The honest measured result: no shipped rule is net-positive at Coinbase's taker rate actually paid

[eaitbrahim]

Stated by the project first, so nobody has to find it themselves: no shipped rule family is net-positive at the taker fee actually paid on Coinbase (~1.2% per leg — the only live venue; a Robinhood adapter exists but is deliberately unwired, issue #198). The viable parameter/fee intersection is empty under production-faithful execution across the Coinbase-listed universe — 0 of 90 configurations in one matrix, 0 of 82 in another — and cost is the binding constraint, not signal quality. As of v0.9.0 the backtester also prices per-product slippage scaled from each asset's real liquidity (5–50bp), so this result can no longer be flattered by thin books.

The full record, with every number and the engine defects that forced a restatement, is in docs/experiments/2026-08-13-restated-under-a-production-faithful-engine.md.

What this means — and does not mean:

• It does not mean the engine is broken. The rails, the attested screening, and the measurement machinery are the project's substance, and they worked: they measured the rules honestly and refused to promote what did not clear.
• It does mean nobody should deploy keel expecting alpha. The honest posture is: this is an auditable compliance engine and a measurement instrument, with a reference agent whose rules currently do not pay for their fees at Coinbase's taker rate.
• It means fee structure is a first-class research question here, not an implementation detail — see the experiment record for where a maker-fee path or a subscription-tier allowance changes the picture.

A visitor who discovers this themselves feels misled; a visitor told upfront can read it as what it is — rigour. That is why this post exists.

For the full picture — what keel is, the fiqh basis, the Arabic entry point, and what shipped in v0.9.0 — see the announcement: keel is open source.