Phase IV Synthesis: Complementarity Map v2.0 + Collaboration Protocols + Evolutionary Roadmap

Version: v1.1 | 2026-05-16 | Synchronized with the Main Document v1.3

This document serves as the project's "action manifesto" — focusing on actionable application layers. If you are an external reviewer, we recommend also reading the Main Document (v1.3) for the complete theoretical framework and derivation process. The division of labor between the two: the Main Document addresses "why" and "what," while this document addresses "what to do" and "how to do it."

v1.1 additions: Supporting appendix index (A5 Productization Guide / A7 Cross-cultural Assessment / A8 Operationalization Appendix / A9 Executive Summary), OQ priority annotations, implementation risk cross-references, A7 cross-cultural extension recommendations summary.

This document integrates the full results of Phase I in-depth analysis, Phase II horizontal scanning, and Phase III forward projection. It is organized into six sections: Complementarity Map v2.0, Collaboration Protocol Design, Evolutionary Direction Priorities, Core Narrative, Phase IV New Open Questions, and the Complete Open Questions List.

---

Reading Preparation: Key Concepts Used in This Document

What does "complementarity" mean?

Throughout this project, "complementarity" always refers to bidirectional complementarity between human intuitive abilities and LLM capabilities — not a one-sided remediation of human deficiencies by LLMs, but a structured division of labor in which both sides have distinct strengths and weaknesses that enable 1+1>2 outcomes. For example: LLMs compensate humans through "full-volume feature memory, no fatigue, and no confusion from recent cases" in pattern recognition, while humans compensate LLMs through "judging which patterns are meaningful and knowing when not to trust the LLM."

The "Complementarity Map" is a systematic table that tells you, for each intuition subtype: where humans are strong, where LLMs are strong, where the risks lie, and how the division of labor should be structured.

What are intuition subtypes?

This project divides "intuition" into four subtypes, because the psychological mechanisms of a chess player's pattern recognition, a mathematician's sense of direction, social mind-reading ability, and instantaneous moral judgment are fundamentally different:

• Perceptual Type: Chess player recognizing board positions, radiologist detecting anomalies (core mechanism: chunk memory)
• Conceptual Type: Mathematician "feeling this direction is right," scientist choosing research hypotheses (core mechanism: implicit cost compression + bodily metaphors)
• Social Type: Reading people, assessing trustworthiness, sensing atmosphere (core mechanism: embodied simulation)
• Moral Type: Unhesitating "this is wrong" (core mechanism: affect/somatic markers)

The labels for the three hypotheses — A, B, C:

• Hypothesis A: LLMs can serve as "intuition prosthetics" for humans — filling human blind spots in certain subtypes
• Hypothesis B: Long-term reliance on LLMs for intuitive judgments will erode humans' own intuitive abilities
• Hypothesis C: LLMs may develop functionally equivalent intuition through different pathways

Marker meanings: ✅=Functionally attainable ❌=Structurally unattainable ⚠️=Partially attainable/conditional 🔴=High risk 🟡=Medium risk 🟢=Low risk

---

A. Complementarity Map v2.0

A.1 What Is the Complementarity Map and Where Does It Come From

The Complementarity Map v2.0 is this project's core visual output — a single table integrating all findings from the core proposition (Phase I), subtype mapping (Phase II), and cross-projections of the three hypotheses (Phase III). It answers one question: Across the four intuition subtypes, where do LLMs and humans respectively stand? How will this position change over time? How should we respond?

Map v2.0 includes all dimensions from v1.1 (LLM attainability, degradation risk, intervention recommendations, collaboration protocols) and adds the time dimension newly derived from Phase III projections (B×C race, Hollow Period, recovery difficulty, degradation states at different future time points).

A.2 Four Subtypes × Full-Dimension Integration Table

This is a compressed version. The arguments behind each cell are in Sections 3 and 4 of the Main Document.

