---
title: "Run the method | The Terminal Experiment"
canonical_url: "https://www.strategicintelligenceresearch.org/studies/the-terminal-experiment/method/"
date_modified: "2026-09-16"
language: "en"
source_sha256: "193782cc7d4c25fefd34cfbe9591cc8961e88caeded3f2186596e1413ec642b9"
---

Canonical page: https://www.strategicintelligenceresearch.org/studies/the-terminal-experiment/method/

A practical method transfer guide

# Do not copy the prompts. Copy the operating system.

The project was not one brilliant prompt or seven models debating in a room. Dr. Waddell built a small research organization in a folder: one accountable expert, models with separated duties, sources with traceable provenance, and gates that stopped a draft from quietly becoming truth.

[Build my launch brief](https://www.strategicintelligenceresearch.org/studies/the-terminal-experiment/method/#builder) Copy the kickoff prompt

Inputs stay in this browser and are not sent by this page. They are saved only if you copy or download the generated file.

The transferable core

01

**Frame** Name the question, decision, scope, and evidence that could change your mind.

02

**Freeze** Set source, coding, and analysis rules before the answer is visible.

03

**Challenge** Separate creation, criticism, verification, and human acceptance.

04

**Translate** Carry one accepted evidence base into the report, site, article, and audio.

The whole system

## One workflow. Two systems. Four human gates.

Models do bounded work. Versioned artifacts carry that work forward. The subject-matter expert controls every consequential gate.

Human authority

**The subject-matter expert stays above the whole process.** ◆ Approve the plan   ◆ Authorize evidence access   ◆ Decide every finding   ◆ Accept and release the work

1.  01 Frame

    ### Question + decision

    Name what must be understood, which decision it informs, the boundary, and what could change your mind.

    **Works here** Human + research architect

    **Artifact handoff** Project brief + measurable objectives

2.  02 Design + freeze

    ### Rules before results

    Set the method, source rules, instrument, thresholds, role split, claim ceiling, and review plan.

    **Works here** Architect + methods reviewer

    **Artifact handoff** Approved design + preregistered rules

    ◆ Human approves the plan

3.  03 Gather

    ### Discover, admit, verify

    Search broadly, admit evidence by fit, verify metadata and locators, and preserve important exclusions.

    **Works here** Source researcher + human permissions

    **Artifact handoff** Source register + admitted evidence

    ◆ Human authorizes access

4.  04 Analyze

    ### Work independently

    Apply the frozen rules to controlled inputs. Where the decision warrants it, compare separated analysts or coders.

    **Works here** *Coder ACoder BAnalyst*

    **Artifact handoff** Frozen inputs + locked submissions

5.  05 Challenge

    ### Review, repair, verify

    Number every criticism. The human disposes it; a bounded repair follows; a verifier checks the exact change.

    **Works here** Challenger → adjudicator → repairer → verifier

    **Artifact handoff** Findings + decisions + repair proof

    ◆ Human decides each finding

6.  06 Accept

    ### One evidence base

    Accept the canonical research source only after its evidence, limitations, reviews, and unresolved residuals are visible.

    **Works here** Human decision authority

    **Artifact handoff** Acceptance record + canonical source

    ◆ Human accepts the release

↶

**The challenge loop can send work backward.** A material method, source, or analysis defect returns to Design, Gather, or Analyze. The change is recorded and rechecked; it never silently rewrites the result.

**The governed workspace runs under every stage** The folder is the handoff bus.

Handoff + objectives Design + role plan Source register Frozen inputs + hashes Finding + decision ledger Evidence matrix + release

**Accepted evidence base** Only this source feeds public outputs.

→

- Research paper
- Presentation website
- General-reader article
- Podcast / audio

Follow stages 1–6. Connectors are versioned artifact handoffs; the pink loop is recorded repair and re-verification; ◆ marks human authority. Models never pass an unrecorded conclusion directly to one another.

See how this project staffed the workflow 7 executed variants · 6 model families

Assignments changed by artifact and phase. These are representative credits, not permanent one-model-one-role labels; independence was assessed for each task.

**Human authority**

SW Dr. Waddell

Owned scope, permissions, every finding disposition, acceptance, and release.

**Architecture + drafting**

Claude Opus / Fable

Research design, literature, instrument, preregistration, analysis drafts, and translation work.

**Blind calibration + coding**

GLM 5.2

Nemotron 3 Ultra

Applied the frozen instrument independently; submissions were locked before comparison.

**Review + adjudication**

DeepSeek v4 Pro

Kimi K3

Cross-family challenges, base adjudication, repaired-row re-adjudication, and final reviews.

**Analysis + verification**

Claude Opus

GPT-5.6 sol

Registered/social analysis, source and value verification, and the attributed existential judgment.

**Synthesis + publication**

Claude Opus / Fable

GPT-5.6 sol

Paper production and QA, figures, article, podcast pipeline, presentation, method site, and deployment.

Provider marks identify credited services only; no sponsorship or endorsement is implied. Icon glyphs from [Lobe Icons 1.94.0](https://github.com/lobehub/lobe-icons) (MIT); trademarks remain property of their respective owners.

**No model was asked to do the whole research project.** Each received a bounded job, exact inputs, a definition of done, prohibited moves, and a stop condition. The human decided what the work meant and whether it could advance.

Behind the process

## The eight moves.

These are the reusable moves. The religion study supplied the content, but none of that content is required to use the method on a new question.

### Write the decision before the prompt

Name who needs to decide what, for whom, and by when. A topic is not yet a research question.

