This is a synthetic persona. It is not a user interview, it is not a transcript, and it is not evidence. Nobody said any of this. Every line below was assembled by a model from public filings in data/real/, by the same person who built the product. Real conversations live in docs/user-research.html and nothing here may be moved there.
Read this before you cite anything on this page. A synthetic user is a hypothesis generator. Its output is a list of things to go and check, never a finding. The one thing this submission cannot fake is having talked to real people, and this is the exact document that could be mistaken for having done so. If a sentence from this page ever appears in the PRD, the MRD or the briefing as a user finding, the submission has done itself active harm.
What it is legitimately for. Three things and no more. Rehearsing questions, so a real twenty minutes is not spent discovering the obvious. Generating candidate pains to check against real sources. Answering a design question when no real user is available, with the guess plainly marked as a guess.
Not from priors about what a regulatory analyst probably thinks. A persona assembled from a model's general sense of a job is a stereotype, and it produces exactly the pains the product already claims to solve, which makes it worthless as a check.
Everything on this page traces to data/real/: 102 public filings, roughly 4,023 pages, across eight commissions — Georgia, Indiana, Utah, North Carolina, Ohio, Missouri, Virginia and Kentucky. Each carries a provenance JSON naming its docket, filer, date, source URL and SHA-256. These are the actual words of the actual people doing this work, several of whom sit on the outreach list.
Three traits below were re-verified by hand against the source files while writing this page, because they are the load-bearing ones: the Utah redline that is wrong in both directions, the Missouri stipulation whose only real change is a page that is missing from the original, and the Utah allocation-factor sentence. The rest carry their file and line reference so anyone can do the same.
Corinne Halloran. Senior Analyst, Regulatory Affairs, at Cordera Power Company — an investor-owned electric utility answering to five state commissions and to FERC. Eleven years in, the last six on the regulatory side after starting in rate design. She is not counsel, not the compliance officer, not the executive. She is the person who opens the filing.
The name and the employer are invented. No person or company in data/real/ is being portrayed. The role is not invented: the corpus is full of people holding it by name, and the ladder Corinne sits on appears in the record twice for the same man in the same commission — Analyst II, Regulatory Affairs in one Indiana cause, Manager, Regulatory Affairs in a later one.
Her desk. Two copies of everything, one public and one confidential. A service list that delivers PDFs by email. Workpapers in Excel with tabs that disagree with each other. A rehearing clock measured in days, not weeks. Recurring obligations born inside orders — an annual filing due 31 March, semi-annual reports, evaluation plans due within 90 days — none of which live in the docket that created them for long.
The moment Verbatim claims to serve her. A second version of something she has already read lands, and she has to establish what changed, whether it binds yet, what it touches, and who has to see it this week.
This is the point of the document. Every trait sits beside the source that justifies it. A trait with no source does not appear. Rows marked ASSUMED say what would confirm or kill them.
