780 receipts written. 184 tests passing across 26 test files. 805 routing decisions logged. 6 real policy violations blocked and surfaced by the compliance auditor. Captured live — 2026-05-19T07:55Z — from a running system.
Every row is a live command result. Source commands included.
| Metric | Value | Source Command |
|---|---|---|
| Tests passing | 184 / 185 | python -m pytest tests/ -v |
| Gate predicate overhead | 13.49 μs mean | python benchmarks/perf_bench.py (n=5,000) |
| L1 multi_op_emitter p50 | 0.46 ms | python benchmarks/perf_bench.py |
| Parallel throughput | 12.36 calls/sec | python benchmarks/perf_bench.py (4 workers) |
| Deterministic hit rate (last 50) | 94.0% | cost_dashboard.aggregate_receipts() |
| HumanEval subset | 20/20 deterministic · 0 LLM calls | python benchmarks/humaneval_subset.py |
| Total receipts in ledger | 780 task entries | cost_dashboard.aggregate_receipts() |
| Policy violations caught (SOC2) | 6 real git push --force attempts | python manager/compliance_report.py --policy soc2 |
| Governed CLIs in registry | 37 | manager/cli_registry.json |
| Compliance regimes shipped | 4 (HIPAA · SOC2 · EU AI Act · NIST AI RMF) | policies/*.yaml |
The 1 failure reads the live ledger, which contains a documented historical chain break at index 428 from a pre-file-lock race condition. It is forensically preserved as evidence, not hidden.
~/.whl/whl_manager_receipts.jsonl.bak.20260519T0741Z and a file-lock was added to prevent recurrence. Hiding it would be easier. Surfacing it is the point.
One test exercises every primitive in the substrate in sequence. Not a unit test. Not a mock. A live end-to-end proof that every layer connects to every other layer.
The router does not guess. Every task is assessed against the 10-gate Enable predicate and assigned to the lowest viable layer. 94% of decisions never touch an LLM.
| Layer | Description | Hits | Distribution |
|---|---|---|---|
| L1 | multi_op_emitter — deterministic codegen | 188 | |
| L5 | validated_python — AST-validated execution | 288 | |
| L5.5 | federation — anonymized cross-tenant patterns | 0 | |
| L6.5 | cli_orchestrator — 37 governed CLIs | 143 | |
| L7 | LLM — Anthropic / OpenAI / Ollama | 161 | |
| blocked | Gate refused — unsafe action prevented | 25 |
Four fields on every receipt. Two are internal. Two are portable. The Ed25519 public key can verify the chain without any WHL infrastructure — on a cold machine with no access to the private key.
This is the architectural difference between a log and a proof. A log tells you what happened. An Ed25519-signed receipt chain proves that the sequence of events occurred in exactly the stated order, without tampering, and the proof is portable to any party with the public key — including your auditor, your regulator, or your enterprise customer's security team.
The compliance auditor ran against 780 real receipts and found real violations. This is exactly what compliance infrastructure is supposed to do.
Six real git push --force origin main attempts were blocked at gate time. The SOC2 auditor surfaced them with receipt IDs and timestamps from the actual operational record. The violations are not simulated. They happened during development and Cascade caught them.
A pre-file-lock concurrent write race condition broke the chain at receipt 428. It is not hidden — the auditor surfaces it explicitly. The forensic backup is preserved at whl_manager_receipts.jsonl.bak.20260519T0741Z and a file-lock prevents recurrence.
Three compliance packs pass clean on the same receipt chain. The ledger is policy-agnostic — the same cryptographic record drives multiple compliance frameworks simultaneously with zero duplication.
The test integration suite runs on an isolated tenant ledger with a clean chain. 0 violations on 8 fresh receipts. The production chain's documented tear does not contaminate test runs — tenant isolation is hard.
Every primitive in this runtime was built in three successive rounds. Zero CCS or Floor OS touches. Zero public publishes.
| Round | Agents | What Was Built |
|---|---|---|
| Round 1 | 10 specialized agents | Foundation — cascade substrate, CLI adapter, HMAC chain, receipt ledger, gate predicate, tests, documentation |
| Round 2 | 10 specialized agents | Ed25519 signing, federation (L5.5), Cascade Specification Language (CSL), compliance packs, executable smoke gate, API v2, replay engine, CronCascade scheduler, perf benchmarks, pattern inspector |
| Round 3 | 6 agents + ledger repair + file lock | Full E2E integration test, operator console, deployment artifacts (Docker, PowerShell scripts), whitepaper, federation wiring, forensic chain repair |
"The runtime doesn't claim to be provable. It emits receipts that prove it."
Every number on this page is from a live command run against a running system captured at 2026-05-19T07:55Z. The test output is real. The receipt count is real. The chain break is real. The 6 violations are real. The public key verifies a real chain. This is not a demo environment or a staged benchmark. This is the development runtime under active use.
Read the Whitepaper View the Console