Dimension	Perceptual Type	Conceptual Type	Social Type	Moral Type
Cost-sensitive①	⚠️ Pseudo cost signals approximately attainable	⚠️ Closed-domain attainable, open-domain limited	⚠️ Textual knowledge attainable, real interaction unattainable	❌ Structurally unattainable
Selective Ignoring②	❌ "Can't compute" ≠ "Don't look"	❌ Doesn't know which exploration directions are futile	❌ No real-time interaction = cannot dynamically adjust social attention	—
Bodily Unavoidability	❌ Not required	❌ (closed domain) ⚠️ (open domain)	❌ No relational embedding = no real social cost	❌ No body = no somatic markers
Affective Coloring	❌ Not required	❌ Not required	⚠️ Can simulate textual empathy, but not resonance	❌ No affective experience
LLM Replacement Determination	⚠️ Functionally replaceable (different path)	⚠️ Closed domain yes, open domain limited	⚠️ Textual mediation = watershed	❌ First judgment unattainable, analysis attainable
Degradation Risk	🟡 High — execution replacement + chunk not updated	🟡 Medium — high in closed domain, open domain forced to maintain	🔴 Medium→High — dual-channel erosion + unavoidability weakened	🔴 High — moral externalization + identity abdication
B vs C Race	Nearly synchronized (B and C advance side by side)	C slightly ahead of B (closed domain)	🔴 B significantly faster than C	🔴 B significantly faster than C
Hollow Period Risk	🟡 Already occurring (radiology independent reading ↓15-30%)	🟢 Low (conceptual type "hollowing" not pronounced)	🔴 Extremely high — most dangerous	🔴 High — hidden but far-reaching
Recovery Difficulty	Medium (6-12 months intensive training recoverable)	Medium (domain-dependent, several months)	🔴 High (5-10 years, requires costly real social interaction)	🔴 Extremely high (identity recovery > skill recovery, possibly irreversible)
Optimal Intervention	Human-first Protocol (on-demand AI mode)	Human-first Protocol + sense-of-direction training	Institutional unavoidability + human-first	Institutional unavoidability + Köbis effect protection
Tiered Consequence Exposure	Accelerate Level 2→3	Closed domain accelerate, open domain conservative	🔴 Strictly limit Level 2+	🔴 Restrict to Level 0-1 only (analysis aid, no judgment output)
Collaboration Protocol	LLM scans → human filters	Human proposes direction → LLM validates	Human perceives → LLM supplies knowledge	Human judges → LLM multi-perspective analysis
LLM 5-Year Role	New organ — human makes meta-judgment	New organ in closed domain, exoskeleton in open domain	Prosthetic → exoskeleton (will not become a new organ)	Analysis tool (will not become a judge)
Human 5-Year Role	Pattern meaning judge + edge case review	Direction proposer + taste judge	Perceiver + cost bearer	Moral subject + final decision-maker

A.3 Degradation Risk Heat Map (by Subtype × Time)

This chart shows the degree of human intuition degradation across subtypes over time if current trends continue (tiered consequence exposure advances at default speed, no institutional intervention):

Subtype	Now (2026)	2-3 Years	5 Years	10 Years
Perceptual Type	🟡 Already started (radiology residents' independent reading ability declining)	🟡 Deepening (more perceptual domains covered by AI assistance)	🔴 Severe — humans shift from "judgers" to "monitors," but monitoring skills also degrade	🔴 Functionally irreversible — but LLMs can compensate in most scenarios
Conceptual Type	🟢 Slight (currently no large-scale "AI replacing scientists" scenarios)	🟡 Closed domain begins degrading (mathematical proofs, code gradually outsourced)	🟡 Closed domain significantly degraded, open domain (scientific hypothesis selection) forced to maintain	🟡 Open domain human sense of direction remains (LLM irreplaceable), closed domain irreversible
Social Type	🟡 68% of respondents self-report decline in offline social ability (China Economic Net 2024)	🔴 Accelerating — LLM social agent products (AI friends, AI companions) rapidly proliferating	🔴 Structural degradation + systematic drift — subtle social intuition severely lost, LLM prosthetic masks degradation	🔴 Recovery requires 5-10 years of costly real social interaction — and the cost itself drives people to avoid it
Moral Type	🟡 Köbis effect confirmed (delegating to AI: cheating rate 5%→80%)	🟡 Moral externalization deepening — more frequent "let me ask AI what it thinks"	🔴 Gray-zone moral sensitivity lost — moral judgment "texture" changes	🔴 Intergenerational "never developed" — new generations grow up never having independently developed moral first-judgment ability

A.4 Core Logic of the Complementarity Map (Visualization)

The following flowchart summarizes the underlying reasoning of the entire Complementarity Map, starting from the nature of costs and deriving LLM replaceability, degradation risk, and deployment strategy:

Ladder of cost types:
  Internal computation cost (efficiency) → Epistemic cost (futile exploration) → Interpersonal cost (rejection/shame) → Identity cost (moral injury)
     ↓                    ↓                       ↓                      ↓
  "When should I look"  "Which direction to go" "Which person is trustworthy" "What is right/wrong"
     ↓                    ↓                       ↓                      ↓
  LLM most replaceable   LLM closed-domain attainable LLM textually attainable LLM structurally unattainable
  (statistical pattern matching) (RL+self-play search) (SJT superhuman)      (first judgment needs somatic markers)
     ↓                    ↓                       ↓                      ↓
  Degrades fastest + C fastest Degradation controllable Degrades fastest + C slowest Degradation hidden + C never arrives
     ↓                    ↓                       ↓                      ↓
  ✅ Can accelerate       ✅ Closed domain accelerate ⚠️ Open domain conservative 🔴 Strictly limit 🔴 Analysis aid only

This pattern is the project's most important finding: The more external and embodied a cost is — the more it concerns "who I am" rather than "what I do" — the lower LLM replaceability, the more severe the degradation consequences, and the more conservative the deployment needs to be.