| Trait | What it means at the desk | Source |
|---|---|---|
| GROUNDED The title is real and so is the ladder | "Regulatory affairs analyst" is not a category invented for a pitch deck. It is a job people hold and file under, and it has rungs. | "Austin J. Baker, Analyst II, Regulatory Affairs for AES Indiana" — in-45911-final-order.txt:78; the same man later as "Austin J. Baker, Manager, Regulatory Affairs, AES Indiana" — in-46258-final-order.txt:135. Also "Kimberly Aliff, Revenue Requirements Manager in Regulatory Affairs" (in-45911-final-order.txt:75), "Chad A. Rogers, Director Regulatory Affairs Indiana" (in-46258-joint-motion…txt:245), "Steven Wills, Senior Director, Regulatory Affairs" (nc-e7-sub1329-dec-direct-testimony-abernathy-to-fogg.txt:181). |
| GROUNDED Her work is numbers in schedules, not prose | When this role sponsors testimony it is about adjustments, normalisation and schedule references — not narrative argument. Verbatim's output has to survive being read by someone who checks arithmetic for a living. | "Mr. Baker addresses the electric operating revenue adjustments from AES Indiana Financial Exhibit AESI-OPER, Schedule REV3 and REV4 to account for test year rider revenues and to reasonably normalize and annualize revenue." — in-46258-aes-indiana-verified-petition.txt:1617-1621 |
| GROUNDED She signs the correction herself | When a filing goes wrong, the person who tells the Commission is this role, in his own name, with his own email address on the letter. Not counsel. This is the single closest thing in the corpus to ADR-01's user doing ADR-01's job. | "/s/ Jeremiah C. Haswell — Director, Regulatory Affairs — Georgia Power Company — jhaswell@southernco.com", on the letter correcting two quarterly reports. ga-55378-large-load-econ-dev-report-q3-q4-2025-revised-pd.txt |
| GROUNDED She finds out informally, not from a check | The trigger for catching a two-quarter error was a person mentioning it. There was no control that caught it. This is the gap the product claims. | "It recently came to my attention that there is an error in the public disclosure version…" — same letter. Error present in reports filed 21 Nov 2025 and 11 Feb 2026, corrected 4 Mar 2026. |
| Trait | What it means at the desk | Source |
|---|---|---|
| GROUNDED Two copies of everything | Public and confidential versions of the same document, maintained in parallel. They can disagree, and the act of producing one from the other is itself a step that corrupts data. | "The error resulted from a formula being inadvertently disturbed by the Company's redaction process used to create the public disclosure version." … "The attachments to the trade secret versions of these Quarterly Reports were not impacted by the error and, therefore, are not being refiled." — ga-55378-…-q3-q4-2025-revised-pd.txt. TRADE SECRET / PUBLIC DISCLOSURE pairs run through the whole Georgia set. |
| GROUNDED Arrival is email against a service list | Not a portal she logs into on a schedule. A PDF in an inbox, served on parties, often scanned. | Certificates of service "by electronic mail" throughout; "cc: (w/ encl. – via email transmission) to Service List in Cause No. 44688" — in-44688-nipsco-compliance-filing-pmc-update-v1-original.txt:236 |
| GROUNDED The text is OCR-damaged | Any tool that matches on exact strings meets "fded" for "filed" and "Sendee" for "Service" on day one. A citation verifier has to survive a corpus that is not clean. | "the Company fded its Application" — va-scc-PUR-2025-00058-company-blackwell-errata-direct.txt:49. "Network Integration Transmission Sendee ("NITS")" — va-scc-PUR-2026-00056-odec-dembek-corrected-testimony-clean-redline.txt:73. And in one Georgia order the same paragraph appears twice, once as "no later than March 31" and once as "no later than March 3 1" — ga-56002-order-adopting-stipulation-2025-irp.txt:234 and :378. |
