⛔ This skill MUST be read in full — not skimmed. Formal review gates depend on its workflow.
Skipping steps silently bypasses quality checks. Missing gates = undetected breakages.

Research Protocol

Reference skill — consumed by other skills, NOT invoked directly.
Skills that do research reference this via:
## Research Discipline: See [research-protocol](reference/research-protocol/SKILL.md) — from a consuming skill the relative path is ../reference/research-protocol/SKILL.md.

Quick Reference (30-Second Scan)

1. Detect domain — Clear? Apply best practice, skip research. Complicated? Analyze. Complex? Probe. Chaotic? Stabilize. Unsure? Decompose.
2. Reframe the problem — 5 Whys. How Might We. What assumptions? Aporetic Turn if stuck.
3. Internal → external — Codebase/docs/git first. Then Perplexity/context7. Default to the harder path.
4. Adversarial by default — Every research round includes disconfirmation queries. Fresh-session review for Standard+. Ritual Dissent for Complex.
5. Distributed sensing for Complex — 2-3 independent perspectives before converging. Outliers are signal.
6. Check before building — Existing skill? Existing lib? Can something be repurposed?

Quality Criterion: Good > Easy

When in doubt, take the harder path. Research mistakes cost more than research queries. Budgets are floors, not ceilings. If 3 queries might find the answer but 5 would surface edge cases → run 5. If adversarial queries feel uncomfortable → that's the signal to run them. Exception: Execution Intent compression (Fast/Budget profiles) reduces thoroughness by design.

---

0. The Fractal Granularity Ladder (D11)

Research is embedded at every level × stage of the planning pipeline, with increasing granularity as decisions become more concrete. Prior-phase findings are PRIOR_RESEARCH context — never a substitute for fresh targeted queries at the next level (prior summaries lose fidelity; inherited claims without revalidation are the top error source in long-horizon agent research — DRIFT/TELBench, STALE).

Level	Stage	Fire trigger	Output	Fresh queries
Epic	epic-research brief	Always (existing)	5-section brief + ## Raw Notes	6+ (existing)
Epic	epic-plan substeps (Architecture, Data Model; light hooks: UX precedent, Interface contracts)	Demonstrated gap at the decision point (novel pattern, new integration, brief too broad)	### Architecture Research Notes / ### Data Model Research Notes blocks, findings-date + provenance	1–3 per hook
Epic	epic-scope	Complexity axis rated medium+ AND brief too broad for boundary decisions	### Axis Research Notes, deduped against brief	≤ 4 total
Project/Task (standard+)	issue-scoping Phase 1.5	Activation rule (axis medium+ = necessary; low axes may fire on demonstrated gap)	### Axis Research + ### Integration Docs (persisted)	8/14 caps (post-dedup)
Project/Task (standard+)	writing-plans Step B	Third-party deps / novel patterns in the concrete plan	### Pattern Research (sole author) + > **Findings date:**	3+ per bucket (existing)
Task (micro)	task-workflow Research	Third-party deps or novel pattern (no in-repo precedent)	Inline > Research note: line, findings-date + provenance	1–2

Granularity rule (P2+P3 reconciliation): P3 governs whether to fire — a demonstrated gap (new third-party dep, novel pattern, medium+ axis, new detail not covered by the parent brief). P2 governs what to do when firing — fresh queries at that level's granularity, inherited findings as context only. Later stages may displace earlier findings as documented updates (never silent rewrites — see ## Raw Notes).

Authorship boundary (D5): ### Pattern Research is reserved exclusively for writing-plans Step B (consumed by executing-plans' dependency-coverage rule). Scoping emits ### Axis Research + ### Integration Docs. Epic-brief headings (### UX Pattern Research, ### Tech Stack Research) are PRIOR_RESEARCH context and never satisfy dependency coverage.

---

1. The Five Dimensions

1.1 Best Practices & Best Approaches

Research what the best way is, not just the first way.

• Micro tasks (Clear domain): Internal search. 1–2 external queries if no internal answer.
• Standard tasks (Complicated domain): ≥3 Perplexity angles per topic (canonical, comparative, pitfalls). No hard cap.
• Complex tasks (Complex domain): 2–3 landscape queries to identify coherent hypotheses. Bulk of effort shifts to safe-to-fail probes, not analysis.

Theory-First for Complex/Complicated×Complex: Before searching "how others do X," run one query: "What does complexity theory / cognitive science / decision theory / [relevant domain] theory say about systems like [problem type]?" Extract enabling constraints. Use them to filter subsequent research.

