There is a version of consultation-response software that reads the regulator pack, calls a large language model and prints a response. It exists. People have shipped it. It does not survive a regulator's reading, and the reason is the same reason a junior analyst's first draft does not survive a partner's reading. The structured questions a human must answer have not been answered yet, so the prose is filling space the evidence has not arrived in.
Probari Change is built around that observation. The chain runs in two passes, and the second pass does not begin until the human-side questions have been answered honestly. That is not a workflow choice; it is what makes the output defensible.
Why a single pass fails on regulator-facing work.
A regulatory consultation is not a piece of text to be summarised. It is a position framed by an organisation that has its own register, its own evidence base, its own commercial reading of the question, and its own legal appetite for what to disagree with and how. None of that is in the regulator's pack. A single-pass model that reads only the pack will produce a response that reads as if the firm had no view of its own, because the model had no view of the firm.
The same problem turns up at the level of stakeholder grounding. A regulator's pack typically frames a small number of stakeholder positions explicitly and assumes the rest. A single-pass model picks up only what is framed. The classes that are absent from the regulator's framing, and absent from the firm's evidence, get written out of the response. In a register the regulator reads carefully for what is missing as well as what is present, that is a finding the firm did not intend to make.
A third problem, mundane but lethal at submission. A single-pass model has nowhere to record uncertainty. It has only the surface of confident prose. Regulator-grade writing carries a register taxonomy, strong, conditional, soft, constructive-alternative, and the choice of register on each disagreement is a position in itself. A model with no representation of register will pick one and run, and the firm's general counsel will spend an evening unwinding the parts that have walked off in the wrong direction.
What the analyst questionnaire is for.
Pass one of Change does the work that does not require the firm to take a view. The research agents read the regulator pack, read the firm's attached files and URLs, and produce a structured PCMS pre-scoring across ten stakeholder classes and seven consideration types. Severity-rated. Disagreement-register tagged. Evidence-anchored.
Pass one ends by surfacing the analyst questionnaire. The questionnaire is not a "talk to the model" surface; it is the structured, lens-tagged set of questions that the research agents could not answer from the regulator's pack and the firm's attached evidence alone. Each question is the precise gap the renderer cannot draft around. Each carries a request for evidence, a citation, a screenshot, a paragraph of reasoning, a "not applicable" with a reason.
The research analyst is a human role. The analyst answers the questions the agents have surfaced, the question of whether the firm has tested the affordability model for protected-characteristic disparate impact, the question of what the firm's position is on the regulator's proposed forbearance baseline, the question of whether a particular operator practice has changed since the last published page. The analyst is not asked to write prose. The analyst is asked to record what the firm knows.
Why drafting is gated.
Pass two does not begin until the questionnaire is complete. The drafting gate is a hard rule, enforced at the route layer and again inside the synthesis prompt. There is no half-answered draft. There is no prose written into evidence the firm has not yet supplied.
This sounds like a friction. It is a feature. The cost of regulator-facing prose written without evidence is not a few rewrites; it is a position the firm cannot defend on the day the regulator asks how it was reached. The gate is what removes that exposure. By the time pass two runs, every claim the renderer is about to make has a recorded answer, recorded evidence and a recorded register.
How PCMS produces register-aware prose.
A generic AI assistant, asked to draft a consultation response, will produce confident prose at one register, usually somewhere between assertive and defensive, and that register will not be the right one for half the disagreements in the response. The Probari Consideration-Material Stakeholder framework is the structure that fixes that.
PCMS scores every claim against ten canonical stakeholder classes, the regulator, the firm, the customer, the competitor, the comparator (the comparison-intermediation actor where one is in scope), the supply chain, the workforce, the public interest, the equality duty, and the technical standard. Each score carries a consideration type, operational, commercial, customer, technical, data, supplier, legal, and a severity rating. Each carries a disagreement register where the firm's position differs from the regulator's framing.
The renderer reads the PCMS JSON and produces prose at the register the JSON specifies. Where the firm disagrees strongly with the regulator on a customer-impact framing and the evidence supports the disagreement, the renderer produces strong-register prose. Where the firm has a constructive alternative to a proposed obligation, the renderer produces constructive-alternative prose. The model is not picking the register; the structure is.
The same structure produces silences-as-findings. When a stakeholder class is not represented in the regulator's framing or in the firm's evidence, that absence is recorded as an ABSENT-prefixed finding in the output. The firm reads what is missing as a finding, not as nothing. The regulator reads it the same way.
Why sealed audit trails matter.
Regulator-facing work has a property that ordinary commercial writing does not. Six months after submission, the firm may need to prove that the version it sent is the version it relied on, that the evidence it cited was the evidence it had at the time, and that the reasoning it followed was the reasoning recorded in its own register. A draft saved to a shared drive does not prove any of that.
The Change Impact Memo PDF carries a hash-anchored cover page. The seal includes a SHA-256 hash of the rendered document, the chain components that produced it, the pass-one output, the analyst responses, the pass-two output, the renderer version, and a verify-seal endpoint that re-computes the hash and compares it to the cover. Two analysts running the same input produce the same memo. The audit trail is reproducible by construction, not by promise.
That property is doing two things at once. It is making the firm's position defensible to a regulator who asks how it was reached. And it is making the firm's reasoning reviewable to its own general counsel six months later, without anyone having to remember which version of the file was the live one. Both jobs require the seal to be on the artefact. Neither is a job a single-pass tool that produces flat prose can do.
The shape of the discipline.
Two passes. A drafting gate. A structured framework that produces register-aware prose. A sealed artefact that carries its own evidence chain. None of these are interesting in isolation. Together, they are the shape of consultation-response work that survives a regulator's reading, and the shape of consultation-response work that a compliance team can run on a Tuesday afternoon and have on the executive committee's desk by the end of the week.