Audit
Real product checks become the source material.
02 / AI Laboratory · Content Engine
Live autonomous pipeline
The loop runs itself. Judgment does not.
Agentability’s research data became a closed content system with a deliberately narrow publishing boundary.
Scores are computed. Explanations are generated from evidence. Editorial work waits at a human gate. Automation can publish only inside the folders it owns.
01
The bounded flywheel
Autonomy is a loop,
not an absence.
Each station has a contract. The system can discover, generate, publish and measure—but the editorial threshold remains a human decision.
Measure gaps
SQL identifies patterns and content opportunities.
Generate
The model explains evidence it is explicitly given.
Review
Editorial content waits for approval or rejection.
Publish
Approved or auto-class content enters owned folders.
Measure
Performance and self-audits feed the next cycle.
02
The reasoning boundary
Calculate what is true.
Reason about what it means.
The reliable parts stay deterministic: scores, tiers, rankings, page selection and priority. The model receives those facts and writes grounded explanation—not numbers it is free to invent.
This split made the system easier to audit and easier to trust. It also made failures diagnosable: data logic and prose generation can be inspected independently.
Calculates.
SQL and code select evidence, compute scores, enforce status and decide what is eligible to move.
Reasons.
Language models turn supplied findings into explanations, recommendations and drafts.
03
What is actually live
A working system,
with a narrow claim.
The source snapshot contains 300 audit targets, 331 audit rows and 9,599 checks. The engine published 100 product pages and now runs eight workflow roles across generation, review, publishing, distribution drafts and measurement.
The often-quoted “about ten minutes a week” is the operating target for approval—not a measured productivity outcome. Social distribution remains draft-only and manual.
The filesystem is the governance model.
Automation can write only to mounted content directories. It cannot reach the homepage, privacy page, report system or machine-readable identity surface. Least privilege is designed into the operating environment.
04
The decision that changed the system
The pre-publish audit gate was the wrong gate.
The audit API can evaluate only a live URL, and its application-weighted principles penalise editorial pages for behavior they do not need. Rather than stage a fake guarantee, the engine publishes inside its bounded surface and records a transparent post-publish self-audit. Measurement remains evidence, not theatre.