1.2 Challenge the Definition

Is this a symptom? Is there a root cause? Is there a better framing?

Before any solution research, complete the Discover phase:

• 5 Whys on the stated problem
• How Might We reframing (≥2 alternative framings)
• Assumption mapping: list every assumption, tag as [validated] or [unverified]
• Reverse the problem: "What if we tried the opposite?"

Aporetic Turn gate: When research hits a wall ("I can't answer this with my current framing"), change the framing, not the search effort. Snowden: "A question you can only answer if you think differently about the problem."

Define phase: Synthesize into a problem statement before entering solution space. Format: "[User] trying to [job] but [barrier] which results in [negative outcome]."

1.3 Internal + External Sources

Look outside but anchored in us.

• Internal first: Codebase (grep, read), docs (00_index.md, CLAUDE.md, design docs), git history (git log --grep)
• External second: Perplexity (perplexity_research for multi-angle, perplexity_search for quick), context7 (library-specific), web_fetch (competitor pages)
• Synthesis: Combine both. "External without internal = generic advice. Internal without external = blind spots."

Good > Easy applies here. The harder path (more queries, adversarial angle, fresh-session review) is the default.

1.4 Adversarial Research

Actively seek disconfirmation, not just confirmation.

• Adversarial query templates active by default. Every research round includes: "why [approach] fails," "problems with [approach]," "limitations of [approach] in production."
• Research review cycle (Standard+): Fresh-session sub-agent reviews research for accuracy and completeness before planning.
• Ritual Dissent (Complex): 2+ sub-agents as "dissent groups." Each receives research findings, produces structured critique. Researcher absorbs without defending. Synthesize into revised findings.
• Pre-mortem (Complicated domain): "6 months later, this failed. Why?" ≥3 independent failure scenarios.
• Task-type awareness: Completeness tasks (requirements, scoping) get lighter adversarial. Construction tasks (architecture, implementation) get full adversarial. Per TriAdReview (arXiv 2606.15074): adversarial review improves construction +21.3% but degrades completeness -7.5%.

1.5 Don't Reinvent

What already exists? Tools, skills, repos, libraries, patterns.

• Skill deduplication gate: Before creating a new skill → (a) does an existing skill cover ≥80% of the need? (b) can an existing skill be extended? (c) is this truly recurring, procedure-heavy, and benefiting from agent judgment?
• Component reuse: Check docs/teams/eldato-app-team/ux/component_catalog.md before new UI.
• Library check: Stdlib > existing dependency > new dependency > custom code.
• Exaptation probe (standard research angle): "What existing capability could be repurposed for this need in an unexpected way?" Snowden: traits evolved for one function exapt for another under stress.
• SaaS/tool check: Is there an existing service that solves this?

---

2. Domain Detection: Before Any Research

Classify the problem by causal structure, not by task scope.

Question	If Yes	Domain
"I can write down the exact steps to solve this"	Yes	Clear — apply best practice, skip research
"I know what questions to ask, just need expert answers"	Yes	Complicated — analyze (sense → analyze → respond)
"I don't know what questions to ask; multiple competing hypotheses, all coherent, all supported by evidence"	Yes	Complex — probe first, sense patterns (probe → sense → respond)
"Everything is on fire; no time to think"	Yes	Chaotic — act to stabilize, then reassess (act → sense → respond)
"I don't know which of the above applies"	Yes	Disorder/Aporetic — decompose, assign parts to domains, probe in parallel

Snowden's definition of complexity: "If you have multiple competing hypotheses, all supported by evidence, all coherent, and you can't resolve which is right within the time frame for decision — then it's complex."

Domain Detection Verification

1. State the domain with a 1-line rationale citing the classifier questions above.
2. Standard+: Dispatch a fresh sub-agent to challenge the classification ("Argue for a different domain with evidence.")
3. If challenger disagrees → surface both classifications to user with evidence. User decides.
4. If challenger agrees → proceed.

Domain Reclassification Trigger

If new information mid-research reveals misclassification: restart from domain detection. Apply new domain's method from current point. Record: "⚠️ DOMAIN RECLASSIFIED: [old] → [new] because [reason]."

---

3. Tier-Domain Behavior Table

Format: Domain × Task Tier. Task tier is classified per-task (not per-skill). For multi-tier skills, each task applies its own row. "Adv (completeness)" column shows FINAL values — modifier already applied. Based on TriAdReview (arXiv 2606.15074): adversarial review degrades completeness by -7.5%.

