EVIDENCE OPERATING SYSTEM FOR AUTONOMOUS MACHINES

Every autonomous decision
should leave evidence.

Deterministic incident reconstruction for mission-critical autonomous systems.

EXPLORE PAMIR →
REAL MISSIONS.
REAL DATA.
REPRODUCIBLE EVIDENCE.
WHY PAMIR

When autonomy fails, engineering teams need more than a score.

They need a defensible reconstruction of what happened — and why the available evidence supports that conclusion.

PAMIR works after the mission. It turns telemetry into inspectable evidence without controlling the vehicle, targeting or engagement.

DETERMINISTICSame input.
Same result.
INSPECTABLEEvery material step remains traceable.
EVIDENCE-GRADEBuilt for engineering, validation and diligence.
AUTONOMOUS DOMAINS
AIR

Unmanned & autonomous flight

From incident to inspectable evidence.

LAND

Field autonomy

Operational reality, reconstructable.

SEA

Maritime autonomy

Complex missions, clear evidence.

CRITICAL SYSTEMS

Defence & infrastructure

Evidence that can withstand scrutiny.

PAMIR EVIDENCE ENGINE

From telemetry
to evidence.

Not a black-box confidence score. PAMIR preserves the path from source telemetry to the material claim — including uncertainty, counterexamples and reproducible replay.

INCIDENT / PUBLIC HOLDOUT 001MODE / POST-MISSION
FIRST DEVIATION / t+468.303565s
NOMINAL FLIGHTsource telemetry
FIRST DEVIATIONestimator output tracking
CHAIN EVENTtemporal evidence
RESULTdeviation detected
FIELD VALIDATION CORPUS V1 · COMPLETE

Evidence over assumptions.

CORPUS / V1

50 real public PX4 flights

50/50 accepted cases achieved exact deterministic replay with SHA256-pinned provenance. No synthetic data.

COMPLETE
COMPOSITION

Independent source classification

6 incident · 33 healthy control · 11 unknown. Labels are retained independently of PAMIR output.

50 FLIGHTS
BASELINE / 0.1

Frozen and reproducible

The validated v0.1 detection policy, thresholds, benchmark corpus and Evidence Pack remain unchanged.

FROZEN
CALIBRATION

Field findings retained, not tuned away

16 of 33 source-labelled healthy controls produced a PAMIR root. This is tracked as a calibration and generalisation target—not as causal ground truth.

48.48%
BOUNDARY

Reproducibility is not accuracy

Exact replay proves deterministic reconstruction for the pinned input. PAMIR confidence remains an anomaly-strength heuristic, not a probability of causation.

EXPLICIT
INSTITUTIONAL / PILOT

Bring us a real autonomy problem.