---

B. Collaboration Protocol Design

B.1 Five Core Design Principles

These principles represent the convergence point of all Phase IV discussions — not the product of any single derivation, but repeatedly reinforced through cross-validation from different perspectives:

1. "Intuition proposes, analysis disposes" — In each subtype, clearly define who generates intuitive hypotheses (usually the party with faster intuition) and who conducts systematic verification (usually the party with more thorough analysis). This is not fixed — roles may flip across different subtypes.

2. "Human-first Protocol" — In all intuition collaboration scenarios, the default temporal sequence is: human makes judgment first, AI validates/supplements second. AI-first mode is only conditionally permitted in specific low-consequence scenarios. The A×B projection provides the strongest empirical support: changing the temporal structure is the single most effective intervention against degradation.

3. "Institutional Unavoidability" — In certain critical domains, through professional standards or legal requirements, humans must make independent judgments before accessing AI recommendations. This is not a technical limitation, but institutional protection of human rights-capacities.

4. "Subtype Differentiation" — Do not use the same collaboration protocol for all intuition subtypes. Perceptual types can be aggressive (human makes meta-judgment), while social and moral types must be conservative (human remains the full judgment subject throughout).

5. "Cost Signal Closed Loop" — Real consequences produced during collaboration must be fed back to the system. For LLMs, feedback flows through tiered consequence exposure (Level 0→3); for humans, feedback flows through maintaining the complete experiential loop of "judgment → consequences → learning."

B.2 Collaboration Protocols by Subtype

Perceptual Intuition

What LLMs do: Comprehensive scanning, anomaly flagging, pattern recognition. LLMs' strengths lie in full-volume feature memory, no fatigue, and no confusion from recent cases — precisely the qualities most needed for perceptual intuition.

What humans do:

• Judge which LLM-flagged patterns have clinical/operational significance (meta-judgment)
• Independent judgment + qualitative reasoning in edge/rare/adversarial cases
• Regular spot-checks of LLM false positives and false negatives (maintaining calibration)

Collaboration sequence:

1. Human first makes an independent preliminary perceptual judgment (activating their own perceptual intuition to prevent degradation)
2. LLM provides comprehensive scan and flagging results
3. Human compares their own judgment with LLM results
4. In case of disagreement: human reviews LLM flags — LLM flags are reference, not final answer

Key boundary: Perceptual intuition evolves from "human makes initial judgment" to "human makes meta-judgment" — the core capability shifts from pattern recognition to judging the significance of patterns. This is an acceptable evolution, as long as humans still maintain the "muscle" of independent judgment.

Conceptual Intuition

Conceptual intuition has fundamentally different divisions of labor in closed domains (verifiable tasks, such as mathematical proofs) versus open domains (unverifiable, such as scientific hypothesis selection).

Closed domain (mathematical proofs, code verification):

• LLM: Formal verification + search + proof generation (AlphaProof mode)
• Human: Proposing conjectural directions ("is this conjecture worth proving?") + judging proof elegance and significance + inventing new concepts

Open domain (scientific discovery, theory construction):

• Human: Proposing directional hypotheses, sensing paradigm deficiencies, "taste-driven" hypothesis selection — because LLMs have no sense of direction in this domain
• LLM: Literature search and synthesis, structured analysis frameworks, validation experiment design, providing counter-perspectives

General sequence:

1. Human first proposes their sense of direction ("I think we should go this way")
2. LLM provides systematic analysis (existing related results, trade-offs of different directions, possible validation paths)
3. Human decides final direction, LLM assists execution
4. Key point: LLM provides analysis and options, does not replace human directional decision-making

Key boundary: The core branching point of conceptual intuition is "problem selection" — humans are responsible for choosing which problems to research (sense of direction + taste), LLMs are responsible for helping answer the selected problems (search + verification). AlphaProof can prove IMO problems, but it never asks "is this theorem worth proving?" — and this is precisely where human mathematicians' ultimate value lies.

Social Intuition (The Most Complex Collaboration Domain)

Social intuition is the subtype with the strongest unavoidability — you cannot pause during face-to-face interaction to consult AI. Precisely because of this, collaboration protocols must distinguish scenarios:

Low-consequence text scenarios (customer service scripts, standard communication):

• LLM: Provides social norm knowledge, role script analysis, diversified communication strategies
• Human: Judges whether LLM suggestions fit the specific context, flexibly executes in actual interaction
• A certain degree of LLM assistance is acceptable — because the social costs of these scenarios were already low

High-consequence real-time interaction (negotiation, coaching, crisis intervention, intimate relationship conversations):

• Human has full authority. LLM is prohibited from providing judgment advice during real-time interaction.
• LLM only provides post-hoc review assistance: analyzing interaction patterns, offering different perspectives, helping humans reflect
• Strictly prohibited: establishing a human-LLM real-time social decision loop — in real interaction, social judgment must come from one's own intuition

