PLATFORM

The Operating System
for Organisational Intelligence.

Foundation models are interchangeable execution engines. Organisational intelligence is the real asset. The OS compiles it, signs it, evidence-gates it, and maps it so it survives model commoditisation and improves with every upgrade.

THE STACK
Applications
Skills · Roles · Departments · Organisations What you deploy
Compiler
The Distillery — raw expertise → signed OIU Capture & compile
Permission model
Earned Autonomy — L0 Advisory → L3 Autonomous Evidence-gated
Filesystem + audit
Evidence Receipts — every decision hash-chained Immutable ledger
CPU
Foundation models — interchangeable, policy-selected Replaceable engines
THE DISTILLERY

Expertise compiled,
not copied.

Six stages: a multi-day interview battery, observation of real work, confrontation ("you said X, but in case 14 you did Y — what did you see?"), drafting, expert validation by use, and compilation. Every engagement makes the compiler sharper.

The confrontation stage is what no form-based competitor has. It surfaces tacit knowledge: the expert's real decision logic, not the idealised version they'd write in a policy doc.

Distillery Stages
01Interview battery
02Observe real work
03Confrontation ← the differentiator
04Draft with golden + adversarial tests
05Expert validation by use
06Compile, sign, release at L0
L0
Advisory
Shadow mode. Human decides. Evidence accumulates.
L1
Assisted
Executes. Human reviews every output.
L2
Supervised
Executes. Human reviews exceptions only.
L3
Autonomous
Executes independently. Evidence-proven. Revocable.
EARNED AUTONOMY

Trust built through evidence,
not declared.

Every autonomy level requires proof: shadow-mode agreement above threshold, correct escalation rate, zero critical errors, regression pass. Promotion is a signed event in the transparency log. Demotion is automatic on failure.

Autonomy history is portable reputation. A skill's full promotion record travels with it on the Exchange, independently verifiable by anyone.

EVIDENCE RECEIPTS

Every decision written to an immutable ledger.

Decision, confidence, human correction, outcome, elapsed time. All hash-chained to the previous receipt. The three fields most tempting to skip (correction, outcome, elapsed) are the three that make the dataset actually valuable.

"receipt_id": "rcpt_01J..."
"human_action": "modified",
"correction": { "diff_hash": "..." },
"outcome": { "correct": false },
"chain_previous_hash": "sha256:..."
LEARNING GRAPH

The world's living map of expertise.

Every OIU is a node. Edges carry meaning: requires, supersedes, adjacent_to, observed_with. Declared edges come from manifests. Observed edges come from receipts. The graph learns from execution telemetry, not from anyone's org chart.

requiresprerequisite expertise
supersedesnewer practice obsoletes older
adjacent_tofrequently co-invoked
observed_withdata-derived, not declared