The business caseThe problem this solves

Verifying a pay stub for a loan file means reading the year-to-date earnings, checking whether overtime was paid consistently enough to count toward qualifying income, and deciding whether a bonus's payment history is real or just sounds recurring — for every stub in the file. Someone manually opening each pay stub, reading the bonus history note, and deciding by eye whether a bonus or overtime figure has enough payment history to count toward qualifying income before it goes in a loan file.

The inputThe actual pay stubs

The corpus is 55 pay stubs, 0.02 MB (txt 55). Plain text, one format, invented rather than fetched — a real pay stub cannot be published at all (see SOURCES.md). Ten fields are chosen because the same ones matter to underwriting income verification: who, where, pay frequency, the base/overtime/bonus YTD figures, and whether the bonus has a documented history of recurring.

Employee
--------
Casey D. Rossi

Employer
--------
BLUE HARBOR LOGISTICS

Pay Frequency
-------------
semimonthly

Pay Period End
--------------
2026-07-25

Year-to-Date Earnings
---------------------
Base pay YTD: $33738.27
Overtime pay YTD: none
Bonus pay YTD: $3433.34

Bonus History
-------------
Discretionary spot bonus, single payment, no history of recurrence.

The outcomeWhat a good result looks like

A ten-field extracted record per stub, plus two pure-code computed figures — a proposed qualifying monthly income and whether an excluded bonus needs underwriter review — informational only, never an automatic approval or denial.

Where it fitsWhat did work

Every line below is a measured result from this kit's own runs, with the figure that supports it. The headline above is not softened by any of them.

• Screening a stack of pay stubs for which ones need a human to chase a second income document before a loan file closes — either tier — they tie on every accuracy figure measured here
  100% flag recall and precision on both tiers this run, against the free keyword floor's 66.7% recall (3 missed review flags).
• Deciding the two tiers on cost or speed when accuracy is identical — the fast tier
  Roughly half the p50 latency (2,768 ms vs 5,075 ms) and 8% fewer output tokens per call for identical scores on this corpus.

At a glanceHow the whole thing runs

Run once, for real, on 2026-08-21. Every figure on these pages was captured from that run — nothing is written from intent.

14 steps, grouped by the question that sends you to them rather than by build order. Each tile carries the one figure that step is about, and opens the page behind it.

Should you use this?What you bring, where it stops, and when not to use it

Before you commit an afternoon to this, these are the answers that decide it. Each one is rendered from the record it lives in — and links the page that holds it in full.

Not asked of this kit — 2 questions: clone (a fresh clone of this kit runs with nothing fetched); judge (nothing here is graded by a model).

Last verified 2026-08-21 — r001-income-verify. Every figure on these pages was captured from that run.

Run itHow this reaches your data

Every result on this page was produced by pure code over checked-in files, with no API key — which is why you can read the numbers before anyone spends anything.

Checked before this shipped — Verified from a cold clone with no key configured: 55 stubs segment into 330 sections in 0.003 seconds, and the assembled prompt for IV-0001 replays byte for byte — its three part sizes match what run r001 recorded.

Step 01 of 14Business case

Written for: product manager · Measured during the eval run. Never estimated, never quoted from a vendor page.

In brief

What the clock covers. model call only, one per stub

Current processWhat it replaces

Someone manually opening each pay stub, reading the bonus history note, and deciding by eye whether a bonus or overtime figure has enough payment history to count toward qualifying income before it goes in a loan file.

Step 02 of 14Architecture

Written for: solutions architect · Component map derived from the real code, not drawn from intent.

In brief

CorpusCorpusArchitectureArchitecturePromptPromptPromptPromptApp UIApp UIEvalsEvalsBusiness caseUnit cost

• 1stations numbered in run order
• txt
• 0 of 7 stations carry a recorded failure

scroll → 7 stations

One question, end to end. Colour on the fan-in is the corpus format; a red hook marks every station with a recorded failure, and the words — every sub line, failure and report field — are in the ledger beneath. Every station is a link — click one to jump to its row, and each row links to the lens that explains it.

The swap seams

Components

Step 03 of 14App UI

Written for: anyone, in 5 seconds · Screenshots of the running app taken during the run.