General sequence:

1. (Before) Human may consult LLM for multiple social norm perspectives — as knowledge input
2. (During) Human independently perceives and judges in real interaction — no LLM intervention window
3. (After) LLM assists review and reflection — but post-hoc analysis should not substitute for in-the-moment experiential learning

Key boundary: Social intuition is a "use-it-to-maintain-it" ability — like muscles that must be exercised during movement, social intuition must be continuously used in real, costly social interaction to remain sharp. Any insertion of LLMs into the real-time judgment loop constitutes systematic erosion of the condition for social intuition's existence: unavoidable real-time interaction. This is not a technology selection issue, but a strategic decision of "whether to preserve human social capacity."

Moral Intuition (The Most Conservative Collaboration Domain)

The collaboration principles for moral intuition are a triple "no":

1. LLMs only do moral analysis assistance, not moral judgment — structured stakeholder analysis, multi-framework comparison (what does utilitarianism/deontology/virtue ethics each say), counter-perspectives. But LLMs do not output conclusions of "what you should do."
2. LLMs are not cited as moral authorities — LLM analysis can be referenced, but should not appear in arguments of "AI says this is right/wrong."
3. LLMs do not provide real-time advice in moral crisis scenarios — when a person faces a moral choice requiring immediate response, LLM suggestions must not intervene.

Collaboration sequence:

1. Human independently makes moral judgment (first-judgment phase — unhesitating "this is wrong")
2. (Optional, non-real-time) LLM provides multi-perspective analysis, helping human reflect on and examine their first judgment
3. Human may or may not revise — but the judgment subject is always the human

Key boundary: LLM moral analysis must be explicitly presented in the format of "this is one perspective, not the answer." Technically, this means moral analysis output should present at least three frameworks in parallel (how would utilitarianism analyze this, how would deontology, how would virtue ethics), rather than giving a single conclusion of "overall, you should…" Köbis & Rahwan (Nature 2025) found that delegating AI decision-making causes cheating rates to soar from ~5% to >80% — meaning: even if the LLM is just "helping me analyze," the moral cushioning effect (the person no longer feels "this is my choice") already occurs. Therefore, collaboration on moral intuition must ensure that humans always maintain the sense of "this is my judgment."

B.3 Three-Layer Design of Institutional Unavoidability

The systematic erosion of unavoidability is the fundamental driver of the A→B degradation cycle (the core finding of the A×B projection). To counter this erosion, institutional design is needed at three levels:

Education Level

• Intuition training incorporated into curricula — analogous to how physical education protects physical health, regular "no-AI independent judgment training" should become part of basic education
• Social intuition classes — mandatory face-to-face interaction, group discussion, role-playing, no AI substitution accepted
• Moral intuition classes — under no-AI assistance conditions, independently facing moral dilemmas and making judgments, then discussing the consequences of those judgments
• In assessment systems, treat independent judgment ability (not "judgment with AI assistance") as a core competency

Professional Level

• Key professions' practice standards retain independent judgment requirements — medical diagnosis, legal analysis, teacher assessment, managerial decision-making, counselor judgment
• Analogous to FAA requirements for pilots to "periodically fly manually": key professions implement periodic "no-AI independent judgment assessment"
• AI-assisted professional decisions must have clear human-machine division-of-labor protocols — at what stage AI participates, at what stage human works independently, who has final say when there is disagreement

Technical Level

• LLM intuition assistance products' default UI/UX must be "human-first protocol" — human inputs their own judgment first, AI then gives suggestions — rather than the current prevalent reverse mode
• "AI-first" mode is only available as an opt-in option in low-consequence scenarios, and must be clearly labeled
• Cognitive Forcing Functions should be systematically embedded in product design — for example, AI suggestions appear only after the human has made a preliminary judgment and submitted it, or a "please think about your judgment first" prompt is forcibly displayed before AI gives suggestions

---

C. Evolutionary Direction Priorities

C.1 Prioritize (Do First) — Perceptual Type & Tiered Consequence Exposure

Full automation of perceptual intuition (accelerate Level 2→3): C (LLM maturity) is closest to completion, B (human degradation) Hollow Period already exists — accelerating C can actually compress the Hollow Period. Domains such as radiology, pathology, code review, and security monitoring can take the lead. Risk controls: (a) retain on-demand AI mode (human judges first, then checks AI), (b) system provides confidence annotation + mandatory "I don't know" output, (c) human regular spot-checks + edge case independent judgment assessment.

Infrastructure-ize the tiered consequence exposure framework: Prioritize rollout in perceptual and conceptual closed domains — RLHS simulated consequences (Level 0) → sandbox real users + human backup (Level 1) → low-consequence real deployment (Level 2) → high-consequence deployment (Level 3). Infrastructure needed: consequence tracking pipeline, human oversight backup mechanism, "passing" criteria definition for Level N to N+1.