| GROUNDED Obligations are born inside orders, with dates | The thing she has to track is not the order. It is the recurring duty the order creates, which then outlives it. | "The Company will continue to provide CCR ARO semi-annual reports to the Commission. In addition, the Company will continue to file the ECS annually with the Commission no later than March 31 of each year." and "detailed evaluation plans for each of the approved DSM programs within 90 days of the selection of the Program Implementers" — ga-56002-order-adopting-stipulation-2025-irp.txt:234, :315 |
| GROUNDED The windows are short and statutory | Thirty days to seek rehearing. Fifteen to respond. Sixty for the Commission to act. Missing one is not recoverable by working harder afterwards. | "a party may file a request for agency review or rehearing with the PSC within 30 days after the issuance of the order. Responses … within 15 days" and "an order within 60 days" — ut-26-035-05-order-approving-settlement.txt:641, :671, :729-734 |
| Trait | What it means at the desk | Source |
|---|---|---|
| GROUNDED The redline she trusts can be wrong in both directions | The redline exists so she need not re-read thirty pages. This one invented three changes that never happened and silently hid one that did. Three sources of truth — the cover notice's list of ten corrections, the redline, and the actual diff — and no two agreed. | Re-verified by hand. Redline reads "WILDFIRE LOSS PROBABILITY DISTRIBUTION ANALYSIS ......................... 86" where original and errata clean both read "… 8" (ut-24-035-04-ellis-…-errata-redline.txt:33 vs -original.txt:31 and -errata-clean.txt:31). And "infrequent-but-extreme" (original, line 192) became "infrequent but extreme" (errata clean, line 192) with the redline showing the corrected form unmarked, and the cover notice not listing it. |
| GROUNDED Missing attachments do not appear in a diff | Twenty-seven pages of prose identical, a changed title, a changed date. A text comparison reports two trivia. The real change is a twenty-eighth page carrying the rate table — the document that sets the prices had been filed without the prices. | Re-verified by hand: original 1,217 lines, corrected 1,232. Added content: "EXHIBIT A / Schedule LLCS Initial Pricing / Customer $412.66 $412.66 / Demand ($/kW) $22.43 $10.66 / Energy ($/kWh) $0.0406 $0.0371" — mo-ET-2025-0184-nonunanimous-stipulation-corrected.txt:1217-1226, absent from -original.txt. |
| GROUNDED One real change hides inside a thousand lines of re-layout | A 378-page reissue whose only substantive edit is a plant name. A mechanical diff is useless unless it can tell an amendment from re-rendering churn. | "The corrections are found on page 95 … and changes original references to 'Brown BESS' to now read 'Cane Run BESS.'" — ky-2025-00113-kollen-correction-cover-letter.txt. Provenance note on the corrected copy: "That one change is buried in roughly 1,370 diff lines of pure re-layout noise." |
| GROUNDED The evidence that would have caught it was inside the same PDF | The body said "Brown BESS" twice. The stipulation bound into the same filing as an exhibit said "Cane Run BESS". Nobody compared the summary against the source it summarised. Twenty-one days to catch. | Body as filed: "the settlement of the Companies' requests for CPCNs for Brown 12, Mill Creek 6, and Brown BESS". Exhibit in the same document: "The Utilities will withdraw their request for the Cane Run BESS without prejudice". ky-2025-00113-kollen-direct-testimony-original.txt |
| GROUNDED Terms of art differ by one letter and mean opposites | "intraclass" filed where "interclass" was meant — within a rate class versus between them — inside a recommendation on revenue allocation. No spell-checker catches it. The same document used the right word correctly twice elsewhere, so the only cheap check available was internal consistency, and it went unmade for sixty-eight days. | "Page 20, Line 14: Replace 'intraclass' with 'interclass'." — ky-2025-00113-walmart-errata-filing.txt. Page 20: "Walmart appreciates the Company addressing intraclass subsidies." Page 21: "to further reduce the interclass subsidies." Filed 29 Aug, corrected 5 Nov. |
