This is not a technical question. It is the question you will be asked — by a client, an inspector, a judge — when something goes wrong.
A system that cannot explain its decision using the judgment that produced it is not a working tool: it is a risk with an interface.
If the answer is "the system", there is no answer. Accountability is not delegated to software — yet today it is not recorded either.
The question your team is already asking under its breath. And the one keeping AI out of the places where it would actually matter.
The problem is not that AI fails. It is that it fails leaving no trace of who decided what, under whose authority, and by what criteria. A UC Berkeley study of multi-agent systems put a number on it: more than three quarters of observed failures were not reasoning failures but coordination and specification failures — nobody had defined who could do what, or within what limits. Cemri et al., “Why Do Multi-Agent LLM Systems Fail?”, UC Berkeley.
In a law firm, an insurer, a bank, that is not a technical fault. It is a case file you cannot defend.
Four capabilities. Not one. And all four leave a record with a name on it.
It does the work end to end: reads the document, interprets it, checks it against your rules, and delivers the complete case file.
Faced with a worse document, it does not do the same thing at lower quality: it uses less, and states which of its tools it left out, and why.
Faced with material that does not allow a reliable judgment, it refuses to work — and that refusal is recorded with its author, its reason in figures, and its timestamp. It does not hand over a half-finished report that looks complete.
You adjust its limits when you need to, within what its contract permits. Without directing every move, and without anything changing behind your back.
At the end of the road, what you are left with is not a software license: it is your firm's knowledge turned into an asset that is yours, with its name, seal and signature. And it is portable: if one day you stop working with us, the asset leaves with you.
Everything this page says can be checked. These are the figures, and they come from running the full suite, not from a slide deck.
2,502 tests across 278 suites, 0 failures and 1 skipped, over the full accumulated stack — run live, and every figure auditable against the repository. Measured 2026-10-19 on 8b536cd.
On the deciding path there is no model at all: they are readable rules. The same case twice gives the same answer twice — and that is what lets you defend it two years later.
Limits are not applied at the end, like a filter. They govern from the moment a part is built until it produces an effect, without changing shape along the way.
Each verdict is chained with keyless SHA-256: anyone can re-verify it with their own tools. Modifying an entry breaks the chain and is detected. Against a full rewrite by whoever controls the database, the defence is external anchoring of the chain — being built.
CCS v0.1.1 · NIST CAISI
Docket NIST‑2025‑0035 · Comment NIST‑2025‑0035‑0180
Tracking mm9‑m8z8‑mjjx · FR Doc 2026‑00206 · under evaluation
Public submission with six downloadable technical documents: specification, comparative analysis, roadmap, valuation and signed letter.
Verifiable at:
https://www.regulations.gov/comment/NIST-2025-0035-0180
Half an hour, with real documents, on the working system. There are no slides: it runs in front of you, and you are also shown what the system decides not to do.
Because it is where error costs most and where proof is worth most. A law firm does not need AI to be fast: it needs to be able to explain, two years later, why a contract was reviewed one way and not another — and who answered for it.
What we learn there serves any regulated activity afterwards. But you start where the demands are highest, not where they are comfortable.
There are tools that observe what AI did and record it afterwards. There are platforms that orchestrate tasks and chain them together. Orchestix does something else: it builds the part that will decide, with its limits inside it, and then governs what that part does. The difference shows on the day someone has to answer for a decision.
The road runs from the legal vertical towards activities where accountability is regulated. No dates, because dates on this page would be a promise, and here there are only things you can check.
When you adjust a limit, the change descends through the entire structure without losing its reason. And every part that receives it answers with evidence of its own.