Verbatim reads successive versions of a regulatory proceeding and works out what changed. Before it will state anything, it re-reads the stored filing at the offsets its own citation gives and compares the two strings for equality, not for likeness. Below is a claim that failed that comparison, and what the product does about it.
Exhibit — MPUC-2026-0142, two claims, one verifier
Withheld
No claim made
quoted text does not match the source at the cited offsets
Asserted
The compliance date for the updated load forecast moved from March 1, 2027 to June 1, 2027.
§ 7.1 · v2 · characters 7577–7750 · verifiedBoth claims come out of the same proceeding, and the same code decided each. The first has no sentence to show you: the slot where one would go is struck through, and what stands in it is the reason and the two strings that differ. They differ by one character. The second survived the same check, so it speaks. Nobody chose which — equality is not an opinion, and there is no confidence score here to soften either answer.
No slides and no mock-up. This is the application running against its corpus. Seventy seconds, five movements: a filing Kentucky published in two versions a million characters long and 127 characters apart, the 144 changes between them, and a claim that will not assert itself — with the source beside the citation that misquoted it.
What it shows, for anyone who would rather read it. Five movements, each ending on one line in the accent colour saying what it was for.
1. The task nobody can do by hand. Kentucky PSC 2025-00113, and one testimony in it that was filed and then corrected. Each version runs to a million characters — 1,024,409 against 1,024,536. That is 94,440 words; at 200 words a minute, about eight hours of reading, each. The two differ by 127 characters. Nobody finds 127 characters in a million by reading.
2. Found. Verbatim compared the two versions and found 144 changes; the whole comparison takes four tenths of a second. One of them is opened: the passage that read “Brown BESS” and now reads “Cane Run BESS”, each side addressed to exact characters of the stored filing. Every change points at the characters it came from.
3. The refusal. The corpus changes here and the film says so first: this failure is planted on a fixture, because no filer publishes a fabricated citation to order. A claim was written about the passage and it is not on the page. Its citation quoted the filing as ten megawatts; the filing says twenty. Nothing the claim would have said is printed in any field — not greyed out, not struck through, absent. The source sits beside it with the cited span marked, and the card's own footer says where the refusal went: to the review queue, under the reason it failed. The quote is not in the source, so the claim is not made.
4. What follows. One layer out: which of the company's duties the change touches and whose desk it lands on. It found a likely duty and its owner, and still routes to nobody, because the mapping was proposed by a word overlap and no person has confirmed it. There are sixteen ways routing can refuse, and the card prints the one that stopped this. It will not name an owner it cannot stand behind.
5. Close. Back on the withheld claim. You act on what it proves. It says nothing else.
What is not in the film, said here rather than left to be noticed. Logins, permissions, approval routes, the security headers and the accessibility sweep, and the review queue as a screen of its own. All of them are in the product and none of them is in this cut: they prove completeness, and the film is an argument. The eight hours above is arithmetic on a word count, not a measurement of anybody — no user of this product has been interviewed, which the user research page says in its first line.
The demonstration workspace is open, and it carries two corpora. One is invented — one docket, three versions, with the traps built in on purpose — and the eval numbers are measured against it, because a trap you wrote is a trap you can score. The other is 102 real filings the state commissions publish themselves, one pair of them loaded, each carrying the URL it came from. Nothing here is confidential and nothing here belongs to a customer: public record is public, and there are no customers.
Five of those filings are read closely in Five readings of the public record — what actually changed between two versions, what it cost to miss it, and which of them this product would have caught. They are the argument for the wedge, made against filings anybody can check rather than against the invented docket.
Three roles to sign in as, so the permission boundaries are visible rather than described. The sign-in page carries a button for each — there is no password to copy and none printed anywhere.
| Role | Sign in as | What it can reach |
|---|---|---|
| Analyst | Reads, proposes, resolves escalations, steers. Cannot approve. | |
| Obligation owner | Approves and rejects the actions the analyst proposes. Cannot propose them. | |
| Administrator | Users, the approval route, the share registry. Cannot approve a claim. |
Each name is a button. It signs in as that person and lands on their workspace — the masthead then names who you are and which role you are holding, so a screen that refuses you says which authority you were missing.
Roles are a shortcut, not the model. Underneath, an administrator grants any one of fifteen permissions to any person in the workspace and names whatever set they end up with. The workspace ships with ten templates drawn from real filings — regulatory analyst, regulatory affairs manager, regulatory counsel, rates and pricing, interconnection, load forecasting, system operations, certifying officer, workspace administrator, auditor — because starting from nothing is worse than starting from something. A template is copied, never edited: a role called “analyst” that one company has quietly changed means something different on every screen that names it.
No template holds both action.propose and
action.approve. A four-person team will want that combination and may grant it;
the product will not hand it over silently, and the conflict report names whoever holds
both.
These accounts hold nothing but the demonstration corpora, and the workspace behind them
is disposable: make reset-demo puts it back the way it shipped, so a visitor cannot
leave it broken for the next one. There was a password printed here until 2026-08-04 and the
argument for printing it — a door held open, not a secret kept badly — was only half
right. The door still stands open; what changed is that nobody has to be handed a working
credential to walk through it, and no string on this page can be pasted into a deployment that
forgot to turn the demonstration off. Sign in as the analyst, then as the owner, and watch the
same screen refuse a different thing.
If you carry regulatory obligations across more than one commission and want a workspace of your own with your own dockets in it, write and say so. Tell me which jurisdictions and what you are tracking; that is enough to know whether this is any use to you.
It reaches one person, and one person answers it. There is no waiting list being managed and no drip sequence behind that link.
One stage calls a language model. Every other stage is deterministic, and the one that decides what may be asserted calls no model at all.
Each filing is stored byte for byte and hashed. Passages carry stable character offsets, because an offset is checkable by reading the bytes between two integers.
Two versions in, typed changes out. A pure function — no model, no clock. When a section is renumbered rather than edited, alignment confidence drops and the change escalates instead of asserting.
The only stage permitted to call a model. It judges whether a change matters. It is never asked whether a change exists.
The gate. It re-reads the stored filing at the cited offsets and compares for equality, not similarity. A paraphrase is what an auditor will not accept.
Draft orders produce monitor and comment. Final orders produce comply, with an effective date. The two never share a code path.
Every decision, approval and threshold change lands in an append-only, hash-chained trail carrying the actor, the timestamp and the citation.
Most of this category is built to always have an answer. A compliance tool that is confidently wrong is worse than one that says nothing, because the analyst who cannot trust the citation has to re-read the source — which is the work the product claimed to remove.
Citations are compared for exact equality after normalisation. Whitespace and smart quotes fold; digits, units and dates never do.
Counts state their denominators. Precision the sample cannot support is a way of being wrong that looks like rigour.
Every summary states how many findings it rests on and how many were withheld, at equal weight. A conclusion that quietly drops what it could not substantiate is the failure this product exists to prevent.
The analyst who interprets a change is never the person who approves the action that follows. Authorship is read from the audit trail, not from a checkbox.
A regulatory affairs analyst at a multi-state investor-owned utility. The person who reads the docket, works out what changed, and tells the business what it must do.
Not counsel, not the compliance officer, not an executive. One user, designed for deliberately. A multi-state holding company does not run one open docket — it runs one per jurisdiction, each on its own procedural clock, each amended independently, and no single filing tells the company everything it now owes.