Project Starfish DevlogGitHub
Devlog

Building a governance-first AI

A build journal: how Project Starfish went from a single decision (deny by default) to governing real agents.

2026-11-07

Every yes was a security finding

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.

2026-10-11

A table is not a control

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.

2026-09-28

The test that didn't cover itself

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.

2026-09-27

Deep dive: why the risk score is a maximum, not an average

Fifty categories, and one arithmetic decision that decides whether the whole thing works. Plus the exactly four things no tolerance setting can lift.

2026-09-27

Deep dive: what a hash chain cannot see

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?

2026-09-27

Deep dive: narrower than your grants

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.

2026-09-27

Deep dive: volume buys nothing

An arithmetic mean is trivially gamed by flooding corroboration. Five mechanisms to replace it, because there are five separate ways to fake agreement.

2026-09-26

Letting a model advise without letting it decide

Admitting a probabilistic signal into a deny-by-default system needs one property: it can raise risk, never lower it.

2026-08-23

Twelve questions, and the three modules nothing called

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".

2026-08-01

Threats as governed evidence, not a signature list

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.

2026-07-22

Turning the audit on ourselves

Four adversarial sweeps over our own core. Twenty-six gaps — and most of them are the code a careful engineer would have written.

2026-07-19

MOSAIC, and the fix that was already sitting there

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.

2026-07-10

Risk in numbers, and a calmer cockpit

A 50-category, 0–100 risk score; an operator Risk-Tolerance dial that never lifts the hard floors; and a calm, one-screen cockpit.

2026-07-04

The 1.0 candidate: ten releases of hardening

Closing the code audit, a durable audit log, and governance as an embeddable surface.

2026-06-20

Governing Claude Code itself

Deny-by-default as an overlay, and a lockdown the agent cannot switch off.

2026-06-19

Going public: Apache-2.0, on npm and GitHub

Free for personal and commercial use, with the trademark as the moat.

2026-06-16

No more 'my agent deleted my drive'

Governed, reversible deletion and a Custodian bound by hard rules.

2026-06-15

Model-agnostic, and proving the work

The runtime spine, model independence, and the Evidence Gate: no unbacked word.

2026-06-14

Burning the boat: from strangler to clean-room

We stopped salvaging the fork, rebuilt clean, and hardened the skin.

2026-06-09

Toby and Hank: the only door in, and a watcher that cannot lie

Capability intake, runtime monitoring, and the portable overlay that became the product.

2026-06-08

The governed core: a deny-by-default PDP, a hash-chained audit, and a boundary engine

Phases 0 through 4: the trusted base everything else stands on.

2026-06-04

Why we are building Project Starfish

Everyone ships skills. Nobody ships governance.

Project Starfish · a governance-first, deny-by-default AI ecosystem · Apache-2.0. This devlog is a backdated build journal reconstructed from the project history.