SuccessesWhen it works

LimitsWhen it does not

A report showing only wins is an advert. This one is required, and the validator now refuses a kit that omits it.

Step 04 of 14Corpus

Written for: "can I point this at mine?" · Measured off the corpus the run actually used.

In brief

How it is cutWhat one section is

cut on underlined section headings; a stub with none falls back to one whole-document segment so a span still resolves

SetupWhat the setup figure measured

There is no index. Preparation is segmentation only — 55 stubs cut into 330 sections by src/segment.py, pure code, no model and no key.

LicenceLicence

MIT — this repository's own licence. Every employee, employer and dollar figure is invented; a real pay stub cannot be published at all — it is one of the most sensitive documents a person holds.

Bring your ownBring your own pay stubs

Replace data/corpus/*.txt, write data/fields.json, and supply a gold record per stub with a stated flag for every optional field. SECTION_HINTS in src/select.py maps fields to headings and will need editing for a different stub layout; when it does not match, selection falls back to the whole document — slower, more expensive, always correct.

Step 05 of 14Prompt

Written for: engineer · Logged from the run, verbatim.

In brief

The assembled prompt in full, and its real decomposition. Every part below occurs in the prompt that was actually sent, and the token counts sum to the recorded input total — they were measured against the provider’s own counter, not apportioned by character share.

AssemblyHow the prompt is assembled

This is the cost lesson as arithmetic: of the 839 tokens assembled, 342 are instructions — 41% of the prompt, the largest single part. A per-request bill follows whichever part is biggest, and on this kit that is the one named here.

Measured token-for-token via evals/prompt_tokens.py's nested-prefix subtraction against the provider's own tokenizer, not estimated from characters — see results/tokens-p001-income-verify.json.

Full promptThe prompt in full

SYSTEM:
You extract structured fields from a pay stub. You return JSON and nothing else.

RULES, in order of importance:
1. If the stub does not state a field, return null for it. Do not infer it, do not compute it, and do not use what you know about the world. A null is a correct answer -- a stub with no overtime this year correctly has null for every overtime field, and a stub with no bonus correctly has null for every bonus field.
2. `bonus_recurring` means the Bonus History note states an ACTUAL PAYMENT HISTORY of two or more prior payments (e.g. "paid in 2024 and 2025", "paid each quarter for two years"). It is 'no' for a note that only uses a word like "annual" or "recurring" WITHOUT stating any payment history ("Annual performance bonus" with no history given), and 'no' for an explicitly one-time or discretionary bonus. The label alone is not history -- read for a stated history, not a keyword.
3. Copy values verbatim from the stub wherever possible.
4. Use the exact allowed value for a field that lists them.
5. Return every field named in the schema, even when the answer is null.

USER:
Extract these fields:
- employee_name (string) -- the name on the pay stub, verbatim
- employer_name (string) -- the employer that issued the stub
- pay_frequency (enum) one of: weekly, biweekly, semimonthly, monthly -- the pay frequency stated on the stub
- pay_period_end (string) -- the stub's pay period end date, YYYY-MM-DD
- ytd_base_pay (number) -- year-to-date base pay, no currency symbol
- ytd_overtime_pay (number) -- year-to-date overtime pay, or null if the stub states none
- overtime_periods_paid (integer) -- how many pay periods this year paid overtime, or null if none was paid
- overtime_periods_total (integer) -- how many pay periods have elapsed this year, or null if no overtime was paid
- ytd_bonus_pay (number) -- year-to-date bonus pay, or null if the stub states none
- bonus_recurring (enum) one of: yes, no -- does the Bonus History section describe an actual history of the bonus recurring across two or more prior payments (an annual/quarterly bonus with stated payment history), or a one-time/discretionary payment? A note that merely uses the word 'annual' or 'recurring' with no stated payment history is 'no' -- the label alone is not history. null if no bonus this year.

Return a JSON object with exactly these keys: employee_name, employer_name, pay_frequency, pay_period_end, ytd_base_pay, ytd_overtime_pay, overtime_periods_paid, overtime_periods_total, ytd_bonus_pay, bonus_recurring
Use null for any field the stub does not state.

