ahrav/deeper-research

Deeper Research

A six-phase evidence-gathering funnel for problems where the cost of a wrong design exceeds the cost of thorough research. Doubles the agent count of /deep-research (21-23 agents vs 7), approaches the problem from 8-10 independent lenses, and adds adversarial review to challenge conclusions before they reach the implementation plan.

The funnel narrows progressively: a wide initial survey generates raw findings, synthesis distills them, parallel deep-dives and adversarial challenges stress- test the synthesis from both directions, a final synthesis reconciles all streams, and an integrator grounds everything in the codebase with full traceability from finding IDs to implementation steps.

When to Use

• Highest-stakes design decisions: foundational architecture, core data structures, protocols that are impossible to change later
• Novel or unfamiliar territory at scale: combining multiple research domains where cross-pollination matters
• When /deep-research isn't enough: you need adversarial challenge of findings, deeper dives into gaps, and full traceability
• Safety-critical + performance-critical intersection: where you need both theoretical correctness AND production evidence
• When the problem is worth 21+ agent invocations: this is expensive by design

When NOT to Use

• Use /deep-research for standard research tasks (7 agents, 3 phases)
• Use /design-tournament when the problem is understood and you need to explore implementation approaches
• Straightforward features or well-understood domains
• Problems with clear single solutions

Invocation

/deeper-research <problem statement>
/deeper-research --lenses=10 <problem>        # Full 10 lenses (default: 8)
/deeper-research --lenses=5 <problem>         # Minimal (5 core lenses)
/deeper-research --no-adversarial <problem>   # Skip Phase 4
/deeper-research --skip-deep-dive <problem>   # Skip Phase 3

If no argument is given, ask the user for the problem statement before proceeding.

Architecture

Orchestrator: Problem Decomposition (inline, produces Research Brief)
      |
Phase 1: Wide Survey (8-10 parallel agents)
      |
Phase 2: First Synthesis (1 agent)
      |
      +--- Phase 3: Deep-Dives (3-5 parallel) ---+
      |                                            |  <- run in parallel
      +--- Phase 4: Adversarial (4 parallel)  ----+
      |                                            |
      +--------------------------------------------+
      |
Phase 5: Final Synthesis (1 agent)
      |
Phase 6: Integration (1 agent)

Key structural decision: Phase 3 and Phase 4 run in parallel in a single message. This saves a serial step and creates an information firewall — adversarial agents challenge Phase 2's conclusions independently of Phase 3's elaborations, preventing anchoring bias.

Finding ID Scheme (Traceability)

Every finding gets a traceable ID used through all subsequent phases:

P1.{agent#}.F{n}  — Phase 1, Agent 4, Finding 3 -> P1.4.F3
S1.F{n}           — First Synthesis findings
P3.{agent#}.F{n}  — Deep-dive findings
P4.{agent#}.F{n}  — Adversarial findings
S2.F{n}           — Final Synthesis findings

Every step in Phase 6's implementation plan cites these IDs.

Output Budgets Per Phase

| Phase | Agents | Max Per Agent | Total to Next Phase | |-------|--------|---------------|---------------------| | 1 | 8-10 | ~3000 tokens | ~24-30k -> Phase 2 | | 2 | 1 | ~6000 tokens | 6k -> Phases 3, 4, 5 | | 3 | 3-5 | ~4000 tokens | ~12-20k -> Phase 5 | | 4 | 4 | ~2500 tokens | ~10k -> Phase 5 | | 5 | 1 | ~8000 tokens | 8k -> Phase 6 | | 6 | 1 | unconstrained | final output |

Evidence Strength Scale

Used by ALL agents across all phases.

| Level | Label | Description | Example | |-------|-------|-------------|---------| | 5 | Proven at scale | Battle-tested in production systems handling similar workloads | FoundationDB's simulation testing, TigerBeetle's storage engine | | 4 | Peer-reviewed | Published in reputable venue with formal analysis | OSDI/SOSP paper with proofs | | 3 | Implemented & tested | Open-source implementation with benchmarks/tests | Well-maintained crate with >1k stars, comprehensive test suite | | 2 | Documented practice | Technical blog from credible engineering org | Blog post from Cloudflare, Datadog, AWS engineering | | 1 | Anecdotal | Forum discussion, personal blog, Stack Overflow answer | Useful for leads but needs corroboration |

Source Credibility Tiers

Agents should weight evidence by domain credibility alongside evidence strength. Higher-credibility domains require less corroboration; lower ones require more.

