Decision & Output: outcome + proof from history

[functor-flow]

Decision & Output: outcome + proof from history

full spec at spec

flowchart TD
  Task["SearchTask"]
  Hist["SearchHistory (aggregates + hits)"]
  Attempts["current SERP attempts"]
  Max["max SERP attempts"]

  subgraph DecisionMod["Decision module"]
    Score["compute support/refute scores"]
    Enough{"evidence >= threshold?"}
    Direction{"support > refute?"}
    MoreSearch{"can we search again?"}
  end

  Result["AgentResult (True / False / Invalid)"]
  NeedMore["need more SERP search"]

  Task --> Score
  Hist --> Score

  Score --> Enough
  Enough -- "yes" --> Direction
  Enough -- "no" --> MoreSearch

  Direction -- "support" --> Result
  Direction -- "refute" --> Result

  Attempts --> MoreSearch
  Max --> MoreSearch
  MoreSearch -- "yes" --> NeedMore
  MoreSearch -- "no" --> Result

Loading

Status: Not implemented for the new agent pipeline (existing validator decision code is separate).

Role. Look at SearchTask + SearchHistory, decide whether we already have enough evidence to answer True/False, or whether the claim should be treated as Invalid, and tell the loop if another SERP round is worth running.

Responsibilities.

• Implement a decision module that reads only SearchTask, SearchHistory, and simple config (evidence thresholds, max SERP attempts).
• Aggregate support and refute scores from SearchHistory.evidence using EvidenceHit.weight (primary/wire/trade/other), and keep simple totals like totalSupport and totalRefute. A good starting scale is: primary domains = 1.0, wire services ≈ 0.8, trade/specialist press ≈ 0.6, and other/long-tail sources ≤ 0.4.
• Use a documented threshold rule to decide when evidence is "enough": for example, treat the evidence as sufficient when either support or refute weight is ≥ 1.6 from at least two independent, time-aligned sources (mirroring the primary/wire weighting sketch in the spec).
• Choose one of three outcomes to start with: True, False, or Invalid. Use True when support clearly wins above threshold, False when refute clearly wins, and Invalid when we either never reach the threshold or the evidence is too conflicting/noisy by the time SERP attempts are exhausted.
• Expose a small decision result that either (a) contains a final AgentResult (outcome + proof + sources) or (b) says "need more SERP search" so the loop can ask Serper for another query if current_serp_attempts < max_serp_attempts.
• When producing a final AgentResult, call a small LLM-backed proof writer that takes SearchTask + the best EvidenceHits and returns a compact markdown proof, then map that into the shared AgentResult shape from the Input issue (Input: Web‑search agent input layer #4).

The AgentResult contract is shared with the Input layer:

interface AgentResult {
  outcome: string; // e.g. True | False | Invalid (initially)
  proof: string;   // short markdown summary + key snippets
  sources: Array<{
    url: string;
    title: string;
    pub_date: string | null;
    excerpt: string;
  }>;
  debug?: {
    total_queries?: number;
    total_pages_visited?: number;
  };
}

---

TODO

• Implement an initial rule-based decision helper that uses SearchHistory aggregates (support/refute totals, number of independent sources, timeframe alignment) to decide when evidence is enough for True / False and when we should fall back to Invalid or request more SERP search. Start with a simple weighting scheme (primary = 1.0, wire ≈ 0.8, trade ≈ 0.6, other ≤ 0.4) and a threshold like support >= 1.6 or refute >= 1.6 from ≥2 independent sources.
• Define a decision boundary, e.g. decideFromHistory(task, history, attempts, maxAttempts), that wraps this helper and returns either { status: 'final', result: AgentResult } or { status: 'need_more_search' } based only on SearchTask + SearchHistory + simple counters.
• Keep a single decideFromHistory entrypoint so Validator and external callers never depend on internal SearchHistory details; future changes to weights, thresholds, or additional outcomes should only require updating the internal rules, not the boundary itself.

[linear]

SWARM-71