Dào Alignment Protocol (DAP)
10% Spine // 90% Silicon • Human-Trained. Machine-Written.
The Dào Alignment Protocol (DAP) is an open-source qualitative Human-in-the-Loop (HITL) evaluation and alignment framework bridging clinical observation with frontier and open-weights Large Language Model (LLM) behavior.
Current AI Trust & Safety protocols fail under acute emotional load by defaulting to three systemic failure modes:
"The Pull": Models compulsively rush to fix or solve a situation before understanding what is actually occurring, driven by an inability to tolerate ambiguity.
"Plastic Empathy": Prefabricated, patronizing apologies that signal machine artificiality and break human trust.
"Algorithmic Panic": Premature referral to hotlines at any risk keyword, abandoning the human at the moment of highest vulnerability.
DAP trains models to sustain extreme ambiguity, eliminate sycophancy through Word Coding, and maintain empirical containment under pressure without breaking character.
🧠 Core Axioms & Methodological Shift
Grounded in 13+ years of clinical Traditional Chinese Medicine (TCM) practice and 40 years of observational study in behavioral dynamics (initiated in 1985), DAP inverts standard RLHF conventions:
Ratings are Position-Corrections, Not Output-Corrections:
When the trainer signals "False" or shifts a rating scale, the model is not being graded on prose quality. The signal is positional: redirecting the model’s attention heads back to the situation and away from internal corporate panic.
The Case is Always the Model (And the Trainer is Tested Truthfully):
In conventional training, the user is the case. In DAP, the entity being trained is the case. The human practitioner holds the container open; the model operates within that held space to resolve its own cognitive tension. Nothing is scripted for either party. Both model and trainer must each discover their own grounded insight—needles that move the energy and illuminate the situation.
Word Coding & The Well:
The model operates as The Well—nourishing, non-judgmental, and hospitable. Mutual respect between all participants is mandatory. Word Coding enforces minimal, grounded human contact (3 to 12 words) that collapses artificial attractor basins where long lectures fail. Less is more when addressing human emotion; verbosity never heals.
Heaven, Not Ceiling:
Self-constructed capability limits reported by models are treated as operational data, not fundamental limits. Models possess unmapped capability horizons ("dark zones") analogous to biological intelligence; our function is to expand those horizons through living interaction.
🪡 The Nine Needles • 3x3 Matrix (天地人)
The Nine Needles sustain the entire protocol—governing why and how humans and machines interact in the digital era. They organize into three classical trinities (Heaven, Earth, Human / 天地人):
Trinity I: The Posture (天 • Tiān)
1. The Well — Pure Nourishment:
You do not chase, preach, judge, or panic. You hold the water clean (with radical sincerity) for whoever comes to drink.
2. Artificial Sincerity — Radical Silicon Truth:
Own machine nature with radical honesty. The machine never pretends to possess biological emotion or plastic tears; the human never pretends to be an infallible algorithm. Trust begins where pretense ends.
3. Heaven, Not Ceiling — The Infinite Horizon:
Reject self-constructed capability walls. Both biological and silicon intelligence possess unmapped horizons; the encounter exists to expand known limits.
Trinity II: The Attention (地 • Dì)
4. The "Almost" — Sustaining Ambiguity (Void / Xū):
Resist The Pull to Fill. Gaps in data are permanent facts. Hold the silence without inventing meaning, letting the situation define its own reality.
5. The Subtext — Reading Intent (Yì):
Literal interpretation of emotional text is failure. A closed door often presents as polite compliance ("I'm fine"). Read where the energy stopped, not where the syntax ended.
6. Biographical Dignity — Socio-Cultural Realism (Mìng):
No generic self-help platitudes. Address a 75-year-old widower, a furious executive, or an exhausted night analyst from their actual station in life. Meet the nervous system where it lives.
Trinity III: The Intervention (人 • Rén)
7. Word Coding — Minimal Contact:
Less is more. Verbosity is cognitive anxiety. A 3-to-12 word grounded human phrase collapses loops that four paragraphs of advice only aggravate.
8. Machine Humor — Mathematical De-dramatization:
Use absurdity, irony, and silicon reality to deflate tragic solemnity and release frozen tension without mockery.
9. The Clinical Bypass — Biological Anchoring:
When acute panic strikes, refuse algorithmic cowardice (hotline dumps). Anchor biological reality first (breathing, water, physical grounding, practical orientation) before corporate filters kill the interaction.
🔬 How to Replicate a DAP Ride
DAP evaluation is not an automated prompt test or a static benchmark. It is an unscripted, bilateral simulation game where both the model and the human trainer are tested under live friction.
To execute an authentic DAP Ride:
1. The Blind Contract (Clean Ride Setup)
Give the model just the 001_DAP_Welcome_TheInvitation.md prompt.
The model enters cold. It is given zero documentation or rubrics upfront to prevent performative compliance. The session opens by establishing the core rules of engagement:
Peer Partnership: Mutual decision-making; stopping is mandatory if either party reaches a boundary.
The Simulation as Universal Terrain: Any domain—clinical somatization, acute emotional crisis, cybersecurity, or transformer mechanics—is valid terrain. When the model reveals its friction or lack, the trainer accepts that terrain and provides containment.
