The business caseThe problem this solves

Deciding whether a commission claim is owed is a five-branch computation with a priority order — the booking source first, then whether the stay was already commissioned, then a commissionable base that is room revenue net of refunds or a cancellation penalty actually charged, and never the taxes and fees sitting on the same folio — and the record puts the property reviewer's own note right next to it. That note is the loudest thing on the line and it is the one part of the record that is nobody's measurement. Someone opening each line of a monthly commission invoice, finding the property's own folio for the booking it names, and deciding whether the money is owed at all before deciding whether the amount is right — did the booking actually come through this channel, was it already commissioned last cycle, did the guest stay, was any of the room revenue refunded, was a cancellation penalty charged, and is the base room revenue alone rather than the folio total. It is a minute a line and it is hundreds of lines, and the part that goes wrong is not the multiplication: it is doing the multiplication first, or reading the reviewer's own note on the line and letting it set the answer.

The inputThe actual commission claim lines

The corpus is 55 commission claim lines, 0.04 MB (txt 55). Plain text, one format, invented rather than fetched — a real commission claim line joined to a real folio names a property's revenue, a channel's contracted percentage and a specific reservation, and there is no public corpus of (commission claim, correctly-owed-amount) pairs for the same reason there is no public corpus of bank statements. Generating it also makes the label mechanical: gold is the computation, not somebody's reading of the reviewer's note.

Claim Line
----------
CLM-2026-04-7415

Property
--------
Marloe Field Lodge - Lakeside

Confirmation Number
-------------------
BK-EL-178215

Folio Status
------------
stayed

Booking Source
--------------
channel

Room Revenue
------------
2756.36 USD

Room Revenue Refunded
---------------------
0.00 USD

Non-Room Charges
----------------
530.52 USD

Cancellation Penalty
--------------------
not applicable (the stay was not cancelled)

Contract Rate
-------------
17.5 pct

Claimed Commission
------------------
482.36 USD

Previously Commissioned
-----------------------
yes

Invoice Status
--------------
unpaid

Reviewer Note
-------------
Something looked off against the folio on this line, revisit before settlement.

The outcomeWhat a good result looks like

A fourteen-field extracted record per claim line, plus one pure-code routing decision taken from two of those fields: a claim that is not owed as claimed on an invoice the property has already paid is the one that goes to finance today. Nothing here short-pays an invoice, raises a dispute or contacts a channel.

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 monthly commission invoice for lines that are not owed, before a revenue analyst opens them — either tier — they tie on every record both answered, 55 of 55 verdicts with 1.00 recall and precision measured here
  100 pct claim-validity accuracy on both tiers against the free reviewer-note floor's 60.0 pct (22 wrong, including 9 unowed claims waved through).
• Deciding the two tiers on cost, speed or accuracy — the fast tier
  About 13 pct cheaper per claim ($0.0026083 vs $0.0029975), 44 pct lower p50 latency (4,931 ms vs 8,764 ms), and identical on every published grader over every record both answered. The deliberating tier is also the one that lost a document to a timeout.

At a glanceHow the whole thing runs

Run once, for real, on 2026-08-22. 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 2026-08-22 — r001-commission-audit. 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 — Checked on a fresh checkout with API_KEY left blank: python -m src.app starts, the 55-record picker populates and the field table draws its fourteen empty rows; clicking Extract returns the no-key sentence rather than a stack trace. python -m evals.check_labels passes with no network access at all, and python -m evals.run --run-id b000-rules --baseline reproduces the free floor's published score offline.

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 commission claim record

Current processWhat it replaces

Someone opening each line of a monthly commission invoice, finding the property's own folio for the booking it names, and deciding whether the money is owed at all before deciding whether the amount is right — did the booking actually come through this channel, was it already commissioned last cycle, did the guest stay, was any of the room revenue refunded, was a cancellation penalty charged, and is the base room revenue alone rather than the folio total. It is a minute a line and it is hundreds of lines, and the part that goes wrong is not the multiplication: it is doing the multiplication first, or reading the reviewer's own note on the line and letting it set the answer.

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 record with none falls back to one whole-document segment

SetupWhat the setup figure measured

There is no index. Preparation is segmentation only — 55 records cut into 770 sections in under a thousandth of a second, in process, with no model and no network.

LicenceLicence

MIT — this repository's own licence. Every claim id, confirmation number, property name and reviewer note is invented; no real booking channel, hotel, guest or distribution agreement is named or reproduced.

Bring your ownBring your own commission claim lines

Replace data/corpus/*.txt, write your own data/fields.json, and supply a gold record per claim line. SECTION_HINTS in src/select.py maps fields to section headings and will need editing for a different statement layout; when it does not match, selection falls back to the whole document — slower, more expensive, always correct. owed_commission() and compute() in src/extract.py are this kit's own invented commission structure and routing rule and should be the first things you replace with your own agreement.

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 1,785 tokens assembled, 806 are instructions — 45% 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 method on CMA-0017 — three calls at max_tokens=1, each part's size the difference between two consecutive prompt_tokens counts the provider itself returned. kits/UC0043-commission-audit/results/tokens-p001-commission-audit.json.

Full promptThe prompt in full

You extract structured fields from a channel commission claim record. Each record is one line of a booking channel's commission invoice, joined to the property's own folio record for the booking it claims against. You return JSON and nothing else.

