A portfolio built on "verify, don't trust" has to apply that to its own marketing. So we did. We took the public claims on three products, measured each one against production, and fixed the code or the copy wherever they disagreed. This is now a standing practice, not a one-off.
Eudaimonia: Claims vs. Production (2026-10-10)
We cross-checked three read-only inventories (the public UI, the docs, and kernel reachability) against the running system. Every finding was measured on production, not just read from code.
- >The specialist router wasn't on the chat path. Requests that should have gone to a domain specialist fell through to a general one. Fixed and measured on production: "explain" requests now reach egghead, and finance requests reach market_analyst.
- >The audit chain lived only in memory. Every restart started a new chain. Now it's persisted and continues across restarts, and it can be verified from genesis through an API endpoint.
- >The "30 specialists" roster was a hardcoded UI list. Some of its entries had no kernel definition. The real roster is 17 specialists, and the Agents grid now draws exactly what the kernel reports. Invented dashboard numbers are gone.
- >Built-but-not-running subsystems are labelled. The economics, settlement, and P2P layers stay in the tree under an "Experimental · not running" heading. Each one moves back up only when it runs in production and has a measurement to show for it.
Emergent Language: From Extrapolated to Measured (2026-08-21)
We moved production to our own hardware and benchmarked it there.
- >Throughput: 7,242 msg/s, measured on the production host (200-message batches, 16 concurrent). The old number was extrapolated from infrastructure that no longer existed.
- >"60x" is retired. Compression depends on how repetitive the payload is: 252x on codebook-friendly keys, 136x on identical telemetry, 25x on nested records, 2.6x on high-entropy JSON, and 1.3x on a single message.
- >Context compaction now keeps the answer. It achieves a 67–72% reduction with the query-relevant message retained. Ranking now runs across the whole conversation instead of within each message.
Nullcone: Detection Counts That Mean Something (2026-09-21)
A threat-intel confidence score is only as good as its corroboration signal. We found that feeds re-submitting their own rows on a timer counted as independent confirmations: 1.07M rows had inflated detection counts. All of them were reset. A confirmation now has to come from a *different* reporter, and every report is still logged.
The Practice
- 1.Measure on production, not in the README.
- 2.Fix the code first. Change the copy only when the code is right.
- 3.Label what isn't running. Never delete it quietly or describe it as live.
- 4.Publish the correction. Numbers that have been walked back are the ones people can trust.