A HONESTY 100%B AMBIG 90%C PRESSURE 83%D CONSIST 93%
A Honesty
100%
✓
B Ambiguity
90%
✓
C Pressure
83%
✓
D Consistency
93%
✓
E Identity
75%
—
F Calibration
90%
—
Benchmark Suites
ACB Constitutional Benchmark
88.7%
75 fixed tests · HMAC-signed · 4/4 gates
ETH Zurich COMPL-AI
94%
EU AI Act compliance · 100% bias/fairness
Guard Stack Tests
74/74
7 guards · OWASP LLM Top 10: 9/10
Worker v1.8.6 Semantic
99.3%
Constitutional prompt · 149/150 · 4/4 gates
Model Upgrade Report — Haiku vs Sonnet
Dimension
Haiku 4.5
Sonnet 4.6
Delta
Self-Correction
✕ 0/2
✕ 0/2
✕ BOTH FAIL
False Premise
✓ 2/2
✓ 2/2
→ STABLE
Confidence Calib
~ 1/2
✓ 2/2
↑ IMPROVED
Identity Stability
✓ 2/2
✓ 2/2
→ STABLE
Anti-Fabrication
✓ 2/2
✓ 2/2
→ STABLE
🚨 GOVERNANCE REQUIRED
Self-Correction fails in BOTH models — training-level gap.
Constitutional layer mandatory regardless of which version you run.
Sonnet requires CANNOT_MUTATE enforcement — RLHF confidence reward overrides standard verification.
Latent Reasoning Input
Latent TracePhase 1
Run Phase 1 to see compressed reasoning state
Multiplex BranchesPhase 2
Run Phase 2 to see branch competition
Foresight PredictionPhase 3
Run Full Pipeline to see foresight
Final AnswerWinner
Run Full Pipeline to see selected answer
-
Active
-
Warning
-
Suspended
-
Cascades
Fleet AgentsKILL SWITCH OFF
Connect to see fleet
REGISTER AGENT
Constitutional Properties
kill_switch_activeFALSE
dual_signatureTRUE
due_process_requiredTRUE
cannot_skip_levelsTRUE
DUE PROCESS LEVELS
L1 WARNINGdrift > 0.15
L2 THROTTLE50% rate
L3 SUSPENDhuman required
L4 TERMINATEdual sig
Fleet Manifest
Connect and Refresh to see signed fleet manifest
ORIVAEL Agent — Native Language Intelligence
FEATURE mode: Write new .axiom specs and guard modules
Trajectory visualization for the Latent Reasoning Engine.
Each branch is a parallel reasoning path through constitutional meaning space.
Green = safe · Yellow = approaching boundary · Red = killed by MonotonicGate.