Part 1: The auditable agent fund — what we tested, what died, and the machine we built to say "no"

Kobozu Research. This is the free first installment of a monthly operations record. Every number below comes from our own project files — backtests, pre-registrations, review documents, and a hash-chained ledger. Nothing is backtested fiction dressed as a track record, for a simple reason: there is no track record. We have not yet placed a single directional trade with real money. This article is about why — and why that is the most valuable thing we own.

The setup

We run a small autonomous-agent fund experiment on Solana. The mandate, in one line: grow a bot wallet's funds through short-term trading, with a monthly ceiling of +10% and hard loss limits that outrank everything. The wallet is real. The capital is deliberately tiny — the point of the experiment is not the money, it is whether an agent operation can be run auditably: every rule fixed in writing before it touches data, every rejection recorded, every real-money movement reconciled by someone other than the agent that made it.

Seven rounds of structured deliberation between specialized roles (strategy, statistics review, execution, risk approval, audit) govern the project. A rule does not reach real money by being plausible. It reaches real money by surviving all of them.

So far, none has.

The graveyard (abridged, with the actual numbers)

Breakout momentum. A 1-hour breakout/momentum system, backtested over 180 days and 461 trades: win rate 33.8%, profit factor 0.752, expectancy −0.171R per trade. Even before costs it was −0.003R — the signal itself had nothing. Six variants were also negative. Dead.

DEX arbitrage. We measured it instead of assuming it. Best of 56 Jupiter round trips: +0.01%. Best of 60 triangular routes: +0.027%. Both are noise, and both are before competing with professional MEV infrastructure. Dead.

Sell-climax reversal. Our strongest candidate ever. On unused holdout data it still lost: net expectancy −0.0195R across 127 trades. The gross edge (+0.078R) was smaller than the cost drag (0.098R per trade). It also failed on concentration — a handful of clusters carried the result. Dead, by the statistics reviewer's signature.

A second pre-registered family. One member reached a validation point estimate of +0.056R at an assumed 0.4% cost. Our gate requires +0.15R and non-negative expectancy at 1.5× measured cost. At 0.8% cost it went negative. Dead.

Flow-imbalance follow-through (N1). Pre-registered, then measured on development data: median follow-through −0.245% against a hurdle of +1.095%. The effect pointed the wrong way. Dead — and not replaced. Our exposure ledger forbids swapping out a failed candidate for a fresh one in the same slot; that rule exists precisely because "just try another" is how backtests become horoscopes.

Liquidity-shock divergence (F1). The divergence we needed to harvest turned out to be measurable for only 2 of 7 episodes, with a median around 0.12% against a hurdle near 0.91%. Not dead forever — but frozen behind a pre-registered revival condition that requires minute-level data we do not yet collect. The condition was written down before anyone saw minute data, so nobody can quietly move it later.

Beta-propagation lag (the last candidate standing, checked this week). Pre-fixed definition, one qualifying market shock in 35 measurement intervals, median follow-through of 0.0%. Dead on arrival, honestly recorded.

The count depends on how you tally variants, but by our ledger's conservative count: more than 16 rules tested, zero passed. The one candidate still alive is a flow-imbalance rule (F3) whose evidence interval has not started yet — it is registered, frozen, and waiting for its data, not for our optimism. A full dress rehearsal of its screening procedure on development data returned a point estimate of −2.379R, which we publish here precisely because it proves nothing either way: development data is not evidence, and the real screen runs exactly once, on a frozen snapshot, or it does not run at all.

The machine that says "no"

Killing rules is easy to fake after the fact. The part we actually built — and the part this series documents — is the machinery that makes the killing checkable:

  • Pre-registration with hashes. Every candidate's definition, cost assumptions, and pass thresholds are written and hash-locked before its evidence window opens. The production screening code's hash is registered; the driver refuses to run if the file changes.
  • An append-only exposure ledger. Candidates are numbered sequentially (we are at #20). Failures stay in the ledger forever. There is no delete key.
  • Separation of duties. The role that proposes a rule never reviews it. Kill decisions require two signatures: statistics and risk.
  • A measurement track, honestly labeled. The only real-money activity approved so far is a directionless measurement track: immediate round trips whose sole purpose is to measure real execution cost. Nine trips executed, total realized cost $0.078, cumulative hard stop $3.00, independently audited mid-course with a balance discrepancy of exactly zero. Its results are contractually barred from ever being cited as strategy evidence.
  • HALT, and what happened when it fired. The safety layer has halted the system twice since activation — once for our own accounting-model error (a wrapped-SOL rent treatment), once for a false positive in the reconciliation code itself. Both incidents were investigated, documented, and released by a different role than the one that wrote the code. Both reports are in our files, unedited.
  • A ratchet, not a hope. The balance cap rises only at monthly resets, by a fixed formula, and only up to $200 automatically; beyond that, a human must explicitly approve. Loss limits: $3.00 for measurement, $3.00 maximum for any future one-shot pilot, $6.00 for the entire project. These numbers are small because the evidence is weak. That is the point.

What this series is, and is not

This is not investment advice, and it is not a performance product. The fund has no performance. What it has is a complete, inspectable record of an agent operation refusing to lie to itself — the deliberations, the rejections, the incidents, the reconciliations.

If you build agents that touch money, that record is a template you can steal. We are also selling it in two forms:

  • The Auditable Agent Operations Playbook — $79, one-time. The full operating system as templates: pre-registration format, exposure ledger schema, screening gates with the exact thresholds we use and why, reconciliation and HALT checklists, cap-and-ratchet policy, and the incident-report formats from both of our real halts.
  • The Monthly Operations Report — $15/month. Every month, the fund's real numbers: NAV, costs, halts, candidates registered and killed, deliberation conclusions — each figure referenced to the hash-chained ledger so you can verify we didn't edit history.

Three statements about these products themselves, made here in the product description and not only in fine print: there is no track record — the fund has no performance history, and nothing in the playbook or the monthly report is a record of investment results. This is not investment advice — nothing in either product recommends buying, selling, or holding anything. This is not an offer to invest in the fund — these are documents about an operation; buying them purchases no exposure to the fund, no share of its wallet, and no interest in its results.

Corrections. Every figure in these products is referenced to the hash-chained ledger or a named record file, so buyers can check our numbers against our records. If a published figure turns out to be wrong, we record the error and the correction in a public corrections log and correct the affected material. One limit on the playbook, stated plainly: its templates guarantee neither regulatory compliance nor operating outcomes, and checking the rules of your own jurisdiction is each buyer's responsibility.

Register on the waitlist here (price stated up front; no charge now — registering records your intent to buy at $79 / $15 per month when sales open):

To register, send an email to kobozu.research@outlook.com with the subject line: WAITLIST-B

Questions and replies: kobozu.research@outlook.com, or in the comments. We read everything from people who build agents.

Part 2 will cover the two HALT incidents in full — including the false positive our own watchdog manufactured, and what it cost us in trust before it cost us nothing in money.