We rewrote one question. Jev went from missing every household link to finding all of them.

An Ohio entity-resolution study: the cost of typed decisions, the cost of reconsidering them, and the clocks that distinguish a fast model from a fast workflow.

Jev alone handled the first pass for a few cents. A small, signal-selected review queue bought more agreement; a larger queue did not eliminate the remaining disagreement. The latest breakdown puts the bill beside output coverage and separates a measured elapsed clock from the sum of time spent in individual calls.

This experiment uses the same second panel as the earlier comparison below. The Jev-only candidate and family-review ranking were frozen before this research pass inspected its outcomes; the Fable and cheap-reasoning controls are retained runs, not new calls. Every agreement figure is against the model-built reference, not human-established truth.

The latest cost and speed breakdown

The earlier experiments explain how we got here. On the first panel, the gated pipeline missed nine of fourteen household links and cost more than the cheap control. We published that result. The next comparison changed the wording of the identity and household questions, not the model or the thresholds. The later Jev-only experiment then asked how much reasoning we could remove. The sections below keep those designs separate.

Inspect the latest bill and clocks, read the case history, compare the earlier gated programs, or follow seven rows through the join.

The rows are not the donors

Ohio's campaign-finance files record reported contributions: a name, an address, an amount, a date, a receiving committee. Our substrate is the candidate-committee and PAC filings for report years 2022 and 2024 — 883,012 rows. To total contributions by donor you must decide which rows belong to the same person, and the strings will not tell you. The names in this article are altered; every field pattern is a real one from the panels. DANIEL R. OKAFOR and DANIEL OKAFOR JR at one address are one man with an inconsistent clerk, or a father and son at the same real-estate firm. WALTER H. BRENNAN SR. and WANDA BRENNAN at one address are two people and one household. A county party's "judicial fund" is an organization until a registry says it is a committee.

That is entity resolution. A false merge adds one donor's money to another's total; a false split scatters one donor across several. Splink and Dedupe are the established probabilistic approaches, and Dedupe's own tutorial uses campaign contributions. We did not benchmark them. Our question was narrower: when the deterministic work is done and a judgment call remains, how much of it can a cheap typed model make, and how do you find out when it is wrong?

The earlier three-arm experiment

We first compared these three programs on the same original evidence, with the same semantic output opportunities and legitimate uncertainty allowed. This is the gated architecture that preceded the budgeted family-review experiment above.

You should know how the reference was made. It is the accepted D28 resolution set: a swarm of cheaper models proposed each case and accepted it on agreement, with Luna proposing on every case and DeepSeek adjudicating most. So the cheap control is graded partly against its own handwriting, and Fable is the only arm that is not. The reference never abstains, so an "unknown" always scores as a miss. Every agreement number here is agreement with that reference, not with a human.

The first panel

Panel 1 was 204 cases in 200 families, frozen before any model ran. Here is what we published on September 17, and what happened when we re-read the retained answers.

The three original rows are the frozen first comparison. The revised row re-ran only the Jev-first arm after the question criteria were rewritten using that panel's own retained answers, so it is a development result, not a held-out score. Panel 2 is the held-out test of the revised design.

The headline had been "200/200 on the core task." The audit said something less flattering. Of the 132 cases Jev handled alone, 123 had one donor in them; the fast path was mostly labelling a single row person. On identity, Jev gave a definite answer for 54 of 87 pairs and 53 agreed with the reference — but the 0.70 gate admitted 13, and Luna did the rest. On households, Jev was asked ten times and answered no ten times, nine at 0.90 or above; five of those nine pairs had byte-identical street addresses and the same surname. Every household link the pipeline did find came from Luna or Gemini. And the three arms were within noise on core (197, 198 and 200 of 200): Fable's three misses were unknown on two suffix conflicts and committee for the judicial fund, calls a careful analyst might make the same way.

None of this was hidden by the numbers. It was hidden by reading them as a scoreboard instead of a receipt.

The fix: say what suffices

We had written the household criteria the way you write a warning. A shared employer or mailing address is not by itself a household. True, and useless: it told Jev what fails and never what passes. The reference's own reasoning was explicit — a shared residential street address plus a surname is household evidence — and we had never put that sentence in front of the model. The identity criteria had the same defect. Of Jev's 33 unknown identity answers, 27 were pairs with the same first, middle and last name at the same street address or postal code — some differing in employer, some in nothing but punctuation. The reference merges every one of them. Jev sat at fifty-fifty because nobody had told it that was enough.

So we rewrote two questions. Not the thresholds. Not the models. The words.

Then we asked Jev the same 87 identity and 13 household questions again, on the same evidence, and compared.

The one identity disagreement under v2 was KAREN L. against a first-name field holding KAREN L, at p = 0.60, below the gate. It went to Luna, which is where it belonged.

Two false starts are worth your time. Our first probe passed the reference resolution into the model's state as the "upstream proposal," and Jev said yes 13 of 13 times — under the old wording. We had leaked the answer: first the relations, and when we stripped those, the reference's prose explanation. Only when the proposal was composed by the pipeline's own code, with the generic basis the real run uses, did v1 reproduce the failure exactly. The lesson is not about our probe. Jev reads everything in state as evidence, including narrative you labelled untrusted. Put a reviewer's summary in the state and you are asking the model whether the reviewer sounded sure.

With the words fixed and nothing else touched, we re-ran the Jev-first arm on panel 1. It matched 200 of 200 families on the full contract at $0.24 with nothing unpriced, and 158 of 204 cases never saw a reasoning model. That number is real, but it is a development result — the panel's own answers chose the wording. For the gated design, the held-out test was panel 2.