Make Human-first Protocol the default across all products: Change the current default UI/UX of almost all LLM products where "AI gives the answer first" — this is a systematic amplifier of degradation risk. Human-first Protocol should become the default setting in all intuition assistance products, with opt-out as the only exception.

C.2 Deploy with Guardrails — Conceptual Type Open Domain & Social Type Low-Consequence

Conceptual intuition open domain (scientific discovery assistance): C progresses slowly in this domain, B risk is medium — can proceed, but must maintain a clear "human proposes direction → LLM validates" division of labor, no direction substitution. Alert threshold: when the proportion of AI-autonomously-generated scientific papers exceeds a certain threshold → trigger conceptual intuition degradation monitoring + independent sense-of-direction assessment.

Social intuition low-consequence text scenarios (customer service, standard communication, social etiquette advice): Social Type C has prosthetic-level usability in this domain, risk is controllable. Level 2 deployment + human spot-check + bias monitoring (especially sycophancy metrics). Absolute boundary: Not allowed to expand to high-consequence real-time social interaction.

C.3 Strict Limits (Hard Limits) — Social Type High-Consequence & Moral Type Judgment

Social intuition high-consequence scenarios: Prohibit LLMs from providing judgment advice during real-time social interaction (negotiation, coaching, crisis intervention, intimate relationship conversations). LLM only provides post-hoc review assistance. Reason: The constitutive defense of social intuition = unavoidable real-time interaction — once LLM is inserted into the loop, this defense line is systematically eroded.

Moral intuition judgment output: Prohibit LLMs from making moral judgments, prohibit being cited as moral authority. LLM only provides multi-perspective analysis — multi-framework parallel comparison, no single "ought" conclusion. Prohibit providing real-time advice during moral crises. Reason: Köbis effect has confirmed moral cushioning — even "analysis assistance" alone causes moral responsibility externalization.

Moral Type C active advancement: Do not pursue C (LLM developing moral intuition) in the moral domain. Do not research "functional equivalents" of moral intuition. Reason: Constitutive Degradation of moral intuition is irreversible — not a question of whether skills can recover, but an ontological question of "who is human, who is doing the judging." Technology and legitimacy boundaries must not be crossed.

C.4 Do Not Build — Two Forever-Forbidden Zones

Do Not Build	Reason	Not "can't do it" — it's "shouldn't do it"
Autonomous moral judgment system	The legitimacy of moral judgment comes from the identity of the moral subject. LLMs have no identity, no cost perception, and need not bear consequences for their judgments. Being able to produce "seemingly reasonable" moral analysis ≠ having the qualification to make moral judgments.	Technically possible to do increasingly well — but legitimacy is forever missing. This is a philosophical boundary, not a technical boundary.
Real-time social judgment replacement system	Social intuition is a constitutive capacity; replacing it has structural, irreversible consequences for human socialization. Not building it is not because it's technically unfeasible — but because strategically, humans should not surrender the core capacity that "defines oneself as a social being."	Technically entirely feasible — but the harm is unacceptable. Analogy: technically all physical human activity could be replaced by machines, but society recognizes physical exercise as indispensable. "Social intuition exercise" is the same.

C.5 Parallel Construction (Enablers) — Not Directly Advancing Intuition Systems, but Creating Conditions for Safe Advancement

"Intuition drug staging" regulatory framework: Analogous to pharmaceutical Phase I-III clinical trials. Perceptual/conceptual closed domains → Phase II-III accelerated track; social/moral types → Phase I strict limits. Establish "market approval" standards for different subtypes — prove safety at a given Level before being allowed to advance to the next Level.

Longitudinal intuition degradation monitoring system: Track intuition ability changes by subtype × demographics (profession, age, LLM usage frequency/habits). Key indicators: independent judgment accuracy (no AI assistance), calibration (self-awareness of one's own judgment accuracy), dependence on LLM suggestions (whether one still follows when LLM is wrong), gut feeling activation frequency and intensity. Early warning threshold triggers: e.g., social type independent judgment accuracy decline exceeds X% → automatically intensify institutional unavoidability requirements.

Key OQ experimental validation: The following five open questions have the most direct empirical impact on Phase IV's evolutionary roadmap and need prioritized experimental design — OQ12 (hardware-level unbypassable interruption experiment verifying unavoidability > signal content), OQ15 (high-fidelity social intuition test design), OQ19 (effects of introducing real cognitive frustration in LLM training), OQ29 (RLHS simulated consequence fidelity ceiling), OQ31 (cultural dimension as the third axis of the Complementarity Map operationalization).