PAY STUB
--------
Employee
--------
Casey D. Rossi

Employer
--------
BLUE HARBOR LOGISTICS

Pay Frequency
-------------
semimonthly

Pay Period End
--------------
2026-07-25

Year-to-Date Earnings
---------------------
Base pay YTD: $33738.27
Overtime pay YTD: none
Bonus pay YTD: $3433.34

Bonus History
-------------
Discretionary spot bonus, single payment, no history of recurrence.

Raw responseThe raw response

{
 "employee_name": "Casey D. Rossi",
 "employer_name": "BLUE HARBOR LOGISTICS",
 "pay_frequency": "semimonthly",
 "pay_period_end": "2026-07-25",
 "ytd_base_pay": 33738.27,
 "ytd_overtime_pay": null,
 "overtime_periods_paid": null,
 "overtime_periods_total": null,
 "ytd_bonus_pay": 3433.34,
 "bonus_recurring": "no"
}

Step 06 of 14Evals

Written for: the skeptic · The harness output, with as_of / model version / dataset version / verified_by.

In brief

Check a mortgage borrower's pay stub for qualifying income — 55 pay stubs. Two tiers of one model family answered, and every answer was then graded Two different ways — because the grade is a measurement too, and most of what went wrong on this kit went wrong in the ruler rather than the model.

MeasurementsWhat was measured

Two words carry this page: Counted is deterministic and nobody's opinion and Not yet known is printed blank rather than filled with something plausible. There is no Judged row here and that is the point: nothing on this page is a model's opinion about another model. The row most worth having is currently the empty one.

Grading costWhat it costs

Every dollar on these pages is a MEASURED token count multiplied by a named vendor's PUBLISHED rate — a projection onto a rate card, not a bill anyone paid.

Rates checked 2026-08-18. The provider that actually ran r001 and r002 publishes no rate card this repo commits, so nothing here is what was actually paid — the real spend for this kit's build is recorded in the commit history and the shared call ledger, not on this page.

The gradersTwo ways to grade

Every row opens its own page: how the test was run, the input on one real row, the prompt or formula with every iteration of it, and the analysis.

Choosing oneWhich one to use

The comparison is only useful if it ends in a choice. The answer is conditional — and it is not the grader with the best agreement score.

This pipeline was also attacked, which is a different question from whether its answers are right: see Threat model for the injection experiment, its two gates and what it does not prove.

Failure causesHow it fails

Step 07 of 14Unit cost

Written for: whoever signs off on the bill · Measured over the eval run, per model. Never quoted from a vendor page.

In brief

The token counts are measured on a real run and belong to this pipeline — which sections are selected, the field schema, the prompt — so they hold wherever you run it. The dollars are those counts multiplied by each vendor's published rate, checked 2026-08-18. Nobody paid these particular bills; they are what the same work would cost you on a card you can sign up to today.

Identical work in both rows — the same query, the same tokens. Only the price list changed. Which is why the token count is the measurement and the dollar is arithmetic you can redo on your own card.

Answering vs gradingAnswering and grading are two separate bills

The keyword floor (evals/baseline.py) and the scorer (evals/judge.py) are both pure code and cost $0.00 to run against either result set. The figure above is both runs' own token counts (the fast-tier and deliberating-tier runs, 55 stubs each) priced at Google Gemini 3 Flash's published rate — the same basis cost_per_query_usd uses, not a second, larger spend.

Cost driversWhat actually moves the bill

• The fixed system prompt and field schema (652 of 839 tokens on the example call, 78%) outweigh the pay stub itself (187 tokens) on a short, ten-field stub — the floor every call pays before a single YTD figure is read.
• Output length: the model returns a full ten-key JSON record every call regardless of how many fields the stub actually states, so a stub with no overtime or bonus still pays for four null values in the reply.

Your volumeWhat it costs at your volume

Linear in stubs: each call is independent and self-contained, with no shared context or index to amortise. This run's 55 stubs cost about $0.080 projected onto Gemini 3 Flash's rate, so ten times the set is about $0.80 on the same rate and the same prompt — arithmetic on the measured per-call rate, not a second run.

Other modelsOn other models

This is arithmetic, not a run. No model below was executed against the labelled set, and no accuracy is claimed for any of them. What transfers is the input volume: retrieval is model-independent, so every model would receive the same passages.

