Verified Work
QualityHarness
A quality layer for your AI work. It makes the model verify facts before using them, write a spec before multi-step tasks, and check its own work for the shortcuts that only look finished. Off by default: turn it on for every message.
Off by Default, On When You Want It
The Quality Harness lives in Settings under AI Providers. It is off until you turn it on, so nothing changes unless you ask for it. Set it to Full for planning, self-checks, and a review pass after work turns, or Light for a quick verify-before-done on every message. Full suits the Plexon model and other fast models; Claude models already do most of this, so Light is usually enough there.
- Full, Light, or Off in one control
- Off by default for every provider and model
- Full adds planning, self-checks, and a post-work review
- Light adds a short verify-before-done to every message
A Verified-Work Contract
At Full, the model works under a plain rule: every claim it makes about its work is something it checked, not something it guessed. It verifies APIs and facts against the real source before using them, writes a short spec before multi-step work, and debugs by reproducing the problem with one hypothesis at a time.
- Verify APIs, config keys, and paths before writing them
- Write goal, done-when, never-touch, and stop-if before multi-step work
- Reproduce, form one hypothesis, change one thing, re-check
- Stop and report after three failed attempts instead of thrashing
The Eleven Fake-Done Checks
Before saying done, the model runs a hostile pass over its own work and looks for the shortcuts that make work only look finished: a relaxed test, a swallowed error, a stub return, an invented API, a task solved that was not the one asked. If it finds one, it undoes it and does the real fix.
- Relaxed tests, swallowed errors, and comment-as-fix
- Stub returns and invented APIs that were never verified
- Happy-path-only work and quiet scope creep
- Off-spec completions: solving a task near the one you asked for
Skills That Load Themselves
Seven quality skills load automatically and stay hidden from the skills panel: verify adversarially, debug systematically, spec first, context budget, tool restraint, subagent fan-out, and sourced claims. Full loads all seven, Light loads the core three. The Software Developer persona adds five coding skills for security review, SQL and migrations, test hardening, git hygiene, and stack traces.
- Seven general skills load with no setup and no clutter in the panel
- Light loads the three core skills, Full loads all seven
- Coding skills ship with the Software Developer persona
- Sourced claims: every fact traces to a source, no invented numbers or links
A Reviewer After the Work
After any turn that changed something, a read-only reviewer re-opens the files and checks the assistant's claims against what is actually there. It never edits. It posts a short pass, fail, or unverifiable verdict to your Inbox. Read-only turns and plain questions do not trigger it, so there is no noise on work that changed nothing.
- Runs only after a turn that wrote files or ran commands
- Reads the real files, never edits, files a verdict in the Inbox
- Pass, fail, or unverifiable, worst findings first
- A mid-turn reminder to re-run builds and tests before finishing
The @loop Agent
Mention @loop in chat to run one task all the way to a verified end. It writes a spec on disk (goal, done-when, never-touch, stop-if), works step by step, checks its own claims after each chunk, and reports honest status with evidence: what passed, what is left, and any decision it made without you.
- Writes a real spec before starting, then works against it
- Self-verifies each chunk with the same fake-done checks
- Reports done, partial, or blocked with the evidence
- Names any decision it had to make on its own
How It Works
Verified-Work Contract
At Full, the model treats every claim about its work as something it checked. It verifies facts against the source and hunts eleven fake-done shortcuts before saying done.
Post-Turn Reviewer
After a turn that changed files, a read-only agent re-checks the claims against the real files and posts a pass or fail verdict to your Inbox. Read-only turns skip it.
@loop Agent
Mention @loop to run a task to a verified end: it writes a spec on disk, works step by step, self-verifies each chunk, and reports honest status with evidence.
Off Until You Ask
Off by default for every provider. Set it to Light or Full to apply it to every message.
Connects to
Nothing here works alone. See the whole map for how the pieces fit together.
Where it goes
This extends AI Chat & Modes.
Turn it on and the model writes a spec before multi-step work and files a verdict after.