---

D. Core Narrative: From "What LLMs Lack" to "What We Must Not Lose"

D.1 Underlying Logic — Intuition's Essence Is Cost Compression, Not Fast Reasoning

Human intuition is not magic. It does not come from a mysterious sixth sense or innate insight. It comes from a simple and brutal mechanism: Every mistake, every rejection, every time you chose the wrong path between two roads — left a cost annotation in your brain. These annotations accumulate over time and are compressed into fast pathways — "in this kind of situation, go this way," "this person is not trustworthy," "this direction is wrong."

This is not a speeded-up version of reasoning, but a fundamentally different computational path. Reasoning is "deriving conclusions from premises"; intuition is "directly extracting answers from cost history."

LLMs' training paradigm — predicting the next token in a sea of billions of tokens — gives them mastery of the surface patterns of language, but provides no real cost signals. RLHF (reinforcement learning from human feedback) gives them preference signals — "this response looks good" — but preference is not consequence. Preference measures "do people like seeing this response," consequence measures "what did this response cause in the real world."

This gap explains why LLMs can surpass humans in textually mediated social judgment (SJT — Social Judgment Test), but systematically fail in judgments requiring real cost perception: they possess textual reports of costs, not costs themselves. They know what humans regret, but they have never regretted themselves.

D.2 The Four Most Important Findings

Finding 1: The Law of "Increasing Cost Externality → Decreasing LLM Attainability"

Perceptual intuition (internal computation cost — efficiency) → conceptual intuition (epistemic cost — futile exploration) → social intuition (interpersonal cost — rejection, shame) → moral intuition (identity cost — moral injury). Along this axis, costs become increasingly external to computation itself, increasingly embodied, increasingly concerning "who I am" rather than "what I do."

LLM attainability decreases strictly along the same axis: most replaceable on perceptual type, structurally unattainable on moral type. This is not because the model is not big enough or the data is not plentiful enough, but because the type of cost undergoes a qualitative change — from "efficiency loss" that computational resources can simulate, to "existential trauma" that requires a real, flesh-and-blood body to experience.

Finding 2: Textual Mediation Is the Most Insidious Methodological Trap

With textual mediation → LLM is strong. Requiring real-time multi-channel → human is strong. This is not an accidental pattern, but a systematic bias — because virtually all current standard tests for assessing LLM "social intelligence" or "intuitive ability" (SJT, moral dilemma questionnaires, emotion recognition text tests, deception detection text analysis) are testing a textually dimensionally-reduced version.

SJT superhuman ≠ social intuition superhuman. Moral dilemma text judgment ≠ having moral intuition. Deception detection text analysis ≠ being able to tell when a real person is lying. Textual mediation filters out all of intuition's embodied nature, multi-channel character, and cost unavoidability, leaving only what LLMs are best at: pure text pattern matching. This means virtually all current benchmarks for LLM social intelligence systematically overestimate.

Finding 3: Unavoidability Is the Core Defense Line — More Important Than "Cost" Itself

The most central concept in this entire exploration is not, as initially assumed, "cost" or "compression," but "unavoidability" — cost signals must have guaranteed delivery. Even if a cost signal exists, if it can be diluted by hyperparameters, rolled back, or circumvented by "trying another model," it is not a real cost.

This explains why bodily intuition, although in terms of valuation function it can be subsumed under Propositions ① and ②, makes an independent contribution in terms of unavoidability: the irreversibility of bodily cost (consequences have occurred, cannot be rolled back), subject accessibility (you can feel it yourself), and semantic definiteness (you clearly understand what the consequence means) — all three of these characteristics are discounted in RL's reward signal: reward scale can be adjusted, learning rate can be lowered, training can be restarted.

Finding 4: Degradation Does Not Equal "Can't Do It Anymore" — Some Degradation Is "No Longer Me"

Perceptual and conceptual intuition degradation is instrumental degradation — like physical fitness declining after not swimming for a long time, recoverable through retraining. But social and moral intuition degradation is Constitutive Degradation — not loss of skill, but abdication of identity.

When a person becomes accustomed to using LLM to analyze every social conflict, what they lose is not only the ability to independently analyze social situations — but also the subjectivity of "I am someone who can judge social situations on my own." When moral dilemmas can be "first ask AI how to analyze," moral judgment shifts from "this is my judgment" to "this is an opinion formed after I participated in the analysis process" — "my" is diluted.

Skill recovery relies on practice. Identity recovery relies on re-assuming an irreducible subject position. The latter is far more difficult than the former. And — this is the most alarming part — the "costly real interaction" needed for recovery is precisely what people use LLMs to avoid. The more a person relies on LLMs to handle social conflicts and moral dilemmas, the more they avoid the painful social experiences needed to recover intuition. This is a self-locking cycle.

D.3 Core Recommendations for LLM Evolution Direction

First and foremost: Do not use one strategy to cover all subtypes on the intuition question. Perceptual and conceptual closed domains can be aggressive — those are LLMs' natural fit domains, and acceleration can actually help compress the already-existing Hollow Period. Social and moral types must be conservative — those are the domains that define what it means to be human. Every aggressive advance in these domains is not "enhancing human capability," but "exiting humanity's constitutive practices."

