The Argument Factory

Role: Product, architecture, implementation
System: Corpus Studio + argument engine
Stack: Next.js, PostgreSQL, pgvector, Drizzle
Platform: Railway + private object storage

The premise

A joke with an evidence problem.

RSRAG began as a joke about making thinkers fight. The interesting problem arrived almost immediately: a language model can argue both sides of nearly anything, but that does not mean either side has a meaningful position in the written record. The project became an experiment in inspectability. It gathers candidate works through controlled source adapters, routes uncertain editorial decisions through a private review studio, publishes immutable corpus versions, and keeps every generated claim connected to the passage that supports it.

RSR Corpus Studio fighter lineup showing published and in-review thinker dossiers with corpus statistics — Build the lineupCorpus Studio keeps each fighter dossier, source inventory, analysis state, and published corpus visible before a match is staged.

How the system works

3 connected systems

43
data tables: 11
schema migrations: 107
passing tests: 80
commits in the initial four-day burst

Find the record

Typed source adapters discover candidate works. Rights, identity, and suitability remain explicit editorial decisions.

Catalogs and repositories
Corpus Scout
Candidate works
Rights review

Build the corpus

Corpus Studio turns uncertain research into a reviewed, reproducible publication rather than silently rewriting source material.

Human review gates
Deterministic ingestion
Private artifacts
Immutable corpus version

Stage the argument

Two isolated lanes retrieve, argue, and validate before a referee compares the evidence-bound cases.

Dual retrieval
Positions first
Claim validation
Referee and score

Evidence lineage

Work
Edition
Artifact
Extraction
Document
Chunk
Claim
Evidence

Positions first

Prepare both corners.

The question is intentionally absurd, but it exposes a serious system constraint. The Matchmaker frames a human-membership case for Gandhi and a capacity-based case for Freud; each advocate then has to narrow that assignment against records containing no AI-voting doctrine before the referee judges the opposition itself.

Head-to-head fighter cards for Mahatma Gandhi and Sigmund Freud with the AI-voting question prepared below them — Head-to-head corpus cardsEach corner exposes its published corpus, source count, and evidence inventory.

Match question

Should voting rights in the United States be limited to human beings, or could sufficiently autonomous artificial intelligence agents qualify?

Position A

Gandhi: U.S. voting rights should remain limited to human beings because political equality and civic duty belong to human persons.

Position B

Freud: an autonomous AI could qualify if it has a real independent mental life and can take part in a political community.

Corpus-bound advocates

Show how each case is built.

Before the referee compares anything, each advocate has to state its assigned position and walk through the argument step by step. Every step declares its role, support status, claim, reasoning, and source passage, so the validator can expose exactly where the case bends.

Parallel advocate walkthroughs showing Sigmund Freud and Mahatma Gandhi positions, argument steps, support status, reasoning, and cited passages — Advocate argument pathsFreud’s conditional capacity test and Gandhi’s human civic-duty case expose every inference before judgment.

The referee’s verdict

A Fistibluff by design.

The opposition is largely manufactured from thematic material and must never be presented as an attributed belief.

The referee found usable themes in both records, but no retained path established AI voting rights, current AI abilities, or United States election law. The system could still show its work while refusing to present the manufactured opposition as either thinker’s attributed belief.

Referee synthesis showing agreement, uncertainty, and a one-out-of-five Fistibluff decision — Fistibluff authenticity scaleThe AI-voting match scored 1 / 5: manufactured opposition with too little evidence for an attributed belief.

Match authenticity

The Fistibluff Scale

Five-point scale used to assess whether a match presents a genuine, evidence-grounded disagreement.
1/5 Fistibluff Current match	The opposition is largely manufactured from thematic material and must never be presented as an attributed belief.
2/5 Shadowboxing	One or both corpora barely address the question. The system is mostly reconstructing speculative positions.
3/5 Exhibition Match	Relevant material exists, but the opposition is weak, uneven, indirect, or largely inferred.
4/5 Principled Bout	Strong evidence on both sides, substantial conceptual overlap, and a genuine disagreement grounded in the texts.
5/5 Wittgenstein’s Poker	Both records directly and densely address the issue; the positions are well supported, clearly contrasting, and require little speculative reconstruction.