| GROUNDED A wrong number can look entirely plausible | A dropped leading digit: 4,248,543 MWh filed where 24,248,543 was correct — an 83 per cent understatement that nothing about the figure flags. Careful reading does not find this. Only comparison against the system of record does. | "in Figure 1, row 'Billed Sales (MWh),' and column '2024,' a digit was inadvertently omitted … The correct figure should be '24,248,543' rather than '4,248,543.'" — va-scc-PUR-2025-00058-company-blackwell-errata-direct.txt. 31 Mar to 5 May: thirty-five days. |
| GROUNDED Fixing the line is not fixing the class | A claim stated in the Summary and again in the body. The correction hedged the verb in the Summary only. The corrected document is now internally inconsistent in a way the original was not, under a cover letter claiming a complete replacement. | Corrected Summary: "how CIAC payments could impact the Network Integration Transmission Sendee". Corrected body: "how CIAC payments impact the Network Integration Transmission". Cover letter: "The clean version should replace the original version in its entirety." — va-scc-PUR-2026-00056-odec-dembek-corrected-testimony-clean-redline.txt |
| GROUNDED Errors surface when an outsider asks about a number | Five of the corrections across two errata were found not by internal review but by answering a Commission Staff data request. Each names the request that forced the second look. Self-review did not find them. | "In responding to data request STF-PIA-1-2, the Company identified updates needed to Section G.3 … should be 14 instead of 12, and … 68 instead of 61." Also STF-DEA-3-14, STF-PIA-4-15, STF-JKA-5-3, STF-JKA-3-24. — ga-56002-irp-2025-errata-2-description.txt, -errata-3-description.txt |
| GROUNDED Every correction carries a separate blast-radius claim | The fix is one obligation. Going on the record that nothing downstream moves is a second one, and it is made by assertion. Eight times across three errata, in the same words. | Repeated verbatim after each item: "The changes in these data do not affect the Company's conclusions or recommendations in the 2025 IRP." — ga-56002-irp-2025-errata-1/2/3-description.txt |
| GROUNDED "Non-substantive" is the filer's word, offered without evidence | Fourteen pages listed by number with no statement of what changed on any of them. To rely on that label she must either trust it or do the diff herself. And the same witness drew two errata covering disjoint pages, neither mentioning the other. | "certain non-substantive corrections … on pages 2-5, 33, 36, 39, 47, 50-51, 53, 58-59, and 77-78" (3 Sept) against an earlier errata for pages 22-23 (15 Aug). — va-scc-PUR-2025-00058-company-miller-rebuttal-errata-correcting.txt, -errata-p22-23.txt |
| GROUNDED A "corrected" filing can be shorter than what it corrects | The errata clean copy runs 37 pages against the original's 69 because it drops the exhibits. The authoritative record is the two documents layered, not the newer one alone. Anything that treats the latest version as the whole truth loses the exhibits. | ky-2025-00113-perry-direct-testimony-errata-clean.txt vs -original.txt |
| GROUNDED Errors come from every side, including the regulator | Five corrections in four months in one Virginia docket, from the company and from Commission Staff. And in Missouri the Commission's own order misdescribed the settlement it approved; the utility that has to operate the mechanism caught it. | "Commission Staff became aware of certain calculation errors contained in Table 2 on page 13 of Mr. Pratt's Prefiled Testimony." — va-scc-PUR-2025-00058-staff-pratt-corrected-page13.txt. "Ameren Missouri states that the Commission's summary leaves out a necessary step … The Commission agrees." — mo-ET-2025-0184-order-nunc-pro-tunc.txt |
| GROUNDED The regulator's own remedy was to stop paraphrasing and cite | Rather than write a better summary, the Commission replaced its restatement with a pointer to the authoritative text. This is Verbatim's thesis, written by a commission about its own order. For nine days the binding statement of an obligation had drifted from the thing it summarised. | Corrected paragraph: "Paragraph 46 of the Agreement and Exhibit D, attached to the Agreement, explain and provide an example of the revenue sharing mechanism…" — mo-ET-2025-0184-order-nunc-pro-tunc.txt. Order 24 Nov, motion 1 Dec, correction 3 Dec. |
