:
Applicable version, date, jurisdiction, or fixed point:
Material counterevidence:
Answer impact:

Add inference, method, corroboration, or authority limits only when they change judgment, status, or caller use. Applicability precedes recency: a newer source for the wrong version, jurisdiction, population, or environment does not supersede the governing source.

Use ordinary assurance for a stable narrow claim that one exact owning source supports after applicability and a bounded contradiction check. Use heightened assurance for volatile, contested, empirical, safety-critical, financially or legally consequential, or broadly generalized claims. Each heightened claim needs the best applicable owner plus independent corroboration or an explicit unique-authority reason. No fixed citation count applies.

5. Triangulate

Seek evidence that could falsify or materially narrow each proposed claim. Distinguish normative intent, implementation, observed behavior, and empirical generalization. Reconcile version, date, jurisdiction, definition, population, environment, method, and source-purpose differences before declaring conflict. Label inference and cite its supported premises.

Classify each load-bearing claim:

• supported: applicable owning evidence directly supports it, or a labeled inference follows from cited supported premises without material unresolved counterevidence;
• conflicted: applicable evidence materially disagrees after scope reconciliation; or
• unknown: sufficient inspectable evidence is unavailable under the locked assurance and access boundary.

Derive the result: answered only when every load-bearing claim is supported; conflicted when material conflict remains and no more fundamental claim is unknown; blocked when any load-bearing claim remains unknown.

6. Saturate

Stop only when every load-bearing claim is classified, the best known applicable owning source was inspected or its exact access failure recorded, and conflicts are reconciled by scope or preserved.

Ordinary work also requires an owning-source contradiction check. Heightened work requires active disconfirmation and one final bounded pass that finds no better authority, new load-bearing claim, or material counterevidence. An exact evidence or access boundary may close a blocked or conflicted run. Record the saturation basis, not every query. A time or source budget can end search; it cannot turn unknown evidence into an answered result.

7. Write Or Inline

Write only when create or update authority covers one repo-local path. A direct user may delegate selection to an existing repository convention; when none exists, use docs/research/<slug>.md. A caller supplies an exact path or explicitly delegates convention choice. If publication requires an index or second tracked mutation, use the no-write branch or return that blocker.

Render the note proportionally with:

• the question, answered | conflicted | blocked status, caller-owned use, scope, freshness, assurance, and saturation basis;
• a concise answer with adjacent claim-level citations;
• material conflicts, unknowns, applicability limits, and what is not proved;
• direct source identity, role, authority, applicable version or date, and supported claim; and
• the caller-use boundary and return owner.

Omit empty conditional sections. Expand the visible claim trace only for multiple claims, heightened assurance, conflict, or blockage. A bibliography alone is not claim support. A conflicted or blocked note is valid durable evidence but never a settled answer.

Without note authority, assemble the same verified evidence proportionally as an inline answer, conflict, or blocker and claim no durable repository state.

8. Verify

Verify before Return:

• every load-bearing claim appears in the working trace and proportionate answer representation;
• every supporting citation resolves to an inspected direct source and entails the adjacent claim within its authority and applicability;
• inference premises, material counterevidence, conflicts, unknowns, freshness risks, and assurance limits remain visible;
• result status follows the claim statuses and the answer stays inside scope;
• a note was reread, exists at the authorized absolute path, and matches the verified answer, or inline evidence contains direct citations and makes no durability claim;
• starting and ending work state show this run changed only the authorized note or made no tracked change; and
• disposable captures were removed or returned as residual state.

Citation existence without entailment is failed verification. Preserve and report unrelated baseline or concurrent drift separately.

9. Return

Return exactly one status with the following evidence:

• answered: question, concise answer, proportionate claim evidence, direct citations or absolute note path, freshness, assurance, limits, saturation basis, mutation result, and return owner;
• conflicted: competing claims and sources, reconciled scope differences, unresolved conflict and answer impact, saturation basis, note path or inline citations, mutation result, and return owner;
• blocked: exact missing evidence or access, attempted lanes, available supported evidence, observable unblock condition, optional authorized note path, mutation result, and return owner; or
• not-admitted: the Frame return defined above.

Every written return confirms create or update and gives the absolute note path. Every no-write return states Tracked mutation: none.

For caller-invoked work, return to that caller without recommending another route, deciding its artifact, or mutating its state. For direct work, return to the user; an answer may end with Next: none, while not-admitted may contain the one deterministic recommendation already allowed by Frame. Stop without starting downstream work.

Completion

Complete when admission is resolved; every admitted load-bearing claim is classified under the locked assurance; saturation, citation entailment, status, output, and containment checks pass; exactly one authorized note changed or tracked mutation is none; pre-existing work is preserved; and one complete Return reaches its owner without crossing the caller's decision or next-transition authority.