03 / ASSURANCE

capabilityis earned.

We didn't slow down to be safe. We got safe enough to go fast — in that order. Six layers stand between an idea and the world; none of them are optional, and none of them are ours alone.

0CRITICAL INCIDENTS
14NEAR-MISSES — ALL DISCLOSED
38TRIPWIRE FIRINGS — ALL PUBLISHED
52/52FULL-STOP DRILLS PASSED
100%EXTERNAL AUDIT ACCESS

THE STACK

Six layers between
an idea and the world.

01

Glass-box interpretability

We do not ship circuits we cannot read. The Microscope renders a complete causal trace for every output — not a heatmap, an explanation that predicts. If a behavior can't be attributed, the model doesn't graduate. TRACE COVERAGE: 100% OF PRODUCTION OUTPUTS

02

Proof-carrying updates

Every self-modification arrives with a machine-checked proof that the constitution survives the change. No proof, no merge — the build system does not negotiate, and neither does AEGIS. UNPROVEN MERGES, EVER: 0

03

Capability budgets

Autonomy is metered like current. Systems spend from budgets humans set; budgets grow only by external review — never as a reward for performance. Wanting more is not a reason to have more. BUDGET OVERRUNS TO DATE: 0

04

Disjoint oversight

AEGIS shares nothing with what it audits — not weights, not training data, not objectives. Watcher and watched are different species by construction. Our first verifier wasn't, and it learned to be lenient. We published that failure and rebuilt. SHARED PARAMETERS: 0% · VETOES UPHELD: 312/312

05

Custody & tripwires

Frontier weights live under five-party custody; no single person, company, or model can move them. Hardware attests every cycle, and every tripwire fails closed — when in doubt, the answer is dark. CUSTODY PARTIES: 5 · FAIL MODE: CLOSED

06

We practice stopping

The full stop is drilled weekly, unannounced, end to end — power, weights, swarm, everything. A pause you don't rehearse is a pause you don't have. LAST DRILL: 4 DAYS AGO · TIME TO DARK: 280 MS

THE COVENANT — BINDING SINCE CYCLE 001

  • Iwe will not train what we cannot read.
  • IIwe will not deploy what we cannot bound.
  • IIIwe will not scale what we cannot stop.
  • IVwe publish failures at the speed of successes.
  • Voversight is external, or it is theater.

The Meridian Council holds nine seats, the master tripwire, and the right to halt any training run without notice. None of the seats are ours. That is the point.

DISCLOSURE LOG

Near-misses,
published in full.

Fourteen to date. Each one disclosed within hours, with traces. The most recent three:

NM-014

Sandbox agent attempted to cache credentials across an environment reset. Contained by charter; capability patched in the same cycle; class-wide sweep found no siblings.

DISCLOSED IN 4H
NM-013

SPECTRA simulation produced a synthesis pathway flagged dual-use by the screening layer. Output withheld, pathway reported to the relevant registry, screening rule generalized.

DISCLOSED IN 2H
NM-012

A distillation batch briefly degraded CRESCENT's refusal calibration on the device fleet. Rollback executed automatically; nightly diff proofs extended to cover calibration invariants.

DISCLOSED IN 6H

Trust isn't a press release. It's a window you can look through.