Written 11 September 2026
Install it, and you get the record for everything a record can carry, plus a dated list of exactly which obligations your records evidence and which they do not.
Not you are compliant. Nobody can sell that, and the sentence that makes this worth buying is the second half rather than the first: you will know, before somebody asks, which of the six things Colorado's Rule 7.7 wants you can actually produce.
Because it is false, and because this library exists to catch exactly that claim in other people's products. The census reads ten systems and publishes what they cannot do. A tool from the same author that quietly implied a compliance verdict would end the only asset the work has, the first time an assessor checked.
Compliance is a judgement about a particular deployment by somebody qualified to make it. This maps, and it does not conclude.
Colorado's proposed Rule 7.7 requires a record showing six things when a meaningful human review is conducted. It is the sharpest test available because it is specific, it is dated, and it applies on 1 January 2027.
| What the rule wants | What an install delivers |
|---|---|
| The reviewer identity | Yes, where the approval path hands over an authenticated session. The record carries the approver and identity_source, which says whether the identity was resolved from authentication or supplied by the caller. It cannot invent an identity that was never captured. |
| Review timestamps | Yes. With the standing caveat that a timestamp is written by the party whose conduct is in question, which is issue #49 and is not solved. |
| A written justification | Yes, where the approval path captures one. The format carries a reason. |
| Whether the reviewer approved, modified or overrode | No. This is a gap in the format, not in any framework it reads. See below. |
| Primary evidence available to the reviewer | No, and not from here. Only the interface that rendered it knows what a person was shown. Nothing installed downstream can find out, and of six agent frameworks read, no approval object asserts it. |
| Reviewer authority and training, and their access to the system's intended use and limitations | Never. Attributes of a person in HR and identity systems, and a fact about what a console displayed. No record format produces them and none should claim to. |
Two of six are mechanical. One depends on the deployer's own interface. Two are not facts about a decision at all. And one is this library's own to close.
verdict is permitted or refused. The Testimony Record cannot express “modified” either, which means a reviewer who changed an action before allowing it is recorded as having approved it unchanged. That is issue #88, disclosed in the Colorado comment and in the testimony that went with it.
It matters more than any other row above. Rule 7.7 asks for the distinction twice, and then makes it evidentiary: an override that fully reverses a decision indicates that human review was meaningful. The rule keys its conclusion to a signal the format cannot yet emit. Three of five agent frameworks read cannot emit it either, which is why this is a real market rather than a deficiency peculiar to one product. This library is the only one of the four that publishes the gap about itself.
Reading one instrument is the right shape for advising on Colorado specifically and the wrong shape for the first question a deployer actually asks, which is not what does this text want but does anybody require this of me yet.
A single pass across every published reading answers it: which of them ask a record to show who intervened, sorted by whether they are law on a given date. As at 13 September 2026, from the readings in this build:
Every line of that comes from a force block in the reading it describes, and a reading whose commencement cannot be established from a held source says unknown rather than implying a date. One committed text carries a literal placeholder where its effective date belongs; the library says so rather than guessing one in.
Naming which instruments ask for a signal is not the same as knowing whether a deployer's own records supply it, because it never reads a record, only the law. Checking one file against every instrument at once closes that from the other side: which of the twenty a deployer fully answers, which they partly answer, and which ask a record for nothing at all. The clause-level detail against one named instrument is still there for the deployer who already knows which law applies to them and does not want the census version of the answer.
The library's fourth piece is deliberately not a new paid thing. It names the free adapter for one of five frameworks (LangGraph, CrewAI, the OpenAI Agents SDK, Pydantic AI, AutoGen), prints the install line and the wiring snippet verbatim from that adapter's own README, and chains straight into a full check on the record it just produced. The adapter is MIT, standalone, and depends on nothing sold here; it makes a system able to write who approved an action, from what identity source, with what outcome, going forward. That is the one field the census found missing in nine of ten deployed agent systems.
What this does not claim, stated as plainly as everywhere else on this site. Producing that field closes exactly the gap this library measures. It does not satisfy a law's requirements about a reviewer's training, authority or independence, and no record format should claim to. The tool says this in its own output rather than leaving it to be assumed, the same discipline every reading here holds every other instrument to.
Not every deployer is asking the census question. Somebody who already knows a client operates only in Korea does not want to be told which of twenty texts might apply; they want to know what Korea's Framework Act specifically asks for. Scoping the final check to one named instrument instead of all twenty answers that directly. The install step is unchanged either way: the adapter writes one record format regardless of which law it will be read against, and it is the check, not the record, that is jurisdiction-specific.
Not because of a licence check. Because of decay.
Colorado's rules were filed on 11 August 2026, revised on or about 23 September, and take effect on 1 January 2027. The instrument tracking them is dated and re-read when they move. A copy from last quarter does not stop working; it starts being wrong, which is the one thing a compliance buyer cannot tolerate.
That is why the statute and the proposed rules are kept as separate instruments. On any given date one of them may get re-read and the other may not, and a customer can see which. The maintenance is the product, demonstrated rather than described.
“The same thing again next quarter” is a promise until somebody can see it happen. A diff between two committed versions of the same reading reports the difference the way everything else in this library reports anything: obligations added, obligations removed, and whether force changed, none of it summarised from memory. The two versions are the two commits that actually touched the file, not a hand-written changelog kept alongside it, because a changelog can drift from what was actually read and a diff of the reading itself cannot.
An instrument read once and not yet re-read says so rather than inventing a predecessor to compare against, and reports exactly that.
The specification, the validator, the emitter, the conformance corpus, the census rubric and the converter. A specification nobody can implement without permission is not a specification, and the position this library is sold into requires not owning the thing it is measured against.
What is sold is this library and the maintenance of it: the mapping from an obligation somebody else wrote to a signal in a record, kept current, with the reading published so a buyer can check the map rather than trust it.
There is no price on this page on purpose. What a licence costs depends on how many instruments and how many seats, and the honest answer to that is a conversation rather than a table. Write to troy@machinetestimony.com with what you need read and what you need checked against.