Overpowered / Know Enough

Core rule

Know enough to act well — no less, no more. Retrieval must have an information objective.

Before each search, be able to complete:

I am looking for [information] because it could change [decision/action].

If you cannot complete that sentence, do not retrieve yet.

Procedure

1. Define the next decision. What are you trying to decide, produce, or verify now?
2. Inventory current knowledge. Separate known facts, assumptions, and material unknowns.
3. Prioritize gaps. Retrieve only gaps that can materially change the next decision. If existing evidence suffices, do not search.
4. Choose the source by role. Prefer an available knowledge-source registry. Distinguish:
   • authoritative / normative;
   • precedent / historical;
   • observational / operational;
   • reference / explanatory.
5. Plan minimal acquisition. Check source access and available functional capabilities, then choose the simplest sufficient strategy using references/adaptive-retrieval.md. Search for the missing fact or evidence, not the whole topic; the harness selects existing means and executes the plan.
6. Evaluate results. Check relevance, authority, freshness, scope, coverage, and contradictions. Preserve source identifiers, passages, and citations; distinguish no result, inaccessible source, conflicting evidence, and insufficient evidence.
7. Iterate only if needed. Take one targeted step to close a material gap or resolve a conflict, using available reconciliation support when appropriate.
8. Stop when sufficient. Judge sufficiency against the task and risk. Additional retrieval should be unlikely to change the next decision; if available means cannot close a material gap, report the limit instead of looping or inventing evidence.

Adaptive retrieval

Read references/adaptive-retrieval.md when a material gap requires choosing an acquisition strategy. It covers direct, structured, lexical, semantic/hybrid, relational, preparation, and optional classification capabilities without prescribing implementations.

For a non-trivial plan, briefly state: knowledge objective → desired capability/capabilities → accessible sources → expected evidence → stopping criterion. Use ordinary language, not a required machine schema. Prefer rules or the active model's reasoning for simple choices; specialized classification is never required. Combine capabilities only when the question and results justify it.

Retrieval budget

Default to the smallest useful search sequence. Expand only when:

• first results are ambiguous or weak;
• a material conflict appears;
• coverage is insufficient for the decision;
• a different evidence role is required, such as current policy plus precedent.

Output contract

Keep a compact knowledge state when the task is non-trivial:

Decision to support
Known
Missing but material
Sources consulted + role
What was learned
Remaining uncertainty
Sufficiency: ENOUGH / NOT ENOUGH

If NOT ENOUGH, say exactly what is missing and whether progress can safely continue.

Knowledge-source registry

Read references/knowledge-source-registry.md when multiple knowledge bases or retrieval tools exist. Use references/knowledge-sources.example.yaml as a portable example.

Gotchas

• Do not retrieve everything because retrieval is available.
• Do not ask the user for a fact that an available source can cheaply establish.
• Do not treat semantic similarity as authority.
• Do not treat precedent as current policy.
• Do not keep researching after the decision is already robust to additional information.
• If the task is purely transformative and all required content is present, do not retrieve.

Composition

These are possible companions when available; the harness or orchestration layer selects installed means, not this knowledge policy.

• Need historical analogs → find-precedent.
• Sources materially disagree → reconcile.
• Needed facts live in structured files → ask-the-data.
• Material claim must be supported → evidence-first.

Stop condition

Stop when the next material decision can be made with explicit uncertainty and more retrieval is unlikely to change it.