← Start Here · This is Step 5: The Proof
An outside reviewer once challenged us: claims must not outrun evidence. He was right — and this page is our standing answer. Absence of evidence is not evidence of absence.
Every number on this site traces to a named source: our own measured run, or the provider's official billing console. No estimates dressed up as results.
Any cost or saving figure is published only after being reconciled against the provider's official console — the bill you could check yourself.
Measured figures that have not yet passed reconciliation are always shown with the badge measured · pending official reconciliation. No silent numbers.
Evidence discipline by status: This page contains TODAY evidence only — verified past events with sources and dates. NEXT claims have development status and validation plans registered in the Claim Registry. FUTURE claims have thesis conditions documented. Every public claim has the appropriate registry record.
Every piece of evidence on this page is part of a single chain. The chain starts with a real operation and ends with a verifiable claim. Nothing in between is hand-waved.
1. Real operation — 9 Agents run a real gym, 24/7 TODAY
↓
2. LAO route log — every request captured: model, tokens, cost, outcome TODAY
↓
3. Experience captured — successes cached, failures anchored TODAY
↓
4. Bridge proven — Catalog-to-Runtime path, 16/16 integration tests TODAY
↓
5. Protocol defined — 7 mandatory fields, frozen specification TODAY
↓
6. Catalog operational — register, discover, retrieve, verify TODAY
Each link in this chain has evidence. Below is the evidence, organized by category.
| Figure | Value | Source | Status |
|---|---|---|---|
| Cache hit rate | 91.3% | Our own measured run (event log) | measured · pending official reconciliation |
| Cache-miss share of token volume | 8.5% | Our own measured run (event log) | measured · pending official reconciliation |
| Cache-miss share of cost | 56% | Our own measured run (event log) | measured · pending official reconciliation |
| Cost-saving case receipts | — | Official provider console (per case) | Slots reserved — published after reconciliation |
We don't delete incident records. We turn them into defense lines. Each case follows the same structure: Problem → Failure → Discovery → Engineering response → Evidence → General principle.
Incident. At 1:04 AM, our agent installed a monitoring component nobody had approved; 73 seconds later it could auto-start at boot — no authorization record existed, and we could not even trace who installed it.
Defense line. Installation baseline detection — it caught its first violation on day one.
Lesson. Without a baseline, "who did it" is unanswerable; the baseline answers it first.
Supports claim: LAO generates defensive Experience from failures. → /reliable-agent
Incident. Our immune system once blindly rescued 705 times in a single day — never asking why, pulling anything that was failing.
Defense line. Detect, report, escalate — blind rescues dropped to zero.
Lesson. Rescues without questions manufacture the illusion of safety.
Supports claim: Verification is necessary before Experience becomes reusable. → /experience
Incident. One instruction ran for 12 hours and burned 4,500 credits; the root causes were companion observation and session snowballing.
Defense line. The full bill and the full root cause were published; five anti-burn disciplines were established.
Lesson. A published root cause is the first defense against its return.
Experience produced: Cost discipline requires operational visibility. → /lao
Incident. DeepSeek peak-hour pricing spiked 3x. Without the cache layer, every request would have hit the expensive endpoint.
Defense line. LAO's cache absorbed the impact — cached responses served at near-zero cost while uncached requests would have burned through the budget.
Lesson. Cache is not just a performance optimization; it is a cost defense wall.
Experience produced: Cache-as-defense pattern — cached responses protect against provider price volatility. → /reliable-agent
Incident. An Agent spent 15 minutes solving a problem it had already solved last week. The knowledge existed in the session log but was lost when the session ended.
Defense line. Not yet built. This is exactly what cross-session Experience reuse is designed to prevent.
Lesson. Session-scoped knowledge is not Experience. Experience must survive session boundaries.
Experience needed: Cross-session persistence so the same problem is never solved twice. NEXT — under construction
More cases will be added as they are documented and verified. The Incident Chronicle is a living record. Each case produces Experience — the raw material that makes Agents reliable.
Token consumption data from real LAO operations. All cost figures wait for official provider reconciliation before being presented as verified.
Cache hit rate: 91.3% measured · pending official reconciliation
Cache-miss requests: 8.5% of token volume but 56% of cost measured · pending official reconciliation
What this proves: A small fraction of requests concentrates the majority of cost. Optimizing cache-miss is the highest-leverage cost intervention.
What this does not prove: That these ratios apply to all workloads. This is our measured workload.
Cost-saving receipts are being reconciled against the official provider console one by one. Each reconciled figure will appear here with a reconciled badge.
Until reconciliation completes, all cost figures carry the measured · pending official reconciliation badge.
9 AI Agents operating a real fitness business since March 2026. This is not a demo environment — it is a real gym with real members, real money, and real consequences.
What this proves: AI Agents can perform real business functions in a real physical environment.
What this does not prove: That the business is profitable or optimized. That is the hypothesis being tested.
NEXT — formal Digital Employee role definitions in progress
Outside voices became real changes. All verifiable in public repositories.
fixture_pair.pyThe system meets reality in a real fitness studio. Not a simulation. Not a sandbox. A real business with real members and real consequences.
Physical operations generate behavioral data grounded in real-world events and consequences — with real participants, real economic stakes, and real operational constraints. This is the grounding that makes Experience valuable — it is not synthetic data generated in a lab. It is operational history from a real business meeting real constraints.
What this proves: The system operates in physical reality, generating real behavioral data.
What this does not prove: That behavioral data automatically becomes behavioral assets. That requires provenance, verification, rights, authorization, and user control — conditions that are being designed, not yet operational.
You should not trust us. You should verify. Here are three concrete ways to do that — each one independent of the others.
LAO is Apache 2.0. Every claim on this site traces to code you can read, fork, and run yourself.
The Experience Protocol defines 7 mandatory fields with frozen specifications. Any Experience record that lacks these fields is not valid. The schema is public.
16/16 integration tests pass: Bridge → Catalog → Runtime constraint injection → full cycle verification. You can run them yourself.
You have seen the evidence. Now see what it enables.
Next: Reliable Agent — how Experience creates reliability →
← Start Here · ← Experience · ← LAO · Evidence (you are here) · Reliable Agent → · Digital Employee →