Doctrine · numbered for cross-examination
The method.
MIT NANDA, State of AI in Business 2025 (Jul 2025, v0.1 preliminary; 52 interviews, 153 survey responses): 95% of enterprise GenAI pilots deliver no measurable P&L impact — the 5% that do were designed for friction: governance, memory, workflows rebuilt around the work, instead of erased. The figure is contested. Report · Critique. North is a designed-friction product. This page is the doctrine: five commitments, each a mechanism in the code, not a mood in the copy.
Determinism first
Same inputs, same evidence. Every time.
Readings are computed by a deterministic scoring model — scenario weights, dimension scores, bands, pre-mortem risks: pure functions over the answers your room gave. No model call sits in the scoring path, so nothing about your reading depends on the mood of a language model on a Tuesday.
Claims about model behaviour are probed, and the probes are reproducible on demand. A run manifest pins the harness version, config, seed, target identity, and served model id; the seed alone drives probe selection and sampling order. Two runs with the same seed produce byte-identical report hashes. Below 30 seeded attempts the suite refuses to run — that is not evidence.
Mechanics · seeded probe suite · re-runnable on demand
Sealed, never blended
The snapshot seals. The scores never feed the machine.
When an instrument finishes, the answers seal server-side into a tamper-evident snapshot — canonical bytes, keyed seal, append-only. The seal carries a contract written into the schema itself: classificationUse: "none". Scores, bands, and archetypes never enter classification, never reach the drafter, never touch the kernel's inputs — any other value fails the seal.
What happens in the room stays in the blob. Identity leaves it only for the partner with an ownership claim — a predicate, not a courtesy: unknown sessions and foreign sessions both answer 404. Sales cannot quietly reprice what the room answered, and the machine cannot quietly re-cut what the room meant.
Humans ratify
Agents draft. Nothing auto-approves.
Every decision gate is an ordered proof chain. The CISO signs first; the auditor signs second — and the auditor's signature covers the CISO's, so the attestation is nested, not parallel. Risk, CIO, and business may countersign afterwards, in roster order: roles may be skipped, never reordered, never repeated.
The order is enforced, not conventional. Sign out of order and the API answers 409 chain_order and writes nothing. Each signature is ed25519 over the canonical payload plus the prior signature, so the whole chain verifies offline. Nothing ships because a model was confident.
Evidence that outlives us
If North disappeared tomorrow, the ledger segment still recomputes.
Claims worth making get notarised. Each tenant keeps one hash-chained notary log: every entry linked to the previous by sha256 over canonical JSON, the whole chain recomputed from its genesis hash on every lookup. Export the ledger segment and the same math runs offline — no account, no API required.
What verifies today: structure and signature — recording, position, chain integrity, and the Ed25519 receipt signature against the key registry NORTH publishes.
Proof, honestly
We publish our misses.
Every prediction North makes is scored at rescore: improved, held, worsened — or inconclusive. The inconclusive rate ships with the record, because the willingness to publish it is the only trust signal a young evidence company actually owns.
Until the track record exists at scale, every number on this site wears its provenance:
No testimonials yet. Here is our method instead.
The door
Run the instrument.
The document read is the door, not a form. Send the governance documents you already have — even blank — and a deterministic read returns in days, every finding pinned to your own file and section, sealed with a receipt.