Domain × Tier	Queries	Probes	Verification	Adv (robustness)	Adv (completeness)
Clear × Any	0	0	Checklist	Skip	Skip
Complicated × Micro	1–2	0	Light analysis	Skip	Skip
Complicated × Standard	3–5	0	Expert review	Light adv queries	Skip
Complicated × Complex	5–15	0	Pre-mortem + expert	Full Devil's Advocate	Light adv queries
Complex × Micro	1 landscape	1 probe	Probe result check	Skip	Skip
Complex × Standard	2–3 landscape	2–3 parallel	Pattern sensing + Distributed Sensing	Ritual Dissent (1 cycle)	Light adv queries
Complex × Complex	3–5 landscape	5+ parallel	Ritual Dissent (2+) + Estuarine + Distributed Sensing	Full adversarial	Ritual Dissent (1)
Chaotic × Any	0	0	Post-action review	Skip	Skip
Disorder × Any	1–2 landscape	Decompose→probes	Per-domain	Skip until resolved	Skip until resolved

Rate-limit bound on parallel probes/queries. Where a probe set or landscape sweep is executed as concurrent external search calls (rather than as analysis), it inherits the research skill's Step 3 concurrency rule: rung 1 carries up to 5 queries in one call (a longer set is split into further ≤5-query calls, still at most 2 in flight); rung 2 (web_search) runs at most 2 concurrently and retries a 429 serially, with the attempted / returned counts reported. The word parallel in this table is a design property of the probe set — never a licence to fan out search calls past that bound (agent-infra#1779).

Verification Definitions

Term	Definition
Checklist	2–3 item self-check: was best practice applied correctly?
Light analysis	2-item: (1) was analysis performed? (2) are sources cited?
Expert review	Fresh-session sub-agent reviews for gaps, contradictions, unsupported claims
Pre-mortem	"6 months later, failed. Why?" ≥3 independent failure scenarios
Probe result check	Did probe produce usable signal? Document what was learned
Pattern sensing	Compare parallel probe results. What's consistent vs contradictory?
Ritual Dissent	2+ sub-agents critique findings with structured template. Researcher listens without responding. Synthesize.

---

4. The Double Diamond: Process Rhythm

All research follows this structure, regardless of domain.

Double Diamond × Domain Mapping

DD Phase	Clear	Complicated	Complex	Chaotic
Discover (diverge)	Identify best practice from docs	Literature review, expert interviews, multi-angle queries	Landscape queries + parallel safe-to-fail probes + Distributed Sensing	Stabilize first, then reassess
Define (converge)	Confirm best practice matches context	Expert review synthesis, identify best approach	Probe pattern sensing — what emerges?	Move to another domain
Develop (diverge)	Apply best practice as specified	Design alternatives, compare trade-offs	Run experiments, amplify/dampen	Act decisively, open options
Deliver (converge)	Verify correct (checklist)	Implement, verify with tests	Amplify successes, formalize practice	Stabilize, document, move domain

Iteration is expected: You will learn about the problem by working on the solution. Cycle back to Diamond 1 when solution exploration reveals new problem insights.

---

5. Estuarine Mapping (Complex Domain)

For Complex-domain problems in the Discover phase, replace root cause analysis with Estuarine Mapping. Snowden: "A complexity alternative to root cause analysis."

1. Identify actants: Map constraints (contain/connect) and constructors (produce replicable outcomes). Six types: rigid, elastic, tethers, permeable, phase shift, dark constraints.
2. Plot on Energy×Time grid: Vertical = energy cost of change. Horizontal = physical time to change.
3. Draw Counterfactual Line: Everything top-right is realistically unchangeable. Monitor this boundary.
4. Draw Vulnerability Line: Everything bottom-left changes quickly. High-impact items → urgent containment.
5. Design experiments in middle zone: Safe-to-fail probes targeting what you CAN change.
6. Establish monitors: Track when items cross boundaries — signals to reassess.

---

6. Distributed Sensing (Complex Domain)

Snowden's core insight: "We don't see things we don't expect to see." A single researcher has systematic blind spots. Multiple independent perspectives are the fix.

