How the firm shows its work
One page to judge whether the firm’s reasoning has held up. A reader who asks “how do I know this is right, and what happens when it isn’t?” should be able to answer it without taking our word for it: the calibration record, the changes of mind, the critiques that landed, and the cryptographic trail under every publication — all here.
Calibration
The firm scores every resolved forecast and publishes the aggregate Brier score, the reliability curve, and the per-method breakdown — with a bootstrap confidence interval and a hash over the exact resolution set, so the scoreboard can’t be cherry-picked after the fact.
Open the full calibration scorecard →
Calibration by forecast horizon → · Knowledge graph →
The same accountability picture, from the other side: the live forecasts feed (every open prediction the scorecard will eventually grade) and the algorithms register (the scored decision procedures behind those forecasts).
Revisions
Theseus treats changes of mind as part of the record. A revision event explains the evidence that moved, the conclusions that changed, and the confidence delta that followed. Publication signatures and revision records are cross-checks — neither should silently overwrite the other.
- Evidence that changed — the source inputs and weights that triggered a material update.
- Conclusions affected — each revised conclusion links back to its public or founder surface, so the causal trail stays navigable.
- Proof and signatures — a revision never erases the signed history it supersedes.
Individual revision events are addressable at /revisions/<event-id>.
Critiques
The firm’s methodological edge depends on inviting the strongest possible external critique. Critics whose challenges land are credited here. Severe critiques (per the rubric below) carry a $500 bounty — paid to the critic, or donated to a charity the critic chose.
Bounty rubric
Severity is scored from structural inputs (cascade weight, claim centrality, curated failure-mode match, source credibility) and an LLM judge that is capped by the structural bracket. The judge can place inside the bracket; it cannot promote a nitpick into a high.
- low— accepted with credit; no bounty. Real, but does not move the conclusion’s confidence by > δ.
- medium — accepted with credit; no bounty. Often paired with a private discussion or an article addendum.
- high— accepted with credit AND a $500 bounty. The critique attacks a load-bearing claim and the firm has updated its position (typically a revision-engine pass changed the conclusion’s headline confidence).
Bounty payment is gated by founder confirmation. The codex queues the payout; the firm’s payouts pipeline is the eventual sender.
How to file
Open any published article and use the Challenge this conclusion affordance. Bring: which specific claim, what counter-evidence, what method you used to derive it, citations.
Accepted critiques
No accepted critiques yet.
Be the first — the response form lives on every published article.
Provenance
Every public Theseus publication carries a cryptographic provenance trail. A reader who asks “is this what Theseus actually said, on this date, with this evidence base?” should be able to answer it without taking our word for it. This section is the contract.
What gets signed
At the moment a conclusion or article transitions to published, the noosphere CLI computes a SHA-256 over a canonical encoding of:
- the conclusion text (Markdown, normalized for whitespace)
- the methodology profile id(s) attached to it
- the citation set
- both discounted and stated confidence
- the methodology quality score (MQS, prompt 01) at the time of publication
- the publication timestamp, slug, and version
The hash is then signed with an Ed25519 key the firm controls. Both the canonical input and the signature are persisted alongside the published row, and served verbatim from a public endpoint:
GET /api/public/signature/<slug>?version=<n>Canonicalization rules
For the verifier to land on the exact same hash bytes as the signer, the canonicalizer is code-defined and identical on both sides (noosphere/ledger/canonicalize.py and the TS mirror in publicationService.ts):
- Markdown is Unicode-NFC-normalized; CRLF and CR collapse to LF.
- Trailing whitespace is stripped from every line.
- Runs of three or more blank lines collapse to one blank line.
- Leading and trailing blank lines are stripped.
- JSON is encoded with sorted keys and tight separators.
- Floats are rounded to 6 decimal places.
- Citations are sorted by (format, block) so reorder is invisible but adds/removes are not.
- Timestamps are rendered as second-precision ISO-8601 Z (UTC).
Key management & rotation
Private signing keys never live in this web app. They live on the operator’s machine under ~/.theseus/keys/publication/. New publications are signed with the active key. Older keys remain in the keyring so historical material continues to verify; rotation generates a fresh active key without invalidating anything previously signed. A key can be revoked, in which case publications signed after the revocation timestamp fail verification, but publications signed before it remain valid (the historical record stays trustworthy).
How to verify a publication yourself
Install the noosphere CLI from the firm’s package, then run:
noosphere ledger verify-publication <slug>
noosphere ledger verify-publication <slug> --from-url https://<host>/api/public/signature/<slug>The CLI fetches the signature, recomputes the canonical hash from the live database row, and checks the Ed25519 signature against the keyring. A mismatch means either the database has been mutated since signing, or the signature is stale (canonical inputs changed and the publication needs to be re-signed). Either is a bug the firm wants to know about; we surface it with a red banner on the publication page rather than silently hiding the failure.
Source
Canonicalizer: noosphere/noosphere/ledger/canonicalize.py · Signing primitives: noosphere/noosphere/ledger/publication_signing.py · Web mirror: theseus-codex/src/lib/publicationService.ts