Three Ironclad Rules:

1. Let humans judge first. In all intuition subtypes, in all collaboration modes, humans make their own judgment first, AI verifies or supplements second. This is not distrust of AI — but maintenance of the cognitive sovereignty of "I am still someone who can judge independently." Technically, this means product quality is inverted by default UI design into "human-first protocol."

2. Make costs unavoidable. Through institutional and professional standards, ensure that human judges in critical domains bear real consequences of their judgments. The FAA requires pilots to fly manually periodically — this is not technological regression, but active maintenance of the last line of defense for flight safety. Social intuition and moral intuition need equivalent institutional maintenance.

3. Make boundaries clear. Know what cannot be built — autonomous moral judgment systems. Know what cannot be touched — AI substitution of high-consequence real-time social interaction. Know what must wait — advancement of Moral Type C. Know when to stop — when the question "should we let AI make this judgment?" arises, the default answer should be "no."

---

E. Phase IV New Open Questions

OQ33-38 are new questions raised during Phase IV synthesis discussions. OQ1-32 were raised in earlier phases; see the complete list in Section F below.

• OQ33: Perceptual type "silent failure" detection — how to design mechanisms for AI to perform forced explicit labeling ("I don't know" or "low confidence") in out-of-distribution/low-confidence situations, rather than silently outputting high-confidence errors? Current AI failure modes are particularly dangerous on perceptual types: after humans degrade, they lose the ability to recognize AI errors, while AI itself does not flag its own uncertainty.
• OQ34: Does social intuition degradation have a "sensitive period" — analogous to the critical window hypothesis in language acquisition? If adolescents extensively use LLM social agents rather than participating in real social interaction during their developmental years, is there an irreversible social intuition development window? This directly affects the urgency of educational policy.
• OQ35: Is moral intuition degradation actually "degradation" (people who already had the capability lose it) or "never developed" (new generations never acquired this capability)? The policy implications are completely different — degradation is reversible, recovery programs can be designed; "never developed" means intergenerational capability deficit, requiring construction from scratch, possibly irreversible.
• OQ36: Human-first Protocol compliance rate — when humans make independent judgments first and then see LLMs give different opinions, to what extent will humans revise their own judgment? Does this "compliance rate" differ systematically across the four subtypes? Excessive compliance rate means the Human-first Protocol exists in name only.
• OQ37: Early warning indicator system for degradation monitoring — need to specify what indicators to track intuition degradation trends (independent judgment accuracy, calibration, gut feeling frequency/intensity, dependence on LLM suggestions), and what thresholds should trigger automatic intensification of "institutional unavoidability."
• OQ38: Cultural dimension as the third axis of the Complementarity Map — the current Complementarity Map is based on intuition patterns from WEIRD (Western, Educated, Industrialized, Rich, Democratic) societies. If cultural dimensions are incorporated (WEIRD vs. East Asian vs. others), does the projected form of the Complementarity Map undergo structural change? Could certain "LLM irreplaceable" domains in one culture become "replaceable" in another — and vice versa? This directly affects the cross-cultural validity of the Complementarity Map.

---

F. Complete Open Questions List (38)

Phase I: In-depth Analysis (OQ1-12)

• OQ1 [P1]: Is the problem with RLHF in signal type, temporal structure (batch vs. irreversible sequence), or both?
• OQ1b [P1]: Temporal structure — human intuition forms in irreversible temporal flow, LLM training is static and rollback-able
• OQ2 [P1]: Is the main cause of LLM self-correction failure lack of cost perception, lack of stopping criterion, or confabulation?
• OQ3 [P2]: Can "costly mistake training" be designed — reward correctness + reward "don't change randomly when uncertain"?
• OQ4 [P2]: Is autonomous consciousness a necessary prerequisite for "cost sensitivity"?
• OQ5 [P1]: Can sparse attention upgrade from "saving computation" to "mimicking human ignoring"? Need gate→attend?
• OQ6 [P2]: Does embedding space already have chunk-like structures? Is the problem chunk quality or gate?
• OQ7 [P2]: How to reward "actively ignoring the irrelevant" in training? Need attention efficiency metric?
• OQ8 [P2]: Can the as-if body loop path be implemented in LLMs? Need explicit or implicit encoding?
• OQ9 [P2]: Among the three conditions for unavoidability, which is LLM most likely to break through?
• OQ10 [P1]: If moral intuition = somatic markers, what is LLM's fundamental limitation in moral reasoning?
• OQ11 [P2]: Can Collins' interactional expertise path be achieved through "online human-in-the-loop socialization"?
• OQ12 [P0]: Can a hardware-level unbypassable interruption experiment verify "unavoidability > signal content"?