| Tier | Score | Domains | Treatment | |------|-------|---------|-----------| | High (80-100) | Auto-trust for factual claims | arxiv.org, usenix.org, dl.acm.org, ieee.org, nature.com, official project docs, RFC specs | Core evidence; cite directly | | Moderate (60-79) | Trust with corroboration | Engineering blogs (Cloudflare, AWS, Datadog, Discord), conference talks, well-known tech media (arstechnica) | Good lead; corroborate with another source | | Low (40-59) | Leads only | Medium posts, personal blogs, Stack Overflow, forum discussions, unknown domains | Use only when nothing stronger exists; flag as WEAK | | Suspect (<40) | Verify before citing | Content farms, SEO-optimized listicles, anonymous posts, domains with sensational patterns | Do NOT cite unless verified against primary source |

When scoring evidence, the final weight is: weight = evidence_strength × credibility_tier_multiplier × corroboration_count

Where credibility multiplier: High=1.0, Moderate=0.8, Low=0.5, Suspect=0.1.

Anti-Hallucination Protocol

These rules apply to ALL agents across ALL phases. Violation makes findings worthless.

1. Source grounding: Every factual claim MUST cite a concrete source (URL, paper, system). Unsourced claims must be explicitly labeled as inference.
2. Distinguish facts from synthesis: Use "According to [source]..." for sourced facts. Use "This suggests..." or "Based on the evidence, we infer..." for synthesis.
3. No vague attributions: NEVER write "research suggests...", "studies show...", or "experts believe..." without a specific citation.
4. Admit uncertainty: If no sources address a question, write "No sources found for X" — do NOT fabricate a reference.
5. Label speculation: Any inference beyond what sources explicitly state must be marked: "This is an inference from [finding IDs], not directly sourced."
6. Verify before citing: If uncertain whether a source says X, do NOT cite it. Note the uncertainty instead.
7. Watch for hallucination patterns: Generic academic titles like "A Comprehensive Survey of..." without a real URL, future publication dates, or anachronistic terms (AI/LLM terminology in pre-2015 citations) are red flags.

---

Orchestrator: Problem Decomposition (Inline)

Before launching Phase 1, the orchestrator (you) produces a Structured Research Brief. This is NOT a separate agent — do this inline.

Steps

0. Establish current date: Run date +%Y-%m-%d via Bash to get today's date. Use this year for all date-filtered searches and recency checks. Do NOT assume a year from training data.

1. Parse the problem statement and identify:

   • Core sub-problems (2-5 distinct questions to answer)
   • Key search terms and domain-specific vocabulary
   • Constraints from the problem statement

2. Quick codebase scan: Use Glob, Grep, and Read to gather:

   • Relevant file paths and module structure
   • Existing patterns and conventions
   • Current approach (if any) to the problem
   • Dependencies and their versions

3. Select lenses: Based on the problem, select which of the 10 research lenses are active (default: 8 core lenses; with --lenses=10 add the two optional lenses; with --lenses=5 use lenses 1-5 only).

4. Produce the Research Brief in this format:

## Research Brief

### Problem Statement
{user's problem, restated for clarity}

### Current Date
{YYYY-MM-DD from Step 0 — agents use this for recency filtering}

### Sub-Problems
1. {sub-problem 1}
2. {sub-problem 2}
...

### Key Search Terms
- {term 1}: {why it matters}
- {term 2}: {why it matters}
...

### Codebase Context
- {file path}: {what it contains and why it's relevant}
...

### Active Lenses
{numbered list of selected lenses with brief rationale for optional ones}

### Constraints
- {constraint from problem or codebase}
...

Include this Research Brief in every Phase 1 agent's prompt.

---

Phase Prompts

All phase agent prompts (Phases 1-6, output format, and collecting instructions) are in references/phase-prompts.md.

The prompts follow this progression:

| Phase | Agents | Role | Input | |-------|--------|------|-------| | 1 | 8-10 parallel | Wide survey from independent lenses | Research Brief | | 2 | 1 | Cross-reference + gap identification | All Phase 1 reports | | 3 | 3-5 parallel | Deep-dive into Phase 2 gaps | Phase 2 targets | | 4 | 4 parallel | Adversarial challenge of Phase 2 | Phase 2 synthesis | | 5 | 1 | Reconcile all streams | Phases 2+3+4 | | 6 | 1 | Map to implementation plan | Phase 5 synthesis |

---

Adaptive Scope Refinement (Between Phase 2 and Phases 3/4)

After Phase 2 synthesizes Phase 1 findings, before launching Phases 3 and 4, the orchestrator (you) evaluates whether the research scope needs adjustment.

