Human+LLM Co-Cognition Methodology: Applicability Guide
Version: v1.0 | Date: 2026-06-08
What This Is
An operational pattern handbook. It doesn't tell you what to think — it describes what works and what doesn't when one human and multiple AIs explore problems where "we don't know what we'll find."
Three Stories
P3 Crisis Cognition — "I know the goal, but not the path"
The task: extract cross-domain, actionable "cognitive resilience design principles" from military command, disaster response, emergency medicine, and HRO — fields that never talk to each other.
He ran an internal critique (Phase D) attacking his own hypothesis, then sent it to two external agents for independent audit with no preset positions. Both audits independently caught a systematic bias he hadn't seen: he was too inclined to find differences rather than shared structures. He had labeled HRO and disaster science as "independent clusters" — both agents independently concluded this was wrong.
Prior: "8 domains share the same crisis cognition degradation structure"
↓ Phase D critique: HRO and disaster science judged "independent clusters"
Posterior v1: "They're different — not a unified meta-domain"
↓ Hunyuan + Kimi independent audit: "You systematically favor finding differences"
Posterior v2: "They share a degradation-recursive sensemaking loop — Phase D's judgment was bias, not discovery"PN Wealth Inheritance — "I'm the domain expert, but expert intuition can be the blind spot"
The task: design an "AI-era family wealth self-check framework" for Chinese households. He knew this domain well.
He started with a physics metaphor — "IEHL half-life" — to measure asset purchasing power decay. Three rounds of external agent review, seven version iterations, two case studies — full methodology deployment.
What worked: 70% of substantive corrections across v0.1→v0.8.3 were triggered by external agent feedback. The sharpest fixes (MCA from "asset" → "multiplier", IEHL from exact years → ⚡ symbols) came from external attack.
What didn't: It took seven full rounds to realize the "atomic decay" metaphor might be fundamentally wrong for Chinese family wealth. Chinese wealth concepts lean closer to "river" (intergenerational flow) or "container" (institutional attachment) — not atomic decay. The methodology didn't catch this anchoring because Step 2 had no mandatory "underlying metaphor fitness test."
Prior: "IEHL half-life is the best metaphor"
↓ External review: MCA should be "multiplier" not "asset"
Posterior v1: "Right direction, needs de-pseudo-precision — ⚡ instead of years"
↓ Round 7 fission crisis: both LLMs favored complex compromise; human framework guardian: "Drop fission, stay honest"
Posterior v2: "The metaphor itself may not fit — atomic decay vs. river/container"P2 Co-Cognition Map — "I don't know what I'll find"
The task: systematically map every position in human knowledge where two disciplines could generate co-cognition breakthroughs. No preset framework, no right answer.
Three parallel search strategies: taxonomic scan (source A), gene radiation (source B), and de-anchored free scan (source C) — with a mandatory quota forcing at least 2 candidate domains from disciplines uncovered by A and B. The framework absorbed 27 suggestions from external agent group audit. The anti-anchoring quota worked: Kimi's source C scan discovered "crisis cognition collaboration" as a meta-domain — which later became its own P3 subproject.
Awaiting evidence: P2 has not closed. MERA has not been executed.Decision Table
| If your project… | Most like | Start with | Biggest pitfall |
|---|---|---|---|
| Fuzzy goal, needs external perspective | P2 | Step 0, tri-source cross, anti-anchor quota | Framework symmetry may be designed, not discovered |
| Clear goal, unclear path, cross-domain | P3 | A/B adversarial briefing, independent audit, weighted self-check | Internal critique may have systematic bias |
| You ARE the domain expert | PN | Tri-source cross, human mixed-mode tags, MERA | Anchored by your first idea (F6) |
| Clear goal, data-rich, execution-focused | — | Not recommended | Time cost > value |
When Not to Use
Quantitative/data-driven projects — if standard scientific workflow (data → model → validate) works, tri-source cross is over-engineering. Solo, single-domain, tight-deadline execution — if you need a competitive analysis in 3 days, you need efficient information gathering, not Step 0 problem framing.
Self-Reference
This guide you're reading was written using this methodology — Step 0 confirmed the scope, Step 2 produced the three-story + decision table structure, external agents audited the draft. If the methodology works on itself, that's the minimum honest test.
Minimal Starter Pack
- Tri-source cross: Ask the same question to two different LLMs — look for what they both missed
- A/B adversarial briefing: Reformat your task as sharp challenges and send it again
- Fatal-level self-check: Q2 (did I over-trust LLM in unfamiliar territory?) + Q6 (did LLM repackage my judgment as its own output?) — 2 minutes
Stop rule: Two consecutive rounds with no new insight = you're done.
Further Reading
→ Rule Pool Governance: How the Lab's 11 document rule types evolve from static documents to a traceable rule evolution system (v2.2, based on JERP)