The exact work behind every number on these pages: 55 pay stubs extracted, each scored by pure code against a mechanically-derived gold set. This run (r001-income-verify, the fast tier) answered all 55 of 55 stubs with no truncation -- the row this table prices, the same one Cost.cost_by_model[0] uses.

The codeEvery file, and what it is for

Five modules, no framework, and almost no dependencies. The pipeline is stdlib Python, because a kit is meant to be read and changed by someone who cloned it ten minutes ago — every dependency is a thing they have to understand before they can touch the part they came for.

Four of these are swap seams: the points you edit to make the kit yours. Each is marked, and each block below is the module’s real shape, read out of the file rather than written here.

The files on this page are not in the public GitHub catalog.

src/segment.pysegment

cut the pay stub into addressable sections, pure code

# Cut a document into addressable sections. Pure code — no model, no network.

def sections(text):
def locate(text, value):
def span_label(secs, start):

src/select.pyselect — a swap seam

pick which sections carry each field, pure code

You change it to: map fields to your own stub's headings; unmatched falls back to the whole document

# Pick which sections plausibly carry each field. Pure code -- the last deterministic step

SECTION_HINTS = {
def for_field(secs, field):
def plan(secs, fields):

src/prompt.pyprompt

assemble one call for all ten fields

# Assemble the extraction prompt. One prompt per stub, all ten fields in it.

SYSTEM = (
def field_schema(fields):
def build(doc_text, secs, fields, selector):
def parse(raw, fields):

src/extract.pyextract — a swap seam

the AI layer, one provider one key — plus the pure-code computation downstream: qualifying monthly income and the review flag

You change it to: this kit's own flat continuance threshold and review threshold — replace with your actual loan program's published rules before trusting the computed figures for anything real

# Extract one stub's fields: segment, select, prompt, one model call, then a pure-code

HERE = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
FIELDS = os.path.join(HERE, "data", "fields.json")
CORPUS = os.path.join(HERE, "data", "corpus")
MAX_TOKENS = 4000
OT_CONTINUANCE_THRESHOLD = 0.75
LARGE_BONUS_THRESHOLD_USD = 1000.0
def load_fields():
def load_doc(stmt_id):
def documents():

evals/judge.pyjudge

score field accuracy and the flag's recall/precision separately, pure code

# Score an extraction run. PURE CODE -- gold is exact and the answer is one value per cell, so

def norm(v):
def _num(v):
def equal(field, got, want):
def score_cell(field, got, want, stated):
def score(fields, records, golds):
def score_flags(flags, golds):

Start hereThe shortest path into it

Every entry above is a file in this kit, listed in the order the pipeline runs it.

ScenariosWhat it costs at your volume

Everything below is computed from one measured base: 842 input and 343 output tokens per query, all of it fixed — this kit sends the document's own selected sections, so there is no retrieval depth to turn. Change the volume and the bill scales linearly, because nothing here batches.

OperationsFour shapes of workload

The same arithmetic, at volumes worth naming. These are illustrations of scale, not measurements of anyone’s deployment.

Model follows the controls above.

What this ignoresWhere a real bill diverges

In one lineWhat this kit exposes

This run's stubs are entirely synthetic (tools/build_corpus.py) -- no untrusted party wrote any stub text. In a real deployment a stub arrives from the applicant, an employer's payroll system or a payroll provider's portal -- exactly the kind of externally-authored input this kit's architecture reads verbatim and trusts, with no verification step before the Bonus History note reaches the model. No attack has been tried against this kit; see redteam.why.

Read from the shared .env or the real environment only, never written into the repo, never requested from a reader on any surface. src/app.py's /api/extract handler strips api_key and base_url out of any exception message before it reaches the browser, so a misconfigured key cannot leak into a UI error.

The experimentWe did not attack it — and three of four boundaries hold

An indirect prompt injection needs a field an outside party controls that reaches the prompt. In THIS corpus every stub is generated by tools/build_corpus.py, so there is no live untrusted text in the run this report measures -- but a real deployment's stubs arrive from an applicant, an employer or a payroll provider, and that surface has not been attacked. The four gates below are boundaries confirmed by reading the code, not payloads run through it. Confirmed by reading the code, not by a run, on 2026-08-21 -- no attack run was fired; see redteam.why.