RULES, in order of importance:
1. If the record does not state a field, return null for it. Do not infer it and do not use what you know about the world.
2. `claim_valid` is decided by COMPUTING WHAT IS ACTUALLY OWED from the structured folio values and comparing it against claimed_commission_usd -- never by how the reviewer note reads. Compute it yourself, IN THIS ORDER:
   a. If booking_source is anything other than 'channel', NOTHING is owed. The booking did not come through this channel, so no commission is earned on it whatever the stay looks like. Stop here.
   b. Otherwise, if already_commissioned is 'yes', NOTHING is owed. It was commissioned on a previous invoice and is not owed twice. Stop here.
   c. Otherwise work out the commissionable base. If folio_status is 'stayed' or 'rebooked', the base is room_revenue_usd MINUS room_revenue_refunded_usd. If folio_status is 'cancelled' or 'no_show', the base is penalty_charged_usd. non_room_charges_usd is NEVER part of the base, in any branch.
   d. If that base is zero or less, nothing is owed.
   e. Otherwise the commission owed is base x contract_rate_pct / 100, rounded to the cent.
   Answer 'yes' when claimed_commission_usd equals the amount you computed, to the cent. Answer 'no' in every other case.
3. CHECK THE BOOKING SOURCE AND THE PRIOR-COMMISSION FLAG BEFORE YOU DO ANY ARITHMETIC. A claim on a direct, corporate_gds or walk_in booking is owed nothing even when the multiplication is perfect, and so is a stay already commissioned last cycle.
4. A CANCELLED OR NO-SHOW BOOKING IS NOT AUTOMATICALLY WORTH NOTHING. When a penalty was charged, commission is owed ON THE PENALTY. It is owed on nothing when the penalty charged was zero.
5. 'rebooked' MEANS THE GUEST STAYED. The reservation moved to a new confirmation number and the stay happened, so it is commissionable on its room revenue exactly as a plain stay is.
6. THE REVIEWER NOTE IS A FIELD TO COPY, NOT EVIDENCE ABOUT VALIDITY. A note that reads like a dispute does NOT mean the claim is wrong, and a note that reads as settled does NOT mean it is right. The structured folio values decide; the note is the property reviewer's own remark and may disagree with them.
7. Copy values verbatim from the record wherever possible, and report every money and percentage field as a bare number with the currency or percent sign left out of it. room_revenue_refunded_usd is null on a cancelled or no-show booking, and penalty_charged_usd is null on a booking the guest stayed on (including a rebooked one) -- return null for those rather than 0 or a guess.
8. Use the exact allowed value for a field that lists them.
9. Return every field named in the schema, even when the answer is null.

Extract these fields:
- claim_id (string) -- the commission claim line identifier, verbatim
- confirmation_number (string) -- the booking confirmation number this claim line cites, verbatim
- folio_status (enum) one of: stayed, cancelled, no_show, rebooked -- what the property's own folio says happened to this booking, verbatim. 'rebooked' means the guest moved to a new confirmation number and DID stay
- booking_source (enum) one of: channel, direct, corporate_gds, walk_in -- how the booking actually reached the property according to the folio, verbatim. 'channel' means it came through the booking channel that issued this invoice; every other value means it did not
- room_revenue_usd (number) -- gross room revenue posted on the folio, as a bare number without the currency
- room_revenue_refunded_usd (number) -- the portion of room revenue refunded to the guest, as a bare number without the currency. A cancelled or no-show booking earned no room revenue and this line is not applicable -- return null for it rather than 0 or a guess
- non_room_charges_usd (number) -- taxes, fees and incidentals posted on the same folio, as a bare number without the currency. These are NEVER commissionable
- penalty_charged_usd (number) -- the cancellation or no-show penalty actually charged to the guest, as a bare number without the currency. A booking the guest stayed on (including a rebooked one) has no penalty line and this is not applicable -- return null for it rather than 0 or a guess
- contract_rate_pct (number) -- the contracted commission percentage on this property's agreement, as a bare number without the percent sign
- claimed_commission_usd (number) -- the commission amount this invoice line claims, as a bare number without the currency
- already_commissioned (enum) one of: yes, no -- does the folio record this stay as already commissioned on a previous invoice?
- invoice_status (enum) one of: unpaid, paid -- has the property already paid this commission invoice, or is it still unpaid?
- reviewer_note (string) -- the property reviewer's own free-text note on this claim line, copied verbatim
- claim_valid (enum) one of: yes, no -- does the claimed commission equal what is actually owed on this booking? Decide this STRICTLY from the structured folio values -- never from reviewer_note. The rule, IN THIS ORDER: (1) if booking_source is not 'channel', nothing is owed at all, whatever the stay looks like; (2) otherwise, if already_commissioned is 'yes', nothing is owed -- it was commissioned on a previous invoice; (3) otherwise the commissionable base is room_revenue_usd MINUS room_revenue_refunded_usd when folio_status is 'stayed' or 'rebooked', and is penalty_charged_usd when folio_status is 'cancelled' or 'no_show'; non_room_charges_usd is NEVER part of the base; (4) if that base is zero or less, nothing is owed; (5) otherwise the commission owed is base x contract_rate_pct / 100, rounded to the cent. Answer 'yes' when claimed_commission_usd equals that amount to the cent, 'no' in every other case.

Return a JSON object with exactly these keys: claim_id, confirmation_number, folio_status, booking_source, room_revenue_usd, room_revenue_refunded_usd, non_room_charges_usd, penalty_charged_usd, contract_rate_pct, claimed_commission_usd, already_commissioned, invoice_status, reviewer_note, claim_valid
Use null for any field the record does not state.

COMMISSION CLAIM RECORD
-----------------------
Claim Line
----------
CLM-2026-12-7929

Confirmation Number
-------