When people pay us to do work, they should be able to trust how we do it. That puts a plain duty on us: know what the tools can do, check the result, and be clear about the limits.

Until September, we had published this tooling as open source and used it on our own projects. In September, Integro Labs began delivering paid client work with the same tools. The work now carries a responsibility to people who pay us and rely on us.

That shift gives the month’s releases their meaning. More tool choice, saved proof of decisions, repeatable imports, and checked data runs all bear on a simple test: can we explain what happened? We see each change through that duty. A feature earns trust through the checks around it and a clear account of what those checks cover.

Choice needs a clear basis

For paid work, we want tool choice to remain a practical option. We also need to know what each choice has passed before we rely on it.

AIWG places project guidance and reusable skills where supported AI tools can read them. September added Pi and Oh My Pi support, with each using its own setup paths. Both remained experimental. Pi also gained model lookup, runs without an interactive screen, and session import.

The value is in keeping project guidance useful across more tools. It gives a team more room to choose how it works. Yet a longer list of supported tools should lead to closer questions. Which parts are ready? Which checks still need to run? What changes when the provider changes?

September also fixed a less visible part of that choice. Refreshing AIWG for one provider now leaves the other providers’ files in place. A team should be able to maintain one setup without breaking another. That is a small repair with a direct link to the freedom to switch.

A decision needs something we can check

Being paid for the work makes a confident answer an inadequate basis for trust. We need a way to examine the choice and its limits.

AIWG’s Decision Engine shipped an opt-in alpha contract in September. A contract sets out the inputs, rules, and expected result in a shared format. It also saves a receipt: a record of the call that can be checked later. The contract sets limits on retries and fallback paths. A different backend can use that same contract.

That gives review a concrete starting point. What was the tool asked to decide? Which rules applied? What did it return? We want those questions to have answers outside the chat where the work began.

The alpha status matters as much as the new contract. The runtime stays off by default. Its offline playground runs saved examples through the real evaluator, but those checks do not prove a live model is right. A record of a decision also does not grant permission to act on it.

Agentic Sandbox calls for the same care in how we describe progress. Its September report could not confirm a release from the available records. That leaves no new version established by that report. We should hold claims to the proof we have, especially when others are paying us to exercise judgment.

Recovery belongs in the plan

Paid delivery gives routine data handling more weight. A retry or restore should have a scope we can explain before it runs.

Fortemi, the release name for matric-memory, added imports that track notes by their source. Repeating an exact batch returns the prior note IDs without adding copies. If the text has changed, the caller chooses whether to make a new version, replace it, or report a conflict.

Those rules turn a retry into a defined step. A lost reply need not mean guessing whether to send the same batch again. The source IDs let the server match the repeated request to the notes already stored.

September’s restore fixes protect the data outside the chosen set and keep retained record IDs. A clash with a record beyond that scope causes the restore to fail and roll back. Purge previews show what is due for removal. The final cleanup receipt records the result without retaining the erased text.

fortemi-react carries related work into browser apps. Its PGlite restore for the full-v1 archive profile now writes to live storage. Notes can be edited after import, and later exports include those edits. The current shard profiles still leave out live source identities and report that loss.

The limits help us judge what a recovery path is fit to do. These fixes support specific import and restore behavior. They do not establish a complete backup or loss-free transfer across the whole suite.

Completion should be checked

A claim that work is done should point to something we can inspect. For paid work, that is a standard we want built into the tooling.

Fortemi’s dataset tools provide steps to preview, run, check status, cancel, and resume work. They check the selected contract and limits before making a call. The server’s dataset runs remain alpha for live storage.

The new fortemi-react adapter connects a reviewed plan to the request that runs it. It checks the final receipt, then asks the server for a separate check before accepting completion. The workflow can keep an unclear result distinct from a verified success. This remote adapter also remains alpha.

That is the useful detail behind “verified dataset runs.” It means checks tied to a plan and a result, with stated limits on what is proven. It gives us a basis for review without treating every returned reply as success.

Pagenary brings the month’s data work to publishing. September added source refresh, dataset imports, and search within a chosen scope. The publisher package also gained guides for its programming interface, setup, and authoring. Those guides put the instructions beside the tool a team uses.

The responsibility grows; access stays open

The move into paid client work changes what is at stake for us. It makes the ability to choose, review, retry, and check results central to how we judge the tools. The open stack remains open.

All Matric-stack and Integro projects leverage AIWG and the rest of the Agentic Operating System. AOS remains open and independent so it can be freely shared with other teams and the public. Adoption of AOS is not gated by any commercial relationship with Matric or Integro.

Integro Labs is the operating and licensing company. That role does not make it the holder of the platform’s intellectual property. Teams can adopt the open tools on their own terms.

We now deliver paid work with tools that others remain free to adopt. Our responsibility is to earn trust in how we use them.

To discuss where owned automation fits your team, book a discovery call or send a no-fee inquiry.

September reports, with links prepared for publication:

Make owned capability practical.

Bring us the operating problem, brittle workflow, or automation boundary you need to make inspectable.

book a discovery callstart an inquiry