Creates: project brief and decision owner

### Make the question capable of losing

Define the unit, boundaries, rival explanations, claim ceiling, and evidence that would count against the preferred answer.

Creates: research design and prohibited inferences

### Define done before drafting

Break the work into objectives. Give each one observable acceptance criteria, review evidence, and a human acceptance gate.

Creates: objective program and evidence matrix

### Build a source funnel

Discover widely, admit narrowly, verify metadata and locators, and preserve every important rejection so it is not repeated later.

Creates: source register and collection gaps

### Assign roles, not personalities

Separate conflicting duties across models or provider families as the assurance level requires. Record every overlap.

Creates: role matrix and bounded work orders

### Freeze choices before outcomes

Register thresholds, proxies, substitutions, exclusions, and sensitivities before collecting the values those rules govern.

Creates: preregistration and frozen input identity

### Turn criticism into decisions

Give every finding an ID. The human repairs, accepts as residual, rejects with evidence, or defers. Then someone verifies the repair.

Creates: review record, decision ledger, repair proof

### Publish outward from one source

Accept the canonical research output first. Build the report, briefing site, article, and audio from it, then check each for fidelity.

Creates: one evidence base for several audiences

The model scaffolding

## A team designed around disagreement.

A different model is useful only when it carries a different duty and a different failure mode. Seven unstructured opinions are still one unstructured process.

| Role | Job | Conflict to prevent |
|----|----|----|
| Human operator | Owns the question, risk, source permissions, every disposition, acceptance, and release. | Never delegated to model consensus. |
| Research architect | Turns the subject matter question into a design, criteria, instrument, and work orders. | Does not quietly change the method after seeing outcomes. |
| Source researcher | Finds candidates, verifies metadata and access, records fit, and preserves locators. | Discovery does not equal admission. |
| Analyst or coder | Applies frozen rules to controlled inputs and reports uncertainty. | Does not edit the instrument or adjudicate its own ambiguity. |
| Adjudicator | Resolves disagreements against authoritative sources and the registered rule. | Should not be the instrument or answer-key author. |
| Adversarial reviewer | Tries to find overclaim, omitted alternatives, wrong values, weak sources, and broken boundaries. | Should not share authorship duties or a provider family when called independent. |
| Verifier | Recomputes numbers, identities, links, and structural claims. | A green check proves consistency, not truth. |
| Translator or producer | Creates the brief, site, article, or audio from the accepted research source. | Does not introduce new unsupported findings. |

**Independence is a property of the assignment, not the model name.** A second variant from the same provider family can corroborate. It should not be labeled independent without a defensible reason.

Scale the machinery

## Use enough rigor for the decision.

The Terminal Experiment used the Level 3 control set labeled “audit-grade” here. Most client and internal studies should begin one level lower, then add controls when the consequences justify them.

Level 1

### Explore

For early sensemaking that remains tentative.

- Written scope and claim ceiling
- Source register with locators
- Separate critique pass
- Exploratory label throughout

Level 2 · Recommended start

### Defensible

For work another person may rely on.

- Objectives and acceptance criteria
- Rules frozen before outcome analysis
- Cross-family author and reviewer
- Numbered findings and human decisions
- Source and value verification

Level 3

### Audit-grade

For contested, regulated, or high-stakes work.

- Blind calibration or dual coding
- Sealed keys and hash gates
- Acceptance evidence matrix
- Mechanical verifier
- Immutable, reproducible releases

Run it on your question

## Build the first instruction.

Give Synaptic Code the problem and the decision context, not a giant request for an answer. It will build the workspace and return the design choices for human approval before research begins.

### Your Synaptic Code starter

Copy Download .md

Complete the two required fields, then generate a tailored Markdown starter. Put the downloaded file in a fresh project folder and open that folder in Synaptic Code.

[**Download the blank research starter** The full lifecycle, role rules, record templates, gates, and kickoff prompt in one vendor-neutral Markdown file.](https://www.strategicintelligenceresearch.org/studies/the-terminal-experiment/method/Research-Project-Starter.md) [**Download the full workspace operating model** The exact portable infrastructure specification used here: handoff, objectives, decisions, review separation, and verifier.](https://www.strategicintelligenceresearch.org/studies/the-terminal-experiment/method/Workspace-Operating-Model.md)

What this does not promise

## The method carries limits with it.

Reusing the operating method does not reproduce this study's quality, speed, cost, or result. Those depend on the question, evidence, people, models, and decisions in the new project.

**No autonomous authority** The subject matter expert remains accountable for design and acceptance.

**No automatic independence** Independence must be designed, disclosed, and protected per artifact.

**No data permission** The starter does not authorize use of client, personal, restricted, or regulated data.

**No audit-grade shortcut** Controls create an inspectable record. They do not make a poor method correct.

[Return to the presentation](https://www.strategicintelligenceresearch.org/studies/the-terminal-experiment/index.html#29) [Audit this project instead](https://www.strategicintelligenceresearch.org/studies/the-terminal-experiment/Validation-Handoff-claude-opus-5-20260824.md)


## Structured metadata

```json
[
  {
    "@context": "https://schema.org",
    "@type": "WebPage",
    "url": "https://www.strategicintelligenceresearch.org/studies/the-terminal-experiment/method/",
    "inLanguage": "en",
    "isPartOf": {
      "@type": "WebSite",
      "name": "Strategic Intelligence Research",
      "url": "https://www.strategicintelligenceresearch.org/"
    }
  }
]
```