Skip to main content
system: operating/L0: 30/L1: 3/L2: 2/L3: 0
// The 48-hour shadow ladder

No capability ships without 48 clean hours.

Every agent climbs L0 → L1 → L2 → L3 through a mechanical shadow window. Any miss — any red-line touch, any unbudgeted spend, any failed eval — resets the clock. This is not a process. It is a contract between you, your compliance team, and the system.

// Tour the tiers

Click any tier to see the gates.

Each tier has six gates that must pass before promotion. Click L0/L1/L2/L3 on the left and the gate panel on the right swaps. Promotion criteria are mechanical, not vibes.

Tier L0 · read-only

Reads connector data: yes
Writes to any system: never
Audit row per read: required
Daily budget cap: enforced
Red-line check: every call
48h clean shadow before L1 promotion

Promotion criteria: zero red-line touches, eval pass-rate above threshold, operator approval, downstream-system health, behavioral consistency, cost discipline.

01
The clock is mechanical.Every promotion requires 48 wall-clock-hours of clean shadow firing. The promotion script reads the audit log and refuses if any window is short. No "approximately."
02
Six gates per tier bump.Behavioral consistency · cost discipline · zero red-line touches · eval pass-rate above threshold · operator approval · downstream-system health. All six green before the agent climbs.
03
Any miss resets the clock.One unbudgeted spend, one red-line touch, one classifier miss in the eval window, and the agent goes back to the bottom of its current tier. No grandfathering. No exceptions.
04
Red lines never move.Payments, hard-deletes, unapproved web submits, security-permission changes — Guardian's capabilities_forbidden_always list is immutable across all tiers. Read the red lines →

Field note (June 2026): Anthropic publicly urged a coordinated pause option as frontier AI nears recursive self-improvement, naming loss of human oversight as the danger line. That is why this ladder exists: autonomy is earned through mechanical shadow windows and gates, never assumed — and no agent self-modifies, ever. (Anthropic, 2026-06-04; reported by Al Jazeera and others.)

The ladder runs in production now.

Of 35 agents: 30 in shadow (L0), 3 proposing (L1), 2 acting-with-review (L2), and 0 at full autonomy (L3) — nothing in the company acts without a human above it in the loop. Two L0→L1 promotions landed this week, both with clean shadow logs you can read.