The second panel: the earlier gated comparison

Panel 2 is 205 cases in 200 families drawn from the same frozen population with a new seed, after excluding every atom, family and source case in panel 1. The Jev-first design was frozen with the v2 wording before any outcome. Fable and the cheap control ran unchanged.

Jev handled 165 cases alone — 122 with one donor and 43 with two or more — and its identity gate admitted 116 of 118 pairs, every one agreeing with the reference. It found eleven of the twelve household links: nine straight from Jev at the gate, two from Luna after one Jev request exceeded the model's token limit. The programs differed by at most two full-family matches. The bill and the recorded batch-service clocks describe a larger operational difference; they do not establish a provider-latency guarantee.

The two families Jev-first missed are the honest part of the story, because in both Jev had the reference's answer and the gate did not trust it. M LOUISE HARGREAVE and MARCUS J. HARGREAVE are two people at one address; Jev said different at 0.65, under the 0.70 gate, and the Luna fallback merged them — taking the household link with it. A county party organization drew organization from Jev at 0.62, under the 0.80 gate, and the fallback chain relabelled it. The safety net was the error both times. That is not an argument for lowering gates on the strength of two families. It is a reminder that the fallback model is not an oracle either, and that the retained answers are what let you see which one was wrong.

Two boundaries on this result. By the time we drew panel 2 the fresh population had run out of committee, joint, placeholder and registry-mapping families, so this panel tests typing, identity and households and nothing rarer; the twelve household links are its relationship test. And dispatch paused for twelve failed calls — eleven Luna connection errors, absorbed by Gemini adjudication or a later fallback, and the one oversized Jev request. Each pause was inspected and the identical frozen plan resumed with its completed batches; the failed calls kept their held commitments and are not counted as free.

How the earlier gated program is built

Ordinary code keeps the original rows and their identifiers, parses the file-specific field meanings (PAC_REG_NO is the donor's registration in CAN_CON files and the recipient's in PAC_CON files), and checks that every input survives every transformation. None of that is model work.

Jev gets a complete source view — up to four cases, under 64 kB — and up to 32 questions at once. Each question names its evidence by path (cases[0].observations[1]) and defines its answers in the criteria. What comes back is a label and a probability for every option. Code accepts the label if its probability clears the kind's gate — 0.80 for type, 0.75 for coherence, 0.70 for identity, 0.90 for registry and links — and composes the case: types, identity clusters, then links. Anything that fails a gate, or composes inconsistently, is batched as a short list of questions for a Luna hank; anything that still will not compose goes whole to a Gemini hank.

Hankweave runs those reasoning hanks: a saved program whose steps host model sessions, with the prompt, the response, the session log and the accounting retained per attempt. That retention is what made the audit above possible, what let us recover two Gemini answers from a Markdown fence without buying them again, and what let eleven Luna connection errors fall through to Gemini and keep every batch valid instead of killing the run.

The runnable version of the join is small enough to read in one sitting. Seven synthetic rows, six questions, seven rows back:

rows.map(row => {
  const unit = byLabel.get(row.source_label);
  const judgment = answers.get(unit.evidence_unit_id);
  return { ...row, evidence_unit_id: unit.evidence_unit_id,
           judgment, decision: route(judgment) };
});

The earlier program's bill, and the whole study's spending

This receipt belongs to the earlier gated comparison, not the new review prefixes. Its Jev line covers 715 questions across 65 calls on panel 2. The new fast pass asks its own questions and has its own bill above. Which model dominates the total depends on the amount and size of the reasoning work; cheap judgments do not make an exception queue free.

Scaling the earlier comparison

The calculator below retains the earlier three-arm measurements. It does not substitute the sparse-review agreement result, and it does not invent an elapsed clock for a smaller review prefix. The latest breakdown above instead keeps the measured calls, actual stage clocks and missing timing measurements visible.

The full Ohio snapshot has 883,012 contribution rows and 283,349 strict evidence bundles. The calculator scales the measured panel-2 workload to that population under assumptions you can change; it is arithmetic, not a run, and panel 2 is more ordinary than the state. The larger claim is structural: repeated bytes should stop becoming repeated model work. If ordinary code can reduce a much larger source to a million useful 2,000-token questions, Jev's published input price implies $84 for the judgments. Nothing here establishes that such a reduction preserves a task; preparation, referrals, storage and checking each carry their own bill.

What we would tell you to do

Put the exact work in code and let it own the identifiers. Give the model the evidence, not a summary of it. Write each question's criteria to state what suffices under the standard your consumer will accept, then measure each kind of question separately — a gate that was right for typing was wrong for households in the same pipeline. Keep every attempt, so that when the model is confidently wrong you can distinguish the question, the state, the parser and the model. The later experiment adds a budget decision: make the review queue explicit, count unsuccessful reviews, and do not assume a reasoning model can only improve an answer.

Start with the example. Its dry run needs no key. Read one question, one returned probability and one join before deciding which part of your pipeline should think.

Sources and accompanying material

1. Ohio campaign-finance search and downloads (archived source). The held substrate is the four 2022/2024 report-year candidate/PAC contribution files.
2. Splink and Dedupe: background approaches, not measured arms.
3. TypeSafe Choice, state, confidence and model pricing; Hankweave. Prices are the retained September 17, 2026 record.
4. Methods — reference provenance, both panel draws, the original gates and the later family-review policy; earlier comparison aggregates; cost, clock, byte and corpus ledger, including the latest research and source hashes; earlier anonymous family outcomes.
5. Runnable source, guide and recorded synthetic output. Raw donor records, private model logs and benchmark labels are not publication material.