Agentic integrity in 0.2.7
QuorX is built so an agent's work is checked, not trusted. What 0.2.7 measured across models, what broke one release earlier, and what changed, in one page with links to the full articles.
QuorX is built around one idea: agentic integrity. An agent's work should be checked, not trusted. Deciders vote on what it may do. A reviewer opens and scores what it built. A safety layer guards every command. Benchmarks show whether a change really helped. And because QuorX runs any model on OpenRouter, all of that has to hold for a frontier model, a cheap fast one, and one that doesn't support tools at all.
QuorX 0.2.7 is 24 pull requests (#35 to #61, not counting #55, held for 0.2.8, and #60, the version bump). This page sums them up. Every number comes from a pull request, a release note or a test run, and one-run results are labelled.
Five releases in two days
- Read the request in full
A prompt enhancer spells out what a request implies, and YOLO waits 15 seconds for you before the deciders answer.
v0.2.3 - Open source
Apache 2.0, with SHA-256 checksums and a signed build attestation on every release.
v0.2.4 - Locked down by default
Much harder to attack, but the data files were left unreadable.
v0.2.5 - Repaired at start
QuorX probes its own files when it starts and repairs them. Nothing was lost.
v0.2.6 - Checked work across models
A cache-friendly prompt, model cards, diff review, staged releases, and a QuorX that answers only this computer.
v0.2.7
How QuorX checks an agent's work
Seeing integrity model by model
Every model behaves differently, so 0.2.7 shows QuorX's checks per model:
- Model cards for 18 models: limits, quirks, prompting tips from each provider's docs, and QuorX's own results. The agent reads the cards of the models it uses.
- Cache-hit share per model in the chat's usage view and the cost card, so it's clear which models really cost less over a long reply.
- Looks before writing, per model, in the On this computer table below: how many steps a reply reads before it writes anything.
- A harder build benchmark with hidden checks, so scores no longer sit at the ceiling.
The new checks are cheap on cheap models. On the same small tests, a fast model got the same correct result as a frontier one:
Both models found both planted bugs on the right lines with no false alarm, and both proposed a memory when expected and not when not. Two tasks each show that the checks work on a cheap model, not how often.
Your own numbers show here when you open this page in QuorX's News.
The numbers
The helpers were offered in all 151 runs, across 8 builder models, and never called. So 0.2.7 offers them only in projects with 40 or more files. A smaller, cache-friendly agent prompt has the rest.
The articles
- A smaller, cache-friendly agent prompt: why the conversation is the cost, and a benchmark that can still tell.
- Deciders vs creativity: literal checkers turn down what a request only implies.
- When a security hardening broke the data: a post-mortem of 0.2.5.
- Shipping safer: tests on every pull request, staged releases, and this computer only.
- Building a release in parallel: how the pull requests came together.
- Four workflow features: diff review, hooks, schedules and memory, for any model.