Once you know the machine gives confident wrong answers, you have two choices: fight it forever, or build something that can't. Lester stops arguing with AI and starts turning his verification ritual into a spreadsheet — and the morning a consultant's flood number lands on his desk, the tool flags it before he's even opened the file. The anti-GFAS protocol becomes a tool.
8:07 AM, Sarah Morrison at his door: a bridge-replacement submission needs hydraulics signed off by Friday so the right-of-way timeline can move. Lester looks up from the legal pad. "The Q100 is wrong." Sarah points out he walked in thirty seconds ago and hasn't opened the report. He knows — but 4,200 cubic feet per second for a watershed that size in East Tennessee feels wrong, and his bet is the consultant applied the wrong regression region. Give him the day.
Every review pulls the same information from the same sources in the same order — regression flows, the FEMA lookup, the model report, backwater rise, structure screening, QC against the standards — and every time he'd started from a blank sheet. So back in January, right after the Pigeon River report, he started building. A dashboard hub. A flow-input tab that auto-selects the governing discharge and flags divergence. And because he needed published flood values without retyping tables out of PDFs, he built the database, then the parser — and discovered eight East Tennessee counties used a new seven-column table format nothing else handled correctly. He fixed that too.
By 11:30 he had four findings. The region misclassification was the big one; the rest were supporting evidence. The backwater tab ran the rise equation independently and got 0.21 feet against the consultant's reported 0.08 — still under threshold, but a mismatch that usually means a boundary condition is set wrong. The velocity check flagged an outlet velocity past the standard with no outlet-protection analysis mentioned. The pressure-flow check showed the structure operating in pressure flow at the design storm — while the summary claimed "open channel conditions throughout." Four cells: yellow, yellow, red, red. He wrote each on the pad, rewrote it cleaner, then typed the memo.
She'd always known he was careful — the frustrating kind, the kind deadlines bent around. But this was careful and fast, and she asked to see the tool. Enter a county once and the whole thing cascades: the right regression region, the published flows queried from the database, the flood-map panel and effective date, all populated. Thirteen QC sections comparing every reported value against what the geometry and hydrology should produce. "How many cells total?" she asked. Three thousand three hundred fifty-two formulas. It handles culverts too — auto-generates every viable configuration and screens them against the standards — plus a scour module, a haul-road calculator he spun off standalone, and a deck-drain calculator that handles simple crown, multi-crown, SAG bridges, and superelevation transitions.
"I've been trying to get consultants to submit drainage like this for two years," Sarah said. "You just built the standard they should be submitting to." She wanted to take it to headquarters. "It's not finished," Lester said. "Is it ever going to be?" He thought. "Probably not." "Then it's ready."
The dam-inventory piece started as a side thought — regulated flow changes everything, and the national database took forty-five minutes by hand to extract one answer. Lester wrote a script that pulls the state's dams, filters valid coordinates, and sorts by storage so the big reservoirs surface first; it runs in under ten seconds. Then he pulled the published regression envelope — the ranges of drainage area, slope, and imperviousness each equation is actually valid within — so that any basin falling outside the range gets flagged as an extrapolation instead of trusted as an answer. Same principle as the pad, same principle as the gauges: don't trust confident answers · verify against the envelope · flag anything outside it. When he showed Earl the sorted dam list, Earl said the government's most boring database finally worked. "It always worked," Lester said. "It just required forty-five minutes to extract one answer." "That's not working," Earl said. "That's suffering."
Three weeks later the consultant's letter came back: region corrected, new Q100 within engineering judgment of Lester's number; pressure flow documented; outlet protection added; backwater rise fixed. Thirteen QC sections, twelve green and one minor note. In the old workflow a full re-review of a revised submission took most of a day — you rebuilt every comparison from scratch. The SUPERSHEET just needed the new numbers dropped in. Sarah brought up the national conference: the QC sheet, the FIS parser, the dam integration — taken together, that's a paper. Lester looked at his pad. He knew the thesis already: the same verification discipline his dyslexia forced him to build for his own brain is exactly what AI-assisted review requires. The failure mode is identical. The tool isn't just a QC sheet. It's an anti-GFAS protocol for an entire profession.
The fountain drink was foamy again — of course it was. Someone swapped the box without bleeding the line. Systems don't stay fixed; they require maintenance, and the fix that held for three weeks needs the same attention on week four. Not a revelation. Just Tuesday. He drank it anyway and opened a fresh note, and under a list of paper titles he wrote the line that had been forming since Sarah said it's ready:
Some things you fix. Some things you work around. The skill is knowing which is which.
The verification crew
The tools, out in the world