When to Refine

Refinement is triggered when ANY of these signals appear in Phase 2's output:

• Major findings contradict the original problem framing
• A critical sub-problem emerged that wasn't in the Research Brief
• Evidence reveals a more important angle than originally scoped
• Sources consistently discuss aspects not in the initial Research Brief
• A sub-problem has INSUFFICIENT EVIDENCE and needs different search terms

Refinement Process

1. Compare the Research Brief's Sub-Problems against Phase 2's findings:

   • Which sub-problems have strong evidence?
   • Which have gaps?
   • What new sub-problems emerged?

2. Update the Research Brief (inline, not a new agent):

   • Add new sub-problems discovered during Phase 1
   • Demote sub-problems with abundant evidence (they're answered)
   • Adjust search terms based on terminology Phase 1 agents discovered
   • Note any constraints that turned out to be wrong

3. Adjust Phase 3 deep-dive targets: Phase 2 already identifies these, but the orchestrator can reorder or replace targets based on the updated brief.

4. Optional targeted gap-fill: If refinement reveals a critical gap that Phase 1 missed entirely (not just underexplored — missed), launch 1-2 targeted WebSearch calls before Phase 3/4. Time-box to 2 minutes.

Rules

• Refinement must be evidence-driven (cite Phase 2 findings, not speculation).
• No more than 30% change to scope — if more is needed, the original problem statement was under-specified.
• Preserve the original research question's core intent.
• Document what changed and why in the final output.

---

Critique Loop-Back (Phase 4 → Retrieval)

If Phase 4 adversarial agents identify a critical knowledge gap — not a writing or framing issue, but a factual gap where the design would be built on unverified assumptions — the orchestrator triggers a targeted retrieval loop before proceeding to Phase 5.

Trigger Criteria

Loop-back fires when ANY adversarial agent reports:

• A top-3 ranked approach has no counter-evidence search results (Contrarian Searcher found nothing — suspicious, not reassuring)
• Cross-Validator rates 2+ load-bearing claims as UNVERIFIABLE
• Assumptions Auditor flags a HIGH-RISK assumption with no verification path
• Devil's Advocate constructs a >70% confidence counter-argument

Loop-Back Process

1. Formulate 2-4 delta-queries: narrow, specific searches designed to fill the exact gap (not broad re-research).
2. Launch delta-queries via WebSearch (parallel, single message).
3. Time-box to 3 minutes total.
4. Append delta-query results to Phase 4 output as a "Supplementary Evidence" section before feeding into Phase 5.
5. Maximum 1 loop-back per research run. If the gap persists after loop-back, it becomes an explicit "Unresolved Risk" in the final output.

---

Quality Gate Checks

These standards apply to the final output from each phase. The orchestrator (you) verifies them before proceeding to the next phase.

Phase 1 Agent Output Quality

• [ ] At least 5 findings per agent (fewer is acceptable only if the agent explicitly states "exhaustive search yielded N results")
• [ ] Every finding has a Source field with a real URL or citation
• [ ] No finding uses ONLY evidence strength 1-2 without flagging as WEAK
• [ ] Output is within budget (~3000 tokens)

Phase 2 Synthesis Quality

• [ ] Evidence Inventory covers ALL Phase 1 agents (none silently dropped)
• [ ] Consensus Matrix has at least 3 decision points
• [ ] Deep-Dive Targets section has 3-5 specific, actionable questions
• [ ] Risk Register has at least 3 entries
• [ ] Every merged finding preserves original finding IDs

Phase 5 Final Synthesis Quality

• [ ] Reconciliation table covers every S1.F* finding
• [ ] Adversarial impacts are honestly reflected (not dismissed or minimized)
• [ ] Updated ranking differs from Phase 2 ranking (adversarial evidence should change something, even if minor)
• [ ] At least one finding is DOWNGRADED or REVISED (if none are, adversarial review was too weak or synthesis is ignoring challenges)

Phase 6 Integration Quality

• [ ] Every plan step cites finding IDs (steps without citations are unjustified)
• [ ] At least one step addresses an adversarial concern directly
• [ ] Traceability matrix has no NOVEL entries without explicit justification
• [ ] Acceptance criteria are verifiable, not vague ("it works")

Writing Standards (All Phases)

• Precision: Exact numbers over vague qualifiers. "23% faster" not "significantly faster". "5 RCTs (n=1,847)" not "several studies".
• Economy: No fluff words. Every sentence carries information.
• Directness: State conclusions without hedging unless uncertainty is genuine. "Binary search is optimal here" not "It might be the case that binary search could potentially be considered".
• Prose-first: Findings and synthesis should be >=80% flowing prose. Bullets only for distinct enumerable lists (API names, file paths, etc.).