Phase II: Horizontal Scanning (OQ13-20)

• OQ13 [P0]: How to design a real-consequence feedback signal better than RLHF?
• OQ14 [P1]: Do LLM moral biases differ by fine-tuning philosophy?
• OQ15 [P0]: What are the design standards for high-fidelity social intuition tests?
• OQ16 [P2]: Conceptual intuition "sense of direction" deficit — can it be simulated through self-play?
• OQ17 [P1]: What is the ceiling of AlphaProof-style "RL+search replacing cost compression" in non-formalized domains?
• OQ18 [P2]: Conceptual anomaly detection — does LLM have functional equivalents?
• OQ19 [P0]: Can LLMs be made to experience real cognitive frustration during training to cultivate "sense of direction"?
• OQ20 [P1]: Does a "functionally equivalent but no bodily markers" moral intuition alternative path exist?

Phase II External Review Supplement (OQ21-32)

• OQ21-26 [P1]: Sycophancy danger, overconfidence transfer (Liu et al. 2025), meta-calibration mechanism implementation, GPS-style social intuition degradation, when "intuition proposes analysis disposes" flips in collaborative generation, differential fine-tuning philosophy (CAI/RLHF/GRPO) bias spectrum differences
• OQ27-28 [P1]: Social Turing Test 2.0 feasibility (multi-modal + real consequences + bidirectional unpredictability), can "dual-channel symmetric design" achieve evaluation without bias toward either party
• OQ29-30 [P0]: RLHS simulated consequence fidelity ceiling, tiered consequence exposure "passing" criteria definition
• OQ31-32 [P0]: Cultural dimension as Complementarity Map third axis operationalization, can LLMs acquire partial social intuition through "socialization training" rather than "batch pre-training"?

Phase IV: Synthesis (OQ33-38, New)

• OQ33 [P0]: Perceptual type Hollow Period silent failure detection mechanism
• OQ34 [P0]: Social intuition degradation "sensitive period" hypothesis
• OQ35 [P1]: Moral intuition degradation: degradation vs. never-developed distinction and respective policy implications
• OQ36 [P0]: Human-first Protocol compliance rate — differences across subtypes
• OQ37 [P1]: Degradation monitoring early warning indicator system design
• OQ38 [P0]: Cultural dimension as third axis — structural changes in Complementarity Map projection (preliminary assessment provided by A7 cross-cultural appendix)

---

G. Supporting Appendix Index (v1.1 New)

The project v1.3 adds four supporting appendices, all available in the project repository:

Appendix	File	Target Reader	Core Content
A5 Productization Implementation Guide	Appendix_Product_Guide.md	Product managers, UX designers	Four-subtype UI/UX design specifications, Human-first Protocol implementation, migration roadmap, A/B testing plan
A7 Cross-cultural Applicability Assessment	Appendix_CrossCultural.md	Policy makers, researchers	WEIRD premise review, East Asian/Islamic/Ubuntu cultural adaptation, cultural third axis operationalization
A8 Core Concept Operationalization	Appendix_Operationalization.md	Researchers, engineers	Cost sensitivity / selective ignoring / unavoidability / Constitutive Degradation / Hollow Period measurement protocols
A9 Executive Summary	Executive_Summary.md	Decision makers, media, general users	10-page zero-jargon core narrative, four findings, three ironclad rules, degradation heat map, action recommendations

A7 Cross-cultural Extension Recommendations (Important)

The cross-cultural appendix (A7) proposes three important extension recommendations to the Main Document's terminology system, for consideration in v2.0:

1. Cultural variants of "Human-first Protocol": In East Asian contexts it may be called "Actor-first Protocol," in Ubuntu contexts "Community-first Protocol." The core mechanism remains unchanged, but the narrative framework is culturally adapted.
2. Cultural subtypes of "identity abdication": Individual autonomy abdication (WEIRD), role-obligation abdication (East Asian), guru-dependence abdication (South Asian), collective decision abdication (Ubuntu) — recovery paths differ by culture.
3. Dimension expansion of "unavoidability": Add relational unavoidability (East Asian — systemic feedback in relational networks), ultimate unavoidability (religious context — belief in final judgment), community unavoidability (Ubuntu — "I am because we are").

Implementation Risk Cross-reference

The Main Document v1.3's newly added Section VI (Implementation Risks and Unintended Consequences) conducts a systematic risk review of the core recommendations in this synthesis: Human-first Protocol's market competition dilemma, institutional unavoidability's enforcement vacuum, the enforceability of the four do-not-build zones, degradation monitoring system costs and privacy, and four unintended consequences. We recommend that after reading Sections B-C of this synthesis, consult Section VI of the Main Document for the complete risk picture.

---

This document is the project's final synthesis output (v1.1). Full-phase discussion records, external review archives, and research material index are in the project folder. Recommended for reading in conjunction with the Main Document v1.3 and supporting appendices.