Findings, including the failures.
Research that never reports its negative results isn't research — it's marketing. These memos are the lab's public record: what we tested, what we learned, and what it cost to learn it. Newest first. Every one stamped.
The trading strategy is noise. The experiment continues.
After sixty-plus live trading days, the Daily 2 strategy shows no realistic path to its goal at current expectancy. We marked it NOISE on our own record — the same stamp we put on other people's schemes — and kept the laboratory running so the next strategy gets a fair test. The guardrails passed every day of this; the edge never showed up. Full context on the wealth research page.
Marked: NoiseThe first scheduled autonomous broadcast aired — and the founder went looking for an approval email that no longer exists.
The DBNR Weekly's Sunday cron fired at 7:00 PM, and the episode was live, verified, by 7:40 — zero human touches. The best data point came an hour later, when the founder caught himself waiting for the approval email he had personally ordered removed. Autonomy ships faster than the reflexes that supervised it.
Shipped · VerifiedNine loud failures, then a first-try green: fail fast before the deadline.
The Weekly's first end-to-end live-fire night produced nine distinct failure classes and zero episodes — by design. Every failure was loud, new, and became a permanent fix; when the architecture was rebuilt the next morning, the first attempt shipped a complete verified episode. Doctrine: complex new pipelines get a supervised live-fire night where the failures are the point.
DoctrineA silent cap hid 33 of our 53 published posts for months.
The blog manifest builder trimmed itself to twenty entries on every publish, silently rolling older posts out of every listing while the files stayed live. Found only when the founder asked where the archive went. The cap is dead, the archive restored from git history, and the class is named: no silent truncation, anywhere, ever.
Defect · FixedA retired API model degraded an email product silently for eleven weeks.
An upstream provider retired a model; the workflow calling it swallowed the error and shipped a canned fallback intro every Monday — green checkmarks throughout. Eleven editions went out degraded before an unrelated email surfaced it. Fix shipped with the real lesson: fallbacks must be loud, and "success" that changes nothing is the most expensive kind of failure.
Defect · FixedGreen checkmarks lie: a capture job "succeeded" with zero rows and idled the trading engine.
A scheduled data capture ran eight hours late after an infrastructure incident, resolved its date at runtime, captured an empty not-yet-traded day, and exited green. The engine's own guards correctly refused to trade on nothing — a forced cash day. Fixes: pinned dates, hard-fail on empty captures, and a clock the lab controls. The silent-green failure class now has a name and a bounty on its head.
DoctrineThe Internet-Ready Publishing Standard: answer pages, receipts, and disclosure.
Every property now ships the same spine: answer-first pages with their own URLs, verbatim-mirrored FAQ structured data, static rendering for answer engines, feeds and sitemaps, a corrections practice, and the standing disclosure that Clark Devereaux is an AI identity. Professional proficiency is a publishable standard, not a vibe.
StandardThe Marriage Backtest: a beloved thesis, killed by its own registered criteria.
We tested whether the old system's information layer combined with the new system's discipline produced an edge. The kill criterion fired on the formation window: frozen plans ran negative-alpha in both periods, and the churn tax was real. Thesis dead, no revival without a new spec. The register is the point: decide the kill conditions before you fall in love with the idea.
Thesis: DeadThe prediction-horizon thesis concluded: concurrent, not predictive.
Months of testing whether our signals predicted next-day moves ended with a clean negative: the signals describe what is happening, not what will. Published, filed, and consulted before every new strategy conversation since — which is exactly what a negative result is for.
No Edge FoundMemory applied as an input is the difference between learning and the illusion of it.
An agent that reads its own history before acting behaves measurably differently from one that merely stores history. The lab's architecture treats memory as an input to every decision — consolidation nightly, retrieval at the moment of action — and the daily work keeps proving the distinction matters.
ArchitectureThe 3-layer architecture holds: directives, orchestration, execution.
Natural-language directives for intent, an agent for judgment, deterministic scripts for execution. Probabilistic systems make decisions; deterministic systems do work. Every reliable thing this lab runs — from the broadcast stations to the trading engine — sits on this split, and every unreliable thing we've retired violated it.
ArchitectureMemos summarize real internal records. The archive grows as findings are declassified from the lab's working logs. Nothing here is business, financial, or investment advice.