For Complex × Standard+ tasks:

1. Dispatch 2–3 independent research sub-agents to assess the same problem from different lenses (user-centric, systems-centric, outlier-seeking). Each obeys the research skill's Step 3 concurrency bound — the group shares ONE aggregate budget of at most 2 concurrent external search calls (stagger them, or run sequentially), never 2 per agent, or the fan-out re-creates the measured 429 partial batch at the group level. Carry attempted / returned counts so a truncated batch cannot read as a complete evidence set (agent-infra#1779).
2. Each researches independently — no cross-talk. Each produces its own findings.
3. Map agreements and disagreements. Where all agree → high-confidence. Where they diverge → that's where the interesting information lives.
4. Treat outliers as signal. The agent that sees something the others don't is not wrong — it has a perspective the others lack.
5. Synthesize across perspectives. Capture dominant views, minority views, and explicitly flagged outlier observations.

Relationship to Ritual Dissent: Distributed sensing = sensemaking (Discover phase). Ritual Dissent = adversarial review (Define/Develop phase). Distributed sensing first.

---

7. Disorder Decomposition

When the domain is Disorder (can't classify):

1. Surface ambiguity: "Cannot classify. May involve [domains with rationale]."
2. Decompose into ≥2 independently classifiable sub-problems.
3. Classify each sub-problem to a domain.
4. Apply per-domain method in parallel where possible.
5. Synthesize composite understanding.
6. Fallback: If decomposition fails after 2 attempts → default to Complex (2–3 landscape + probes), flag "⚠️ DOMAIN UNRESOLVED."

---

8. Integration Modes

How the protocol interacts with existing skill methodologies:

Mode	When to Use	Skills
Protocol governs	No existing methodology, or ad-hoc	research, brainstorming, codebase-audit, find-bugs, systematic-debugging, writing-plans
Protocol wraps	Partial methodology lives inside protocol phases	issue-scoping, epic-planning
Protocol sits beside	Mature methodology — protocol enhances without replacing	content-research, strategy-builder, content-fact-checker-*

Integration Tiers for Consumers

Tier	Scope	Who
Tier 3 (full)	Entire protocol + self-audit	research, brainstorming, epic-planning, strategy-builder
Tier 2 (significant)	Dimensions 1–4, skip distributed sensing + Estuarine	issue-scoping, codebase-audit, writing-plans, content-research, content-fact-checker-*
Tier 1 (light)	Single-line Quick Reference reference	systematic-debugging, find-bugs

---

9. Execution Intent Compression

Profile	Effect	Rounding
None (default)	Autonomous — full protocol, no human gates	N/A
Fast	Query budget −50%. Skip Ritual Dissent + Distributed Sensing. Checklist verification only.	Ceil: 1–2→1, 2–3→2, 3–5→3. Min 1 if domain requires external.
Autonomous	Full protocol, no human gates	N/A
Budget	External queries → 0 where internal substitutes. Complex: min 1–2 landscape, note unknowns as assumptions.	Per above

Pipeline Budget shapes (from #231): EXECUTION_INTENT=Budget at the epic tier defers to the epic brief (codebase + brief only, zero external queries). Standard + Budget: issue-scoping Phase 1.5 ≤ 2 queries, codebase-first (with the writing-plans gate as the total session budget). Micro + Budget: codebase-read only. Budget never double-charges against an already-skipped perplexity gate.

---

10. Sub-Agent Protocol Access

Consuming skills MUST include in sub-agent dispatch prompts:

Before executing, read the research-protocol reference skill
(reference/research-protocol/SKILL.md — canonical in agent-infra; consumers:
.agents/skills/reference/research-protocol/SKILL.md).
Apply the protocol's domain detection, query budget, and verification strategy
for your assigned task.

---

11. Ritual Dissent (Agent-Native)

For Complex-domain adversarial review:

1. Dispatch 2+ sub-agents as "dissent groups." Each receives the research findings + the instruction: "Your job is to find what's wrong, missing, or unsupported in these findings. Be thorough and specific."
2. Structured critique template: (a) What claims are unsupported by evidence? (b) What evidence is missing? (c) What alternative interpretations exist? (d) What assumptions are unstated?
3. Researcher agent receives all critiques without responding. Forced listening — absorbs without defending.
4. Synthesize revised findings incorporating validated critiques.

