Literature Survey

State Ledger

When a tactic returns a Yield Report, ADD its counts to the Current column. </HARD-GATE>

Procedure — Iterative Loop

WHILE (gaps.length > 0 AND iteration < MAX_ITERATIONS):

  STEP 1 — QUERY GENERATION:
  Generate 3 academic queries + 3 web queries based on:
  - Current gaps (from previous reflection)
  - Research brief initial queries (first iteration only)
  - Knowledge accumulated so far

  STEP 2 — SEARCH + READ:
  Call academic-research tactic with 3 academic queries
  Call web-research tactic with 3 web queries (in parallel)

  STEP 3 — REFLECT:
  Use prompts/reflect-gaps.md to discover new gaps from what was just read
  Update knowledge with key findings

  STEP 4 — EVALUATE:
  Use prompts/evaluate-answer.md to judge sufficiency
  Remove gaps that are now answered

Budget Gate

1. Papers read >= 80% of target? (e.g., Medium → >= 32 papers)
2. Web pages fetched >= 80% of target? (e.g., Medium → >= 40 pages)
3. Queries issued >= 60% of target?

If ANY condition fails → CONTINUE iterating. Do NOT evaluate stop conditions. The Budget Gate takes absolute priority over convergence signals.

Only when ALL three conditions pass may you proceed to the stop check. </HARD-GATE>

  STEP 5 — STOP CHECK:
  - All gaps cleared? → STOP (success)
  - No progress for 3 iterations? → STOP (plateau)
  - Target paper count reached? → STOP (coverage)
  - MAX_ITERATIONS hit? → STOP (budget)

END LOOP

Completeness Probe — Lit-Survey

Sub-area coverage:

• Have I searched at least 3 distinct sub-areas of this topic?
• For each sub-area, do I have at least 3 papers?
• Are there obvious sub-areas I haven't explored?

Citation tracing:

• Did I check the references of the top-cited papers I read?
• Did I find and read at least 2 seminal/foundational papers?
• Are there highly-cited papers (>100 citations) I missed?

Research community diversity:

• Have I found papers from at least 2 different research communities? (e.g., ML community + domain-specific community)
• Are there conferences/journals I haven't searched?
• Did I search for both recent (last 2 years) AND foundational (5+ years) work?

Depth check:

• Did I actually READ (not just skim titles of) my top papers?
• Can I explain the key contribution of each top paper in 2 sentences?
• Do I have enough detail to identify genuine research gaps?

If ANY checkbox fails → add 1-3 targeted queries and run ONE MORE iteration. Maximum 2 extra completeness iterations allowed. </HARD-GATE>

State Management

• knowledge: cumulative string, grows each iteration
• papersRead: string[] of paper titles, deduplicate across iterations
• urlsVisited: string[] of URLs, deduplicate across iterations
• gaps: string[] discovered via reflection, removed when answered

Output

{
  "knowledge": "Accumulated knowledge string...",
  "papersRead": ["paper1", "paper2", ...],
  "urlsVisited": ["url1", "url2", ...],
  "gaps": ["remaining gap1", ...],
  "facets": [...],
  "iterationsUsed": 7,
  "stopReason": "all_gaps_cleared | plateau | target_reached | max_iterations"
}