Built to be checked.
BRIER is the firm's own system. It keeps every call on a hash-chained ledger. Each morning it seals that ledger with a public timestamp anchored in Bitcoin. It grades each call against a source named in advance. Anyone can check the result without trusting us.
Every diagram on this page shows what runs today, read from the record at build. Parts still being built are marked.
- The ledgerAppend-only and hash-chained. Each approval names who, when and why.Live
- Daily anchorSealed at 07:00 each day on the public OpenTimestamps calendars, committed to Bitcoin.Live
- Public record223 calls · fingerprint 10dfa35cc0e9Live
- Snapshots46 published versions, each named by its fingerprintLive
- Selection437 statements reviewed · 223 enteredLive
- Domain desks6 desks · 64 reportsLive
- World data42 series · 14 early-warning indicatorsLive
- Forecast LabBlind framings, preregistered. Shipped 21 October 2026Trial
- LineageEach role carries forward what it learnedIn build
- Agent cellA disposable cell per job, default-deny networkIn build
Append-only. Sealed every morning.
Every act that matters is an entry: a call locked, an outcome recorded, an approval given. Each entry carries the hash of the one before it. The record can only grow. Nothing edits it. A correction is a new entry.
What the seal rules out
- Edit an entryThe chain breaks at that entry. Every entry after it no longer matches.
- Remove entriesThe count no longer matches a statement already made.
- Rewrite and resealA new statement cannot be backdated. Its Bitcoin block shows when it first existed.
A call moves forward only.
Each call passes through five states. What is fixed at each one never changes afterward. The counts are today's.
- 01Candidate437 reviewed
A sentence from a report, tested against the standard. Private working material, never shown.
- 02Locked223 entered
The statement, its probability, its date and the rule that will settle it are fixed.
- 03Open186 waiting
The window runs. A named public source has not settled it yet.
- 04Resolved37 graded
A named person records the outcome from the source the rule named. The evidence is linked.
- 05Scoredfrom 30 families
Eligible calls enter the public score once enough related groups are graded.
Sources of equal standing disagree. The call stays open and adds nothing until the founders rule.
A call found wrong in its probability, premise or outcome is voided in the open. It keeps its reason and a correction record. It is counted separately.
What did not qualify stays in the count.
A forecaster can look good by keeping only the easy calls. So every statement reviewed is counted. Each one that failed the standard is counted by its reason. Selection cannot hide in the denominator.
One square per statement reviewed, 437 in all.
- Entered the record223
- A scenario, not a probability108
- No unambiguous window67
- No probability21
- Not publicly resolvable8
- Another failure of the standard7
- Duplicate or overlapping3
Check it without taking us on trust.
The whole public record is one file with one fingerprint. Your own browser can compute that fingerprint and compare it. Try it here.
- DownloadYour browser fetches the public record, one JSON file.
- FingerprintIt computes the file's SHA-256 itself. No server does it for you.
- CompareThe result must equal the published fingerprint, character for character.
- Go back in timeEvery earlier version is kept as a snapshot named by its own fingerprint: 46 so far.
10dfa35cc0e9e6ca257acb14db0ef714a1a9837a4b92cb658fb6130e0588b81cYour browser's answernot yet computedOne core remembers. One core reasons.
The reasoning core does not work from a model's general knowledge alone. It anchors each case on a reference class from the memory core. Each class carries its sample size and sources. A named person decides what becomes a call.
What it knows
- SemanticBase rates for defined reference classes, each with its sample size, window and sources
- EpisodicThe firm's own scored experience
- ProceduralThe operating standard: how the work is done
- WorkingThe live problem
The calibration engine
- Anchoron the reference class the case belongs to
- Weighthe evidence against the firm's scored experience
- Runthe futures, and the strongest case against
- Expressa probability or an interval
A named person
Every probability is a proposal until a named Principal of the firm locks it. Outcomes return to memory only as scored evidence.
Testing whether independent framings forecast better.
TrialShipped 21 October 2026
The trial was preregistered before any outcome. Nothing it produces enters the firm's public score. It reports no result before its final read.
Graded by a rule that punishes confident misses.
Every resolved probability call is scored with the Brier score: the squared distance between the probability and what happened. A calibrated forecaster's 70% calls come true about 70% of the time.
score = (probability − outcome)2
A call at 80% that comes true scores 0.04. The same call that does not scores 0.64. Saying 50% to everything scores 0.25. Lower is better. Confident misses cost the most.
Shown from 30 graded families.
Related calls count as one family. One event cannot pass for many tests. With few graded families, even a skilled forecaster's score can land almost anywhere. Move the slider to see how the range narrows.
An illustration with stated assumptions: a forecaster whose true score is 0.17 and a typical spread of single-call scores. It is not the firm's record.
From public data to a locked call.
Each desk covers one domain. Every figure it uses carries its source and how far that source can be trusted. A reading with no live source says so.
- Named sourcesEach with a reliability tier
- Collected daily42 series, 14 indicators
- Six desksAnalysts and AI, one domain each
- Reports64 so far
- Calls223 on the record
| Desk | Reports | Calls | Graded |
|---|---|---|---|
| Economic & Financial Markets | 13 | 53 | 16 |
| Supply Chain & Logistics | 11 | 31 | 4 |
| Mining & Manufacturing | 10 | 39 | 4 |
| Infrastructure & Catastrophe Risk | 9 | 24 | 1 |
| Political Risk & Governance | 12 | 40 | 4 |
| Defense & Security | 10 | 36 | 8 |
AI drafts. It never approves.
AI agents do much of the work, inside limits they cannot change. The system enforces each layer below. None depends on the agent behaving.
- Tests fixed firstbefore any code is written
- Builderin its own workspace
- Independent checkerfresh context; never reads the builder's account
- Two fix roundsat most
- A named Principalmerges and installs
A disposable cell for every job. It mounts one workspace and nothing else. Its only way out is a proxy that denies by default and logs every request where the agent cannot reach.
Lineage: a system that remembers what it learned.
Each role in the system, the analyst, the checker and the red team, will carry forward what it learned. Every lesson is earned on evidence, checked by another lineage and accepted by a person. It will be measured against a control group that starts with none.
- 01InheritA session starts from its role's accepted lessons, read only
- 02WorkIt does the job inside its fence
- 03ProposeIt may suggest new lessons backed by evidence the agent could not have written
- 04CheckA reviewer from another lineage tests each one
- 05AcceptA named Principal accepts, edits or declines
- 06SignThe app signs a new version, hash-chained
Lineage is data, never instruction. No agent can write to it. A share of sessions always runs with none as a control. Both are graded on the same questions.
Run on your own numbers.
BRIER Intelligence puts this system to work on your own business, in a walled space of its own. Every recommendation is written down and graded. Your people decide.
Our own record is open to anyone.
Most advisors tell you what they think. We write it down first and show you how often we were right.