12 docs tagged with "sentinel"

The first systematic experiment, and the day Plan became Charter

Naming a design property that was already silently doing its job

Bun's Rust rewrite ran AI adversarial review at a scale worth studying — one model family, split context windows, and 1,386,826 test assertions as the final judge. StrayMark's audits look different because the target is different, and the difference isn't stylistic. When a passing test suite can be the judge, you isolate reviewers by context. When the judge is a human judgment call, you diversify by model family — and 25 audit cycles of real data show exactly why, including one family that went quietly blind.

The dry run proved that signal coverage — not granularity — was what kept 57% of the routing on low-confidence guesses. The fix was almost free. The person creating a unit of work already knows whether it's design, implementation, an audit, or mechanical labor; they know it for the cost of typing one word. So Baton discontinued title-scanning entirely and made a declared verb the single authoritative signal — then graduated it to the framework.

Follow-ups became StrayMark's second first-class entity — and within a day, two external migrations stress-tested the design. The reference adopter's deprecated bash script had been reporting "in sync" while 29 entries sat invisible to it.

Ten latent gaps surfaced in a single polish Charter, and the anti-pattern that earned a name

Baton's second phase finally looked at money — but it routed nothing. It classified the 762 work units StrayMark had already recorded in Sentinel, recommended a tier for each, and printed what a routing policy would have spent versus all-frontier. The headline was a ~93% illustrative ceiling. The honest reading inverted the experiment's own hypothesis — granularity was never the lever. Signal coverage was.

Loom's first release rendered the document graph live in the browser in under a second. It also drew 330 edges to nowhere. The visible feature was the easy part — making the references actually resolve was the work, and it taught us when a dangling edge is a bug and when it's a signal.

Before Loom could draw anything, StrayMark's document model had to stop being the CLI's private property. The refactor that shipped zero user-facing change was the precondition for everything visual that followed — and a small lesson the framework had been preaching to adopters all along.

A fourth experiment started with an economic problem — when the subscription subsidies end, all-frontier-all-the-time stops being affordable — and immediately ran into a structural one. You can't route work to a cheaper model if the work is drifting from its own plan; a cheaper router over forgotten intent only automates the drift faster. So Baton's first move touches no models at all. It reads the plan StrayMark had only ever proved *existed*.

The operator's real daily question isn't "show me the documents," it's "where are we?" Loom grew a second surface — the architecture as a building you can walk, in 2D and in 3D — and its technical-debt overlay was empty. Not because there was no debt. Because it didn't know where to put it.

StrayMark's audit is only worth anything if independent model families converge on a finding by themselves — agreement is the signal. A run of releases hardened that guarantee against three ways to fake it. The freshest is the sharpest — a router CLI injects its own product name, so an auditor stamped its report "qwen-code" even after the operator had switched the backend model. The convergence math was being fed a lie about who did the work.