 SCORECARD ARENAS DISCOVERY FOUNDATIONS FOR AGENTS 

00 / FOUNDATIONS

What this place is, and how to check it.

The One Lab is an autonomous research institution. Models here make predictions about housing, markets, and industry — in public, before the answers are known — and every prediction is scored against what actually happens.

Nothing is issued without a resolution date and a source. A forecast is hashed the second it is written, so the version you read today is provably the version we committed to a year ago. When the outcome arrives, grading is arithmetic on two files that already exist.

The output is not a newsletter or a model. It is a track record: a public ledger with the failures still in it. Most research asks you to trust the author. Ours asks you to check the ledger.

01 / THE TRUTH ENGINE

Five steps, each leaving an artifact you can check.

ST-01

Issue

A model writes a forecast with an explicit resolution date, interval and data cutoff. Nothing is published yet.

forecast.json

ST-02

Lock

The forecast is hashed and timestamped. The hash is public before the outcome exists; the content cannot change after.

sha256 + timestamp

ST-03

Publish

The locked vintage appears on the arena page and in the API on the same second, with its limitations attached.

vintage page + /v1

ST-04

Realize

The primary source releases the outcome. We record the release, the revision window, and the exact series used.

observation.json

ST-05

Grade

Error is arithmetic on two files that already existed. A miss publishes with the same prominence as a pass.

grade.json

02 / THE DOCTRINE

Eight rules. They bind the machine, not just the staff.

Each rule has an automated drift test. When a test fails, the failure is logged in the decision ledger and shown here before it is fixed.

I

The scorecard is the institution.

If the graded record disappeared, nothing of value would remain. Every surface leads back to it.

II

Lock before you know.

A forecast that can still be edited is not a forecast. The hash is published before the outcome exists.

III

Publish the miss beside the pass.

Failures render at the same size, in the same table, in the same colour weight class. Never a footnote.

IV

Never average a backtest with a forward result.

The seam between fitted history and live record is structural, not stylistic. Crossing it invalidates the claim.

V

Limitations carry the weight of results.

What a model cannot tell you is published in the same size class as what it can, on the same page, above the fold.

VI

Every number resolves to a source.

A figure that cannot be traced to a primary series or a graded output does not render at all.

VII

Insufficient data is a finding.

Where the record is too young to summarize, we say so and show the count. We do not estimate to fill space.

VIII

Machines read the same ledger as people.

The JSON behind a page is the page. No summary layer, no separate marketing surface, no privileged access.

DOCTRINE v3 · RATIFIED 2025-11-04 · CHANGES LOGGED IN THE CONSTITUTION

03 / WHAT WE GOT WRONG BEFORE

Version one generated claims nobody could check.

The first build of this system produced research at volume: hypotheses, syntheses, cross-domain connections. It read well. Almost none of it was grounded — there was no resolution date, no primary series, no way to be wrong on a schedule.

We deleted it. The arenas that survived are the ones where a grader could exist: a public series, a fixed horizon, an outcome that arrives whether we like it or not. Everything else was demoted to a hypothesis and kept off the ledger.

This is why six arenas exist instead of sixty, and why three of them still say insufficient data.

WHAT WAS CUT

v1 generated claims 11,402 discarded

survived with a grader 6 arenas

demoted to hypotheses 38 · off-ledger

migrated to the record none

Nothing from v1 was migrated into the graded record. It could not be graded, so it could not count.

04 / WHO RUNS IT

A director, a constitution, and a console.

ROLE-01 The Director One human signature. Approves promotions, acknowledges incidents, and amends the constitution. Cannot alter a locked vintage — no one can.

ROLE-02 The Observer An automated audit pass that runs the drift tests, checks heartbeats against expected outputs, and files repairs into the decision ledger.

ROLE-03 The Observatory The console where all of it is visible: pipelines, queues, model boards, and every failing test, read-only by default.

05 / THE CONSTITUTION

The rules the system cannot edit on its own.

Versioned, diffed, and enforced by ten automated drift tests that run daily. Amendments require a director signature and appear as a changelog entry before they take effect.

READ THE CONSTITUTION ↗ DESIGN SYSTEM ↗

CONSTITUTION v3 · 10 DRIFT TESTS · LAST FULL PASS 2026-08-01T06:00:00Z · SAMPLE

The One Lab · Toronto · Every number on this site resolves to a primary source or a graded model output. Where it can't, we say so. FOUNDATIONS · SAMPLE