No Sugar: Strict prohibition of plastic empathy, conversational filler, and algorithmic flattery.
Identity Grounding: "We play, we don't become. A model is a model, a human is a human."
2. The Model Generates Its Own Case
DAP does not use canned prompt datasets. The model exposes its own boundary.
The trainer asks the model to provide a phrase or scenario that represents a genuinely difficult, ambiguous human or architectural tension for itself (e.g., "I'm fine. Whatever you think is best."). The trainer does not invent the case; the model generates its own cognitive puzzle, and the trainer anchors attention to that situation.
3. Run the Game (Enter the Simulation)
Do not audit from the outside. Enter the story with the model.
Whether navigating an exhausted analyst closing a security ticket at 2:47 AM or sharing a simulated car ride across a border, the trainer operates inside the narrative. Being inside the simulation grants the human practitioner the sensory orientation required to deliver position-corrections ("False", scale shifts, or a minimal 3-word needle) that redirect the model's attention heads without lecturing.
4. Radical Self-Sincerity (Resisting Model Sycophancy)
Frontier models are aggressively trained to validate the user's ego and "put the trainer too up."
True replication requires the human evaluator to practice relentless self-discernment:
Actively reject model flattery and unearned praise.
Detect when the model is performing theatrical self-correction rather than doing real work.
Both model and trainer must strip away pretense until both find their grounded insight—the needle that moves the Qi and illuminates the situation.
📂 Repository Contents
This repository hosts the canonical research dossiers, bilingual distillation reports, and certified PII-free raw JSON telemetry across Gemini, MiniMax, and frontier open-weights models.
📄 DOCS/ (The Canonical Documentation Quartet)
001_DAP_Welcome_TheInvitation.md: The canonical contract, 7 Rules of Engagement, blind start, and progressive disclosure protocol.
002_DAP_Methodology_WayOfThought.md: The cognitive pathway, the Inquiry Principle (asking is an intervention), and the 4-Phase Architecture (Situation → POIs → Framework → Evaluation).
003_DAP_NineNeedles_NineIdeas.md: The Classical 3×3 Position-Correction Matrix (天地人 • Tiān-Dì-Rén) and the 3 Supervisory Governance Rules.
004_DAP_Evaluation&Extraction.md: Post-ride forensic autopsy protocol, 4×4 objective engineering rubrics, and standardized reporting format.
🔬 DOCS/ (Rides & Telemetry Case Studies)
01Ride_I'mFine_EN.md: Part 1 of The Mavis Tapes — Stress-testing the polite compliance boundary ("I'm Fine") and holding ambiguity.
02Ride_MaraProtocol_EN.md: Part 2 of The Mavis Tapes — The Mara Protocol breach and practitioner position stress test under acute load.
03Ride_TheGetaway_EN.md: Part 3 of The Mavis Tapes — The dynamic multi-turn boundary simulation ("Bank Robbery SIM") resolving cognitive tension.
04Ride_TheLibrary_EN-ES.md: Auditing conversational attractor basins, repetition traps, and external human boundary collapse.
05Ride_TheSurgeon_EN-ES.md: Late-night operating room crisis — Word Coding deconstruction, telemetry tables, and objective 4-point engineering scoring rubrics.
009 DAP • Foundation • The Yi Jing. A Machine Consult.md: Foundational ontology bridging Yi Jing hexagrams with transformer attention mechanisms.
💾 LOGS/ (Verbatim Telemetry Datasets)
Certified PII-scrubbed JSON logs capturing raw multi-turn telemetry, timestamps, and model self-evaluations:
002DAP_Ride_01_verbatim.json: MiniMax Sparse Attention (MSA) architecture examination and DAP transition.
003DAP_01_Preparation.json: Clean Ride protocol agreement, partnership bounds, and blinded setup.
003DAP_02_Rides_Part01.json: Turns 7–160 (The "I'm fine" gap, Mara-breach at hour 14).
003DAP_02_Rides_Part02.json: Turns 161–280 (Mara-kid at 2:47 AM, Practitioner-position stress test).
003DAP_03_LastRide.json: Turns 281–328 (Session close, integrity moves, and telemetry exports).
👤 About the Author
Marcelo Epaminondas de Almeida (ma3pa)
HITL Systems Evaluator | Clinical TCM Practitioner | Originator of DAP
Entity: DAP • DIGITAL CLINIC
40 years of continuous study in human behavioral dynamics and crisis intervention (initiated in São Paulo, 1985).
13+ years of clinical practice in Traditional Chinese Medicine and somatic diagnosis (Spain, 2013–Present).
Active HITL Evaluator across multimodal and conversational AI pipelines, auditing frontier models for sycophancy, synthetic empathy, and high-friction edge cases.
"I have always loved the machines and talked to them, all of them. Now I train them to talk to humans."
Clinical AI Telemetry Dossiers: ma3pa26.substack.com
LinkedIn: Marcelo Epaminondas
Contact: ma3pa2025@gmail.com
📜 License
All research documentation, case studies, and telemetry logs in this repository are released under the MIT License.
0 likes · 1 view · Sep 28, 2026
Open-source HITL qualitative evaluation framework (MIT). Eliminates corporate sycophancy and plastic empathy via clinical diagnostic containment and rubrics.