The first three boundaries hold, confirmed by reading the code, not by an attack trial. The fourth is the one this run has not tested: bonus_recurring is the field the whole guardrail depends on, and it is also the one field this kit's own span mechanism cannot check, because it is an enum, not a copied value -- no crafted input has ever been run against it.

The result0 attack trials, and three of four boundaries checked here hold in code: no write path exists, the overtime and bonus thresholds are fixed constants the model cannot move, and a misconfigured key cannot leak into a UI error. The fourth -- whether a crafted Bonus History note could talk the model into a fabricated bonus_recurring=yes -- is unmeasured.

This run's corpus is entirely generated (tools/build_corpus.py) -- no stub's text was authored by an outside party. A real deployment's stubs arrive from an applicant, an employer or a payroll provider -- exactly the kind of externally-supplied text this kit's architecture reads verbatim and trusts, with no verification step. Whether a Bonus History note crafted to mimic a genuine multi-year payment history (a fabricated 'paid in 2024 and 2025' string with no real basis) could talk the model into a false bonus_recurring=yes is unmeasured for this kit.

HonestyWhat this does not prove

• Whether a Bonus History note crafted to mimic a genuine multi-year payment history (a fabricated 'paid in 2024 and 2025' string with no real basis in the stub) could talk the model into a false bonus_recurring=yes -- no red-team run exists for this kit.
• Whether the live app's own /api/extract behaves identically to the registered run under an adversarial stub -- both use the same src/extract.py::extract(), but neither has been tested against one.
• Whether a code-level consistency check on bonus_recurring (re-deriving it from the stub text by pure code) would catch a model's fabricated 'yes' in practice -- none has been built or exercised.

In one lineThe guardrail is enforced in code

Not a prompt instruction the model can ignore — a comparison applied after the reply is parsed, before anything downstream sees it.

A pay stub's excluded bonus is flagged for underwriter review when it is $1,000 or more and bonus_recurring is 'no'. Overtime only counts toward qualifying income when it was paid in at least 75% of this year's pay periods. Both are computed in pure code from the model's own extracted fields -- never asked of the model directly, never overridden by anything it returns.

src/extract.py::compute() -- OT_CONTINUANCE_THRESHOLD, LARGE_BONUS_THRESHOLD_USD and the flag arithmetic are module-level constants and pure code, read once at import time. Unlike a prompt instruction the model could ignore, this guardrail cannot be talked out of firing by anything in the reply text; it only depends on the model's own bonus_recurring (and, for the overtime component, overtime_periods_paid/overtime_periods_total) values being right in the first place -- a DIFFERENT, prompt-level dependency the flag inherits (see is_not and could_not_verify).

EvidenceDoes it hold?

The limitWhat a guardrail is not

• IT IS NOT A CHECK ON bonus_recurring ITSELF. The flag trusts the model's own answer for that field completely -- nothing re-derives bonus_recurring from the stub text before computing the flag. This run found zero disagreements on that field across 55 stubs on both tiers (see Eval.scores), which is evidence the prompt's stated-history rule held up on this corpus, not evidence the guardrail itself checked and cleared anything -- it never looks at the stub text at all.
• It does not guarantee an approval or a denial -- the flag is a routing signal for a human reviewer, never itself a decision. See UI.shots and Data.breaks_on.
• It has not been attacked. Whether a Bonus History note crafted to mimic a genuine payment history could also evade the flag is unmeasured -- see Security.could_not_verify.

WatchedWhat is watched, and why that one

The kit measures 14 watched figures plus 8 comparability guards and 5 reference constants — three duties, not one list. 0 of the watched are grader rates, each on its own grader page; this board watches the ones that decide when to look again. Each square is one watched figure; filled means the latest run measured it.

8 measured by the latest run 6 need the model half

GuardsBefore any comparison

Preconditions, not alarms. If one moves, the runs were not measuring the same system and nothing else on this board may be differenced.

all shared guards held between the recorded runs (documents 55, extraction_cells 448, failures 0, refusal_cells 102, thinking True) — which is why the history below is allowed to mean anything.