| Trait | What it means at the desk | Source |
|---|---|---|
| GROUNDED Every number has a basis, and the sentence rarely says which | The plainest form of the multi-jurisdiction failure: the number was right, the jurisdiction was not, and nothing in the sentence flagged it. Total-company was used where Utah-allocated was required. | Re-verified by hand. "On a total-Company basis, PacifiCorp seeks $185.7 million, which is $82 million on a Utah-allocated basis using the System Overhead allocation factor from the 2020 Multi-State Cost Allocation Protocol." — ut-24-035-04-laconte-phase3-amended-direct-testimony-clean.txt:102-104. Cover letter: the recommendation "needed to be updated to reflect only the Utah-allocated amount." |
| GROUNDED One wrong input reaches two conclusions ten pages apart | Fixing it meant finding every place the number had been carried forward. At the second site the derived total was dropped rather than recomputed. Adding one footnote renumbered six later ones and moved the page breaks, so line 189 sits on page 10 in one version and page 11 in the other. | Body: "$112.1 million" became "$112.1 million (total Company)… Utah-allocated share is approximately 44.258% or $49.6 million." Summary: "reduced by $30.9 million to $81.2 million" became "reduced by $5.8 million." — ut-24-035-04-laconte-…-amended-…-clean.txt vs -original.txt |
| GROUNDED The docket number is not a stable key | A recurring obligation outlives the proceeding that created it. Georgia's quarterly report cites Docket 55378 in Q1 and 56002 from Q2. In North Carolina one sentence names three docket numbers for one obligation, and the regulator ordered the entire back catalogue refiled at a new address. Tracking by docket silently loses the series at the moment it moves. | "The Companies previously filed the Reports in Docket No. E-100, Sub 207 and are now refiling in Docket No. E-100, Sub 208A pursuant to the Commission's Order … Docket No. E-100, Sub 208 (March 13, 2026)" — nc-e100-sub208a-refiling-letter-spring-fall-2025.txt. Georgia: ga-55378-…-q1-2025-pd.txt vs -q3-2025-pd.txt. |
| GROUNDED The cure can land somewhere other than the disease | The Georgia correction letter names Docket No. 55378 and says the reports were filed "in the above-referenced docket". The two reports it corrects both name Docket No. 56002. The string "56002" appears zero times in the correction. | ga-55378-…-q3-q4-2025-revised-pd.txt against -q3-2025-pd.txt and -q4-2025-pd.txt |
| GROUNDED Jurisdiction is not only states | The same company answers to a state commission and to FERC for different parts of the same business. | "AES Indiana is also subject to the jurisdiction of the Federal Energy Regulatory Commission ('FERC')." — in-46258-aes-indiana-verified-petition.txt, para 6 |
| GROUNDED Two versions cannot be compared by section number | Competing stipulations thirteen days apart. The same subject sits at III.E in one and III.G in the other; III.J means exit fees in one and capacity assignment in the other. The parties could not even agree what to call the two documents. Comparison has to be by topic. | "To avoid unnecessary disputes about naming, Staff recommends using the neutral terms '10/10 Stipulation' and '10/23 Stipulation.'" — oh-24-508-EL-ATA-healey-staff-testimony-comparing-stipulations.txt |
| GROUNDED A full clause-by-clause diff is expensive enough to abandon | A Commission Staff witness whose entire assignment was comparing two stipulations twice asserted a difference existed and declined to say what it was. That is the cost of the manual diff, stated by someone paid to do it. | "includes a process for new data centers to locate in AEP Ohio's service territory, though it differs from the 10/23 Stipulation." and "though the eligibility and process are different from what is included in the 10/23 Stipulation." — same file |
These are in the persona because a persona needs them, not because a source supports them. Each says what would confirm or kill it. None may be used to justify a product decision on its own.
