One Change, Whole Company
Your AI agents will make thousands of decisions in a regulated plant. When one fact changes, you must show what rested on it, across every department and every agent, as it stood on the day.
The problem. Each system (lab, batch records, quality, ERP, projects) keeps its own records. A failed standard, a withdrawn supplier certificate or a lapsed analyst qualification crosses all of them, and tracing it is a manual investigation. Agents acting across departments multiply the reach.
What this demo shows. Seven planted changes in a fictional company, each followed across departments with a verdict per record (still supported, needs review, recall assessment for released product, refuted): 463 records reached across three companies, 0 missed against the answer key, while each system alone missed 413. Agents trying to act on changed records were stopped every time; with the silos 21 went through. An inspector's question answered as known on the day and as known now.
How it works:
- One layer across the silos. Every record is a claim of its own system, lots resolved across their id forms, so a change follows the real genealogy.
- What still stands. A lot confirmed by an independent test stays supported; a retest on the same faulty standard does not; released product goes to recall assessment.
- Agents held to evidence. Every agent action names what it rested on, is checked before it acts, and is recalled when that changes.
The honest part. The company is synthetic and there are no connectors to real systems yet. The answer key and the layer follow the same rules, reviewed by an engineer with manufacturing experience, so the run shows them applied, not that every plant judges the same way. Fabric itself would need validating before GMP use; the existing systems stay the systems of record.
Fabric is the layer that lets a regulated company put AI agents to work and still answer what every decision rested on.
A biotech manufacturer under GMP and SOX runs a lab system, batch records, a quality system, an ERP, a regulatory register and a project tool, and now has AI agents working in every department. Each system keeps its own records. When one fact changes, the damage crosses all of them. Fabric holds every record as a claim of its system and follows each change across the whole company: what it reaches, what still stands, which agents must stop, and what was known on any date.
A fictional company, generated with planted changes and an answer key; every record, lot and number on this page is synthetic. Regulatory points are general knowledge, not legal or compliance advice.
1. Seven changes, and how far each one reaches
One company from the judged run. Each change's reach is every record already written that rests on what changed, with a verdict: still supported by independent evidence, needs review, recall assessment for product already released, or refuted. The red line is what each system working alone would have found.
2. Agents held to what they rest on
3. The inspector's question
4. The auditor's question: revenue in doubt at each quarter end
5. How it was judged
What this does not show
- The company is synthetic. There are no connectors to real lab, batch record, quality or ERP systems yet; a pilot starts with one chain, such as a calibration failure through to batch release and inventory.
- The answer key and the layer follow the same rules, so the run shows the rules applied exactly, not that they match every plant's judgement. The verdict rules were reviewed by an engineer with manufacturing experience: released product in doubt goes to recall assessment; a tightened specification leaves released lots unaffected; an excursion reaches the lots filled in that room that month.
- The silo baseline follows each system's own links only. Real teams bridge some gaps by hand, more slowly and less completely.
- Fabric itself would need validating before use in a GMP setting. The existing systems remain the systems of record; the layer runs across them.