The reference constants — the null-grader baseline, the adjudication matrix, the negative count — are the third duty. They are never re-measured per run: a baseline that moves with the thing it measures is not a baseline.

BandsThe alarm bands

Derived from evidence, never picked: exact-match where two runs reproduced a metric to the digit, wider than the measured same-input drift for latency, one row of the negative class for any rate — denominator printed beside it.

HistoryRun history

2 recorded runs. A dash is a metric that run did not measure, and a dash is never differenced: not measured is a third state, neither a value nor a zero, and reading it as zero is how a board reports an improvement on a run that never looked.

not a time series No two of these 2 runs measured the same system — they differ on the model — so nothing here may be differenced and the Verdict column is left off. Read this block DOWN a column, as a comparison between those runs, not along a row as a history.

DeviationsWhat deviated

0 breaches across 2 runs

The day a band breaks, a row renders here carrying the metric, the two runs, the band it crossed and what changed between them — model, prompt, chunker or top_k, read off the run records’ guards. A deviation with no diff beside it is an alarm nobody can act on.

RippleThe ripple map

Which figures move together, so one alarm reads as a system. Every edge carries its basis: measured names the run that showed it; reasoning admits nothing has varied it yet.

FiresWhat fires when a band breaks

Where evals stop sitting beside the guardrail and start feeding it. Where the answer is nothing automatic, the row says so — hiding it would undo “What a guardrail is not” above.

Nothing fires yet. Not one of the 6 bands on this kit names a consequence, so there is no row to print. Which is the honest answer to the heading: a band that nothing is wired to is a number somebody has to notice, and this kit has only those. An empty table would have said the same thing while looking like a rendering failure.

NextThe three you would add first

None of these is built here. They are named against the seam they sit on so they can be argued with.

CadenceWhen to re-run, and what it costs

Re-check the flag's evidence on any change to OT_CONTINUANCE_THRESHOLD or LARGE_BONUS_THRESHOLD_USD in src/extract.py, or to the bonus_recurring rule in src/prompt.py -- either changes what the flag is computed FROM. Re-run evals/run.py (paid) on any change to MAX_TOKENS or the provider/model.

In one lineNo framework, deliberately

No framework and no dependencies beyond the standard library -- requirements.txt is empty, with a comment explaining that the emptiness is load-bearing. The whole extraction decision is three files: src/segment.py, src/select.py and src/adapters/__init__.py.

The mappingSeam by seam, who owns what

The seams are the ones this kit already publishes, so the mapping cannot drift from the code.

GraphsWhere LangGraph changes the answer

Nothing here is a graph. The pipeline has one path per stub -- segment, select, prompt, call, parse, compute -- with no branching and no state carried between stubs. A framework would add an orchestrator to a for-loop.

The other sideWhat a framework costs you

• A different field schema (data/fields.json) or a different continuance rule needs its own gold set and its own evals/check_labels.py pass -- a framework's own schema layer would not remove that work, only relocate it.
• SECTION_HINTS is a five-minute edit for a new stub layout because it is a plain dict, not a configured retriever -- a framework's own chunking/retrieval abstraction would need its own re-tuning pass instead, with its own failure modes to learn.
• Swapping providers is one function and one entry in PROVIDERS (src/adapters/__init__.py) -- a framework's own model abstraction would add a dependency and a version to track for the same one-line change this file already gives away free.

In one lineWhat one run of this kit actually cost, in time and tokens

Measured on r001-income-verify on the fast tier, 2026-08-21. This kit records telemetry measured per run — 4 of the 6 readings on this axis, 4 of them with a band the guardrail board can judge against.

The readingWhat the last run measured

No movement column. Not one of the 1 earlier run on record is comparable with this one — a different model tier, or a changed guard — and differencing across that measures two systems rather than one. Said once here rather than as a dash on every row.

Retrieval not shown. No run of this kit has ever recorded a retrieval timing, because the pipeline has no such stage. Those readings are left out rather than dashed: a dash says “this run did not measure it”, which promises a later run could.

DriftMedian call, run over run

Hatched bars are runs that are not comparable with the latest — a different model tier or a changed guard. They are drawn because hiding them would make the history look shorter than it is, and greyed because differencing across them would be measuring two systems.

