Thesis Arena

Turn any claim into a multi-agent stress test that exposes weak logic, missing evidence, and stronger arguments.

Thesis Arena is a reasoning instrument for pressure-testing an idea before you publish, teach, debate, or make a decision around it. It decomposes one thesis into claims, examines those claims through independent analytical lenses, and rebuilds the result into a more defensible formulation.

This is not a chatbot thread and not four personalities producing unrelated opinions. Every stage returns a structured artifact that is validated, saved, rendered progressively, and reused by the final synthesis.

OpenAI Build Week track: Education
Live demo: protected judge access is supplied privately in the Devpost submission.

Product flow

1. Claim Map separates empirical, causal, interpretive, predictive, and value claims using a Toulmin-inspired structure.
2. Red Team finds hidden assumptions, counterexamples, omitted alternatives, and failure modes.
3. Steelman produces the strongest defensible version of the argument, including qualifications and missing warrants.
4. Logic checks consistency, causality, scope, and inferential jumps.
5. Fact Check separates externally verifiable claims from opinion and returns evidence statuses with source links.
6. Synthesis records what survived, what failed, the hidden assumptions, the key crux, and an improved thesis.

The compact cards make the result scannable. Each card opens into a full report whose count is derived from the actual findings array. The final analysis can be exported as a print-ready PDF or sent through another stress-test cycle.

Product walkthrough

1. Map the argument

The Claim Mapper separates a complex thesis into individually inspectable claims and labels the kind of reasoning each claim requires.

2. Run independent analytical lenses

Red Team, Steelman, Logic, and Fact Check examine the same claim map from four different methodological perspectives. Results arrive progressively, so completed modules remain readable while the rest of the workflow continues.

Inspect a cited Fact Check report

Each externally verifiable claim receives an evidence status, explanation, and inspectable source links rather than an unsupported confidence score.

3. Turn critique into a stronger thesis

The final synthesis records what survived, what failed, the hidden assumptions, and the Key Crux before producing a more defensible formulation that can enter another stress-test cycle.

Architecture

The browser starts a run and receives an execution ID immediately. n8n saves each checkpoint to PostgreSQL while the browser polls a lightweight status workflow. Completed modules appear as soon as they are available, so the user can read Red Team or Steelman while the remaining analysis continues.

If a run is interrupted, the retry workflow resumes the saved n8n execution instead of repeating already completed paid model stages.

Technical highlights

• Six specialized GPT-5.6 Luna nodes, including cited online Fact Check.
• Strict shared JSON contract across n8n, frontend, mocks, storage, and reports.
• Progressive delivery of Claim Map and analytical lenses.
• Saved-execution retry from the last successful checkpoint.
• PostgreSQL run archive with model, duration, stage, error, and final payload.
• Atomic account quota check before any paid model node runs.
• Nginx isolation, Basic Auth, rate limiting, and server-side secret injection.
• Dedicated A4 HTML/CSS report instead of screenshot-based PDF generation.
• Dependency-free browser application and deterministic local mock mode.

Run locally

The local version requires no API keys, package installation, or build step.

node server.cjs

Open http://127.0.0.1:4173.

The default config.js uses deterministic mock data, including detailed lens findings and Fact Check sources clearly marked as mock evidence. This is the recommended path for reviewing the UI and data contract without spending model credits.

Connect n8n

1. Apply database/schema.sql to PostgreSQL.
2. Create OpenRouter, PostgreSQL, Header Auth, and n8n API credentials inside n8n.
3. Copy .env.example to .env and provide the corresponding n8n credential IDs. Do not put API keys in these variables.
4. Generate credential-bound workflow exports:

node scripts/build-n8n-workflow.mjs
node scripts/build-n8n-control-workflows.mjs

5. Import and activate the three files from n8n/.
6. Configure the browser API boundary in config.js or use the same-origin Nginx example in deploy/.

The committed workflow JSON files are deliberately credential-free. After import, n8n will ask you to select credentials for the affected nodes.

Models

The live analytical layer uses:

Claim Mapper   openai/gpt-5.6-luna
Red Team       openai/gpt-5.6-luna
Steelman       openai/gpt-5.6-luna
Logic          openai/gpt-5.6-luna
Fact Check     openai/gpt-5.6-luna:online
Synthesizer    openai/gpt-5.6-luna

OpenRouter is used as the transport layer. The analytical model is GPT-5.6, and the final run metadata records the model identifiers that were configured.

Data contract

analysis-schema.json is the canonical contract. UI rendering, report rendering, mocks, PostgreSQL payloads, and n8n outputs use the same structure.

Important invariants and integration notes are documented in docs/DATA_CONTRACT.md.

Built with Codex and GPT-5.6

ChatGPT was used for product methodology, the analytical framework, positioning, visual direction, and review. Codex handled the implementation loop across the frontend, n8n workflow generation, PostgreSQL, Nginx, deployment, debugging, and production verification.

GPT-5.6 Luna is also the runtime reasoning layer inside the finished product.

The complete build narrative, key decisions, production challenge, and lessons learned are documented in docs/BUILT_WITH_CODEX.md.

Repository structure

index.html                 application shell
app.js                     state machine, rendering, and API client
styles.css                 responsive visual system
mock-data.js               deterministic contract-compliant fixtures
analysis-schema.json       canonical response schema
report.html / report.js    dedicated printable report
n8n/                       credential-free workflow exports
scripts/                   reproducible workflow generators
database/schema.sql        PostgreSQL persistence and quota schema
deploy/                    optional Nginx deployment example
docs/                      contract and Codex build documentation
assets/                    architecture and product media

Current scope

The Build Week MVP intentionally supports one complete interaction:

one thesis → one claim map → four lenses → one synthesis → one report

It does not include social feeds, public profiles, subscriptions, collaborative workspaces, or a general-purpose chat interface.

Security notes

• API keys and credential values never enter browser code.
• The public workflow exports contain no production credential IDs.
• Authenticated account identity is injected by the reverse proxy rather than trusted from a browser-supplied header.
• PostgreSQL records quotas and run ownership; authentication secrets are stored separately.
• The demo is rate-limited because every live analysis has a real model cost.

License

MIT