A build journal: how Project Starfish went from a single decision (deny by default) to governing real agents.
We set out to show which agent session is waiting on you. Wiring it up turned up a defect: every tool call a person approved in Claude Code was audited as a no-silent-execution violation.
Seven risk categories are marked as floors no policy can lift. We asked whether any real input could trip one. DROP TABLE users was being allowed.
We built a test to catch correct-but-unwired modules. It works — and it's a hand-written list, so it could only catch the instances someone remembered.
Fifty categories, and one arithmetic decision that decides whether the whole thing works. Plus the exactly four things no tolerance setting can lift.
A chain proves what came before each record and says nothing about what came after — so truncation is invisible from inside. One number fixes it. Then what if that number's own write fails?
The Scope Contract. Sealed at approval, amendable only by someone who is not the proposer, and tighter than the grants it derives from in exactly one place.
An arithmetic mean is trivially gamed by flooding corroboration. Five mechanisms to replace it, because there are five separate ways to fake agreement.
Admitting a probabilistic signal into a deny-by-default system needs one property: it can raise risk, never lower it.
An external reviewer asked what each control does when absent, disabled, failed or lying. Ten findings — and a green suite could not tell "works" from "runs".
A detection earns authority through a lifecycle: reproduced, corroborated by a quorum, approved by someone who is not the publisher, sealed, on the hash chain.
Four adversarial sweeps over our own core. Twenty-six gaps — and most of them are the code a careful engineer would have written.
A CLI command-composition attack paper sent us back to our own threat model — and found a hardened defense we'd built a month earlier but never wired in.
A 50-category, 0–100 risk score; an operator Risk-Tolerance dial that never lifts the hard floors; and a calm, one-screen cockpit.
Closing the code audit, a durable audit log, and governance as an embeddable surface.
Deny-by-default as an overlay, and a lockdown the agent cannot switch off.
Free for personal and commercial use, with the trademark as the moat.
Governed, reversible deletion and a Custodian bound by hard rules.
The runtime spine, model independence, and the Evidence Gate: no unbacked word.
We stopped salvaging the fork, rebuilt clean, and hardened the skin.
Capability intake, runtime monitoring, and the portable overlay that became the product.
Phases 0 through 4: the trusted base everything else stands on.
Everyone ships skills. Nobody ships governance.