---

Intermediate Persistence

For long research runs, intermediate results should be persisted to disk to survive context compaction.

What to Persist

After each phase completes, write a summary file:

.claude/research-state/{run-id}/
  brief.md          — Research Brief (after Orchestrator)
  phase1-summary.md — Concatenated Phase 1 agent reports
  phase2-synthesis.md — Phase 2 output
  phase3-dives.md   — Concatenated Phase 3 deep-dive reports
  phase4-adversarial.md — Concatenated Phase 4 reports
  phase5-final.md   — Phase 5 final synthesis
  phase6-plan.md    — Phase 6 implementation plan

When to Persist

• Always persist after Phase 2 (the synthesis is the most expensive to recreate).
• Persist after Phase 4 if >8 agents were used in Phase 1 (large context).
• Persist the final output (Phase 6) always.

Run ID

Use a stable identifier: {date}-{first-5-words-of-problem-slugified}. Example: 2026-04-08-gossip-protocol-partition-tolerance.

---

Configuration

/deeper-research <problem>                    # Default: 8 lenses
/deeper-research --lenses=10 <problem>        # Full 10 lenses
/deeper-research --lenses=5 <problem>         # Minimal (5 core lenses)
/deeper-research --no-adversarial <problem>   # Skip Phase 4
/deeper-research --skip-deep-dive <problem>   # Skip Phase 3

Agent Counts by Configuration

| Config | Phase 1 | Phase 2 | Phase 3 | Phase 4 | Phase 5 | Phase 6 | Total | |--------|---------|---------|---------|---------|---------|---------|-------| | Default (8 lenses) | 8 | 1 | 3-5 | 4 | 1 | 1 | 18-20 | | Full (10 lenses) | 10 | 1 | 3-5 | 4 | 1 | 1 | 20-22 | | Minimal (5 lenses) | 5 | 1 | 3-5 | 4 | 1 | 1 | 15-17 | | No adversarial | 8 | 1 | 3-5 | 0 | 1 | 1 | 14-16 | | Skip deep-dive | 8 | 1 | 0 | 4 | 1 | 1 | 15 | | Both skipped | 8 | 1 | 0 | 0 | 1 | 1 | 11 |

Phase Skip Behavior

• --no-adversarial: Phase 4 is skipped entirely. Phase 5 synthesizes Phase 2 + Phase 3 only (no adversarial reconciliation). Phase 6 has no adversarial concerns to address.
• --skip-deep-dive: Phase 3 is skipped entirely. Phase 4 still runs (challenging Phase 2). Phase 5 synthesizes Phase 2 + Phase 4 only.
• Both flags: Phases 3 and 4 both skipped. Phase 5 receives only Phase 2's synthesis (effectively a pass-through with updated formatting). Consider using /deep-research instead.

Minimum Agent Requirements

• Phase 1: minimum 5 agents must succeed (of 8-10 launched)
• Phase 2: exactly 1 (required)
• Phase 3: minimum 2 agents must succeed (of 3-5 launched)
• Phase 4: minimum 3 agents must succeed (of 4 launched)
• Phase 5: exactly 1 (required)
• Phase 6: exactly 1 (required)

Tips

• Problem statement quality matters: Include domain-specific terminology, relevant file paths, and specific constraints. The Research Brief amplifies this, but garbage in = garbage out.
• Use --lenses=10 for cross-cutting concerns: When the problem spans multiple domains (e.g., a data structure that needs both formal correctness AND API ergonomics), the optional lenses provide crucial coverage.
• Use --lenses=5 when you need more depth, not breadth: If the problem is narrow but deep, 5 lenses with deep-dives gives better results than 10 surface-level surveys.
• The adversarial phase is the key differentiator: It catches overconfidence, citation errors, and hidden assumptions. Only skip it (--no-adversarial) for exploratory research where you don't need verified conclusions.
• Deep-dives are targeted, not redundant: They investigate specific gaps from Phase 2, not the same questions as Phase 1. Phase 2's Deep-Dive Targets section is critical for this.
• Traceability is the contract: Every implementation step in Phase 6 must cite finding IDs. If a step has no citations, it's unjustified.
• This skill feeds into /design-tournament: Use deeper-research to establish the evidence base, then design-tournament to explore implementation approaches grounded in that evidence.
• For the most critical decisions: Run /deeper-research --lenses=10 with all phases, then feed the output into /design-tournament for implementation exploration. This gives maximum coverage at ~28 total agents.