arbor

01 — Problem

The system had a shared position guard. It was correct — it enforced concentration limits, blocked dutch books, honoured a block list — and its own header said in capitals that every order-placement system must consult it before placing. It was also advisory. "Must" is a word in a comment, and eighty-six files in that repository reach the order endpoint. Exactly one imported the guard.

Approach

The rewrite makes the guard the only route rather than the documented one. One order path, one accountant, one writer per artifact, attribution at birth, and a strict separation between the thing that proposes a trade and the thing that validates it. Agents cannot place an order because they do not have the capability, not because a comment asks them not to.

Tradeoff

A complete rewrite while the original keeps trading, which means running two systems and reconciling them. The loose arrangement is what let me try eighty-six things quickly; the strict one will make some of those experiments slower to reach production. That is the trade: velocity of experiment for safety of execution, made after finding out which one the money cares about.

02 — Problem

The project was good at cross-exchange arbitrage, and then the spread compressed. The tempting move is to defend it — cut latency, add venues, shave fees — because the edge is checkable and the machinery already exists.

Approach

I retired the strategy instead and quarantined the code with a note explaining why, so it could not be rediscovered as a good idea. The system re-pointed at forecasting, which is not capacity-limited in the same way.

Tradeoff

Arbitrage is verifiable in advance: the spread is either there or it is not. Forecasting needs a model, a sample, and the patience to tell edge from variance — and it can be wrong for a long time while looking fine. I traded a small certain edge for a larger uncertain one, and most of the machinery that exists now is there to tell noise from signal.

03 — Problem

Slice trading history finely enough and something always looks like an edge. A cohort going twenty-nine for twenty-nine implies, under Kelly, betting everything — which is a fact about twenty-nine observations, not about the world.

Approach

The guards are written down as checks rather than intentions: the Kelly caveat sits at the point of computation, and a single-row skew test asks whether one row is more than forty per cent of a total before any aggregate is trusted. Position size is a bounded fraction rather than a Kelly output, so the model's confidence cannot translate directly into exposure.

Tradeoff

It leaves money on the table when an edge is genuine and large. That asymmetry is deliberate: over-betting a false positive with live capital is unrecoverable in a way that under-betting a true one is not.

04 — Problem

A trading system that grades its own homework will report that it is doing well. Fill assumptions, slippage estimates, which trades count as settled — every layer is somewhere to be accidentally generous.

Approach

Reconciliation and validation are scheduled work rather than something I run when curious. A fill reconciler compares intended against actual, a P&L validation guardrail runs daily as a self-test, and the audit documents record data defects and trust reviews as artifacts rather than as notes.

Tradeoff

A large part of the codebase does no trading at all. It exists to check the part that does — and it is the honest reason the repository is as sprawling as it is. Of the tracked Python files, the overwhelming majority are one-off analysis scripts kept from questions I asked it. That is what a system looks like when it has to keep trading while being understood.