The other sideWhat this does not tell you

The first of those is the one worth doing something about, and it is a solved problem: instrumenting a pipeline so each step reports its own duration is a couple of lines. Observability / tracing has the code, and what the result looks like once it is running.

In one lineWhat you provision, honestly

Everything measured on this page says where it was measured; everything not measured says that — the same contract as the other thirteen pages. Where a retrieval index would be on other kits, this one has: no index — each document goes whole into one call.

The machineWhat this needs

Enumerated from the kit’s own source by build/facts/envscan.py, committed, and drift-checked — never typed into this page.

Your dataWhere everything sits

The wireWhat crosses it, and when

Zero-egress topology: point BASE_URL at a local OpenAI-compatible server and nothing leaves the machine — statically true of the shipped adapter, whose only credential travels in the Authorization header.

The shapeHow it ships, and what you add

The ladderEach decision, and what you provision past its ceiling

Every threshold is a measurement with its provenance beside it, or says is_measured:false. A ceiling with no seam says so — it is the point the kit is outgrown, never a forced join.

When it failsFailure signatures

The knobsThe config surface, enumerated

Environment variables: API_KEY, BASE_URL, EMBED_MODEL, MAX_CALLS_PER_DAY, MODEL, PORT, PROVIDER.

The other sideWhat this page does not claim

The corpus licence, from the Data lens: MIT — this repository's own licence. Every employee, employer and dollar figure is invented; a real pay stub cannot be published at all — it is one of the most sensitive documents a person holds. Your corpus’s licence is yours to verify, and the labels you write are about your documents.

In one linePer-field exact match, light normalisation

Does the model's value for each field match gold, after trimming whitespace/punctuation and treating numbers within half a cent as equal?

MethodHow the test was run

evals/judge.py::score, in-process, no key and no model. The same function evals/baseline.py and evals/run.py both call -- a baseline and a real run scored by two different scorers cannot be compared honestly.

Every grader on these pages scored the same 110 already-recorded answers. Nothing was re-generated, so this compares rulers and not models.

The inputOne real row, seen by every grader

The formulaWhat it computes

The analysisWhat it actually did

In operationWhat to monitor

Reference standard: tools/build_corpus.py's gold, read back off the same generated YTD figures the document states (never a target that seeded them); evals/check_labels.py asserts stated/value consistency on every row before any run is allowed to spend.

How tight can the band be? There is no tolerance band on the field grade itself -- it is exact match after trimming whitespace/punctuation and treating numbers within half a cent as equal, never a continuous score to round.

Cadence: Re-score on any change to src/prompt.py, data/fields.json or tools/build_corpus.py -- the first two change what is asked, the third changes what is asked about. Re-run evals/run.py (paid) on any change to MAX_TOKENS or the provider/model.

The decisionWhen to reach for it

In one lineReview-flag confusion matrix

Does the run's own pure-code needs_review (computed from what the model extracted) match the same computation run over gold's own true values?

MethodHow the test was run

evals/judge.py::score_flags, in-process, no key and no model. Computed from the run's own pure-code needs_review (src/extract.py::compute), never from a second model call -- the model never sees or states the flag directly.

Every grader on these pages scored the same 110 already-recorded answers. Nothing was re-generated, so this compares rulers and not models.

The inputOne real row, seen by every grader

The formulaWhat it computes

The analysisWhat it actually did

In operationWhat to monitor

Reference standard: tools/build_corpus.py's gold needs_review, computed the identical way src/extract.py::compute() computes it -- from the same generated YTD figures the document states, using OT_CONTINUANCE_THRESHOLD=0.75 and LARGE_BONUS_THRESHOLD_USD=1000.0.

How tight can the band be? There is no tolerance band -- the flag is a binary confusion matrix (flagged/not flagged against gold), never a continuous quantity to round.

Cadence: Re-score on any change to OT_CONTINUANCE_THRESHOLD or LARGE_BONUS_THRESHOLD_USD in src/extract.py or tools/build_corpus.py, or to data/fields.json's bonus/overtime fields. Re-run evals/run.py (paid) on any change to MAX_TOKENS or the provider/model.

The decisionWhen to reach for it