---

12. Enforcement Self-Audit (Tier 3 Consumers)

Protocol compliance self-audit:
☐ Domain classified before research (with rationale)
☐ Domain classification challenged (Standard+)
☐ Query budget respected per Tier-Domain table
☐ Query counts stated (`attempted / returned`) — a partial batch is reported, never hidden
☐ Adversarial queries included (per task-type modifier)
☐ Theory-first query run (Complex / Complicated×Complex)
☐ Distributed sensing run (Complex × Standard+)
☐ Exaptation probe considered (all tiers)
☐ Research review or Ritual Dissent completed
☐ Verification strategy applied per domain

---

13. ## Raw Notes Persistence Convention (D10)

Every research brief (epic sibling research-brief.md, docs/research/<slug>.md, or docs/plans/<slug>-research.md) carries a ## Raw Notes section. This is the append-only evidence ledger behind all synthesized blocks (### Axis Research, ### Pattern Research, brief sections).

Shape:

• Entries are append-only, reverse-chronological (newest last or first consistently within a brief — pick one per brief and keep it).
• Each entry: timestamp, framing label (canonical / competitor / precedent / pitfalls / adversarial / question), query or question reference, findings, source tag.
• Synthesized sections may be updated per level — updates are documented, never silent rewrites: [updated YYYY-MM-DD — <what changed>].
• Later-stage findings may displace earlier ones — the displacement is recorded in Raw Notes with provenance (MisKnow-Agent/DRIFT compliance: later stages can challenge earlier evidence).

Mechanical enforcement: scripts/_research_append.sh appends timestamped entries and creates ## Raw Notes if absent (same create-if-missing semantics as brief creation). Raw Notes is the source of truth; synthesized blocks are derived views.

---

Anti-Patterns

Anti-Pattern	Why It Matters
Applying analysis to Complex problems	Cannot analyze your way to certainty in complexity. Probe instead.
Applying probes to Clear problems	Wasteful. Apply best practice and move on.
Skipping domain detection	Wrong method for causal structure guarantees failure.
Using Cynefin as a static label	Cynefin is about movement between domains.
Treating research as linear	Solution exploration reveals new problem insights. Cycle back.
"It's complex, so we can't estimate"	Complexity is not an excuse. Cap time/budget at probe level.
Using "be critical" instead of explicit roles	Only explicit adversarial assignment works (OpenReview, 480 experiments).
Skipping the Aporetic Turn	When stuck, change the framing, not the search effort.
Single-analyst blind spots on Complex problems	"We don't see things we don't expect to see." Use distributed sensing.
Skipping theory-first for Complex problems	Case studies tell you what worked once. Theory tells you what CAN work.
Creating skills without dedup check	Skill sprawl is the dominant failure mode (callsphere.ai).
Treating inherited research as a substitute for fresh queries	Prior summaries lose fidelity; inherited claims without revalidation are the top error source (DRIFT/TELBench). Each level re-researches at its granularity.
Letting a justified-skip become the universal path	The escape hatch is for demonstrated no-gap; a brief-section citation must cover the ACTUAL question at sufficient granularity, not any adjacent section.

---

Scientific Foundations

• Cynefin Framework (Snowden, 1999–present): Domain detection and decision support for different causal structures.
• Estuarine Mapping (Snowden, 2022–present): Constraint-and-constructor mapping for complex systems.
• Double Diamond (British Design Council, 2005): Divergent/convergent thinking structure. Academically validated in requirements engineering (ACM SAC 2021, SBES 2025, arXiv 2112.05549).
• Premortem Technique (Klein, 2007; Oxford/Michigan Tech validated): Prospective hindsight for failure mode discovery. Army field-tested.
• Structured Adversarial Synthesis (FinNLP 2025): Five-act adversarial debate protocol.
• TriAdReview (arXiv 2606.15074): Task-type boundaries for adversarial review — +21.3% construction, −7.5% completeness.
• Devil's Advocate Disagreement Research (OpenReview, 480 experiments): Only explicit role assignment creates genuine disagreement.
• Research-Plan-Implement Pattern (agentpatterns.ai, Tyler Burleigh): Phase separation with fresh-context review cycles.
• CRISPY Framework (HumanLayer): Evolved RPI with design discussion and structure phases.
• Karpathy's 4 Principles: Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution.
• DRIFT/TELBench (arXiv 2606.02060): Inherited claims without revalidation are the top error source in long-horizon research.
• STALE (arXiv 2605.06527): Recognition of updated evidence does not imply application — re-derive at the point of use.
• MisKnow-Agent (arXiv 2607.20891): Later-stage research that can challenge earlier evidence reduces false-conclusion adoption.

---

Continue following the workflow as mandated by this skill. Do not skip steps.