| Assumption | Why it is here anyway | What would confirm or kill it |
|---|---|---|
| ASSUMED Name, employer, tenure, the specific figure of five commissions | A persona needs a person. The corpus fixes the role and the ladder; it does not fix any of these. | Nothing. These are scaffolding and carry no weight. The count is set at five only because it forces the two things that are grounded — the allocation problem and the moving-docket problem. Any number above one does that. |
| ASSUMED — and it is ADR-01's own open question That the analyst, not counsel, does the reading | ADR-01 says plainly this persona is a hypothesis until real interviews happen. This document cannot settle it and does not try. | The corpus cuts both ways and that is the honest report. For: a Director of Regulatory Affairs signs the Georgia correction himself, in his own name, with his own email. Against: the Utah amendment cover letter is filed by Holland & Hart LLP; the Indiana joint motion is signed by Taft Stettinius & Hollister LLP. Counsel is on the paper constantly. Killed by two of three interviews saying the real reading is done by counsel — in which case ADR-01 is rewritten and the PRD with it. |
| ASSUMED What is on her screen — Excel, Outlook, a shared drive, some tracker | The corpus proves .xlsx workpapers with disagreeing tabs exist ("Hydro_Modernization_Budgets_TRADE SECRET.xlsx", "Capital_Details" tab) and that service runs by email. It does not prove what she personally opens, or what her company's system of record is. | One question in one real interview: "What did you write down, and where does it live now?" If the answer is a spreadsheet, the PRD has to say so. |
| ASSUMED How long the manual diff takes her | Needed to size the claim. The corpus gives only an indirect signal — a Staff witness paid to compare two stipulations gave up twice — which is an inference about cost, not a measurement of time. | Ask about the last one specifically, not in general. A number from one real person beats this row entirely. |
| ASSUMED Her tone, her patience, what she would say out loud in a meeting | The agent has to speak. Filings are written to be filed; nobody writes their irritation into a certificate of service. | Twenty minutes on a call. Until then the voice in .claude/agents/regulatory-analyst.md is a costume over grounded facts, and it is deliberately made blunt so it disagrees rather than flatters. |
| ASSUMED That the obligation owner is a distinct person who approves her work | ADR-01 names this as a load-bearing second role and a security control. The corpus shows verification pages and named signers; it does not show any company's internal approval chain, because internal process is not filed. | Ask who had to look at it before anything happened. The real routing graph, not the org chart. |
Be generous here rather than defensive. The list of what a synthetic user cannot answer is longer than the list of what it can, and pretending otherwise is how a hypothesis gets promoted to a finding.
Whether the real reader is the analyst or counsel is ADR-01's own open question, and this document does not close it. Worse, the method is biased against ever closing it: the corpus is a record of who files. A prior pass over 39 outreach targets found only 13 utility-side and matching the ICP — 5 commission staff, 9 consumer advocates, 10 outside counsel or consultants. Docket mining finds people who file; the ICP mostly reads, and often never appears on a service list at all. Treat those 26 as adjacent voices, not as the user. Only real interviews decide this.
Not what a seat costs, not what a budget line looks like, not who signs a purchase order, not whether this is bought by the department or by IT. There is nothing in a public filing about any of it, and a number invented here would end up in the MRD and be indefensible at the panel.
Willingness to pay, switching cost, and what tool already sits in the workflow are all unobservable from the record. H2 in the PRD — that nobody reliably diffs versions today — is exactly the kind of claim this persona must not be used to confirm. If she already has a redlining tool she trusts, the corpus would not show it. It shows that redlines get filed and that at least one of them was wrong in both directions; it does not show what she uses to make one.
The corpus was assembled by hunting for errata and version pairs. Every rate on this page would be inflated. "How often does this actually bite you" is a question for a person.
The public record is the tail end of the work. Everything upstream — the meeting where materiality was argued, the mail that routed it, the deadline that was nearly missed and was not, the pain that never became a document — is invisible here. If the largest pain in the job never generates a filing, this method cannot find it, by construction.
This is the failure that matters most. The persona was built by the same mind that built Verbatim, from a corpus chosen by that mind, and it will produce pains that Verbatim already claims to solve, because that is what it was assembled from. The evidence table above is the defence — every trait traces to a filing rather than to a prior — but the defence is partial, because the selection of which filings to read was not blind. Treat agreement from this persona as worth nothing. Treat disagreement as mildly interesting. Treat a real user's five minutes as worth more than all of it.
The consultable form lives at .claude/agents/regulatory-analyst.md. It is built to be blunt, to say "that is not my problem" and "we already do that in Excel and it is fine", and to separate what it knows from what it is guessing in every answer. A synthetic user that likes everything is worse than none, because it ratifies every bad idea put to it.
Three rules on its output:
docs/user-research.html. That file is for real conversations with real people, and this is the one kind of evidence this project cannot fake.