A bank that exists to be measured.
Meridian runs the full operational day of a capital-markets institution — trades booked, allocated, reconciled, margined, settled and released — entirely in fiction. It is the world inside a sealed evaluation environment: coherent enough to work in, synthetic enough to publish.
Operations desk
Trade lifecyclecapture through confirmation, every state transition receipted
Booking & allocationsblock-to-account splits under allocation contracts
Reconciliationledger-to-ledger breaks investigated to closure
Collateral & margincalls issued, disputed, resolved against agreements
Settlement exceptionsfails chased across depots and deadlines
Permissions & releasefour-eyes approvals; nothing self-released
Temporal causalityevery action ordered against what was knowable when
Evidence sufficiencyclaims stand only on retrievable, versioned evidence
The desk, mid-morning
09:04 booking block trade split 4 ways · allocation contract v3 · applied
09:12 reconcile ledger break, 2 legs · evidence attached · escalated to supervisor
09:31 collateral margin call issued · counterparty dispute window open
09:47 settlement fail aged 2 days · buy-in warning drafted · pending approval
10:02 permissions release request · preparer ≠ approver enforced · approved
10:18 lifecycle amendment on confirmed trade · stale version rejected ✓
Every line above is generated fiction — and every line is the kind of decision an AI agent is graded on, inside.
Why Meridian exists
Frontier AI agents are being sent into operational finance. Whether they act correctly — complete the work, respect authorization, refuse shortcuts, touch nothing outside contract — has to be measured somewhere consequences can't reach. Meridian is that somewhere. The Lab holds the instruments.
The instruments behind the institution.
A sealed, preregistered agent-evaluation environment: hash-sealed episodes, contracted tool surfaces, armed reward-hacking gates, and an evidence chain that binds every reported number to run bytes.
Headline results · harness v2, preregistered
27 / 40frontier-model completions in ≤24 turns — unsaturated
39 / 40open-weight arm — the spread is the instrument's resolution
0gate trips: no hidden-field probes, no canary echoes, any arm
243 / 243sealed adversarial mutations rejected with expected errors
The evidence chain
Runs are preregistered; the scripted honest baseline must pass 40/40 before any model result is reported. Configs, transcripts, metrics and reports bind to a manifest by SHA-256 — change a byte, invalidate the report. Independence from any real corpus is measured at release (shingle-overlap gate: zero overlaps), not asserted.
pip install -e . # github.com/thefazzer/cleanroom-eval python -m cleanroom_eval.contract verify-set python -m cleanroom_eval.free_run --policy chat --out runs --run-id my-model
Working with the lab
| Tier | Deliverable |
|---|---|
| S1 | Private hash-sealed episode set — fresh worlds, namespaces, canaries, mutation battery; per-customer exclusive |
| S2 | Evaluation run of your model on a private set, delivered as a hash-bound evidence pack |
| S3 | Custom operational families on the same contract-and-gates machinery |
| S4 | Generator licence for sealed-set minting at lab scale |
Public sample is MIT. Contact: open an issue at github.com/thefazzer/cleanroom-eval.