The reusable part of inquiry

Theseus publishes its conclusions, but the more durable public object is the discipline that produced them. Before the catalog and the console, this page teaches two things a first-time reader needs: the contract every published conclusion meets, and the single pipeline that produces it. Nothing here is private; everything is filtered for public visibility before it reaches this page.

Skip to the methods →

The contract every conclusion meets

The reasoning contract

Every conclusion the firm publishes keeps five things apart, so a reader can audit the reasoning instead of trusting it. The methodology exists to hold them distinct:

Claim: The single proposition being asserted, stated plainly enough that it could turn out to be wrong.
Evidence: The cited sources and live observations the claim rests on — never an appeal to authority.
Method: The named procedure that turned the evidence into the claim, so an outsider can reuse or contest it.
Objection: The strongest recorded challenge the claim has survived, kept visible rather than buried.
Revision condition: What would have to be observed for the firm to change its mind — written down before the fact.

How a claim becomes a published conclusion

The pipeline, end to end

One staged pipeline turns a corpus into principles, principles into algorithms, and live observations into graded conclusions and bets. It is the same diagram the home page and manifesto show — there is only one.

   corpus ──▶ synthesizer ──▶ principles ──▶ algorithms
                                                  │
                              live observations ──┤
                                                  ▼
                                            conclusions
                                                  │
                                                  ▼
                                               memos
                                                  │
                                                  ▼
                                         portfolio agent
                                                  │
                                                  ▼
                                                bet

Appendix — for readers who want the full console

The methodology console

Everything below the pipeline is the firm's internal methodology console, kept public and reachable but no longer the first thing a reader meets: the meta-method, the methods catalog with current status, and the empirical record those methods have earned. The deeper historical theory has moved to the methodology appendix — it is relocated, not removed.

Read the methodology appendix →

Layer 1 — what the firm believes about inquiry

The meta-method

Before any single method, the firm holds a method for judging methods: five working criteria — Progressivity, Severity, Aim-Method Fit, Compressibility, Domain Sensitivity — applied to each method so a reader can see what it is, how it has calibrated, where it composes with other methods, and where it has failed. The three surfaces below are that meta-method made inspectable.

Layer 2 — the methods, with current status

The methods catalog

Sortable. Filterable by domain. Status is the method's current standing; calibration slope is shown only for methods whose track record clears the firm's publish gate — below that, the cell is left blank instead of dressed up.

Search13 of 13 methods

	Description
classify_claim_type v1.0.0	Classifies a claim into discourse categories (METHODOLOGICAL, SUBSTANTIVE, etc.).	active	—	—	OK	2018-10-20
contradiction_geometry v1.0.0	Detects contradiction via Hoyer sparsity of embedding difference vectors.	active	—	—	OK	2018-10-20
contradiction_probe v1.0.0	Predicts the embedding-space neighborhood where a new proposition's logical contradiction should lie, then surfaces nearby existing propositions as unconfirmed candidates.	active	—	—	OK	2018-10-20
decompose_voice v1.0.0	Decomposes a founder's voice into an intellectual profile with orientation scores.	active	—	—	OK	2018-10-20
external_claim_match v1.0.0	Ingests external literature and matches claims against internal positions.	active	—	—	OK	2018-10-20
extract_claims v1.0.0	Extracts atomic truth-apt claims from a text chunk using an LLM.	active	—	—	OK	2018-10-20
extract_methodology v1.0.0	Extracts portable methodology profiles from a source text.	active	—	—	OK	2018-10-20
extract_prediction v1.0.0	Extracts falsifiable world predictions from a single claim using an LLM.	active	—	—	OK	2018-10-20
method_candidate_extractor v1.0.0	Scans ingested artifacts for passages describing a methodology and extracts structured method candidates via regex + LLM.	active	—	—	OK	—
nli_scorer v1.0.0	NLI cross-encoder scorer for claim pair coherence using DeBERTa.	active	—	—	OK	2018-10-20
six_layer_coherence v1.0.0	Six-layer coherence aggregation with 4/6 majority voting.	active	—	—	OK	2018-10-20
suggest_research v1.0.0	Generates research topics, empirical anchors, and reading lists after a discussion.	active	—	—	OK	2018-10-20
synthesize_conclusion v1.0.0	Registers a substantive conclusion and returns method calibration feedback.	active	—	—	OK	2018-10-20

Layer 3 — the empirical record the methods have earned

Benchmarks, calibration, and the tournament

A method is only as good as its record. This layer is the evidence: the firm's first-run benchmark, the cross-model results, the adversarial tournament, and the published failure modes — plus the raw manifest for outside replication.

Public failure modes

9 entries published across all methods. Each method's full catalog is reachable from its page; the firm holds private entries until the framing matures.

Public boundaries

The explorer is deliberately incomplete in one respect: it does not expose raw deliberation, private transcript text, or unreviewed chain-of-thought. It exposes the method at the level needed for critique and reuse. Material revisions create a new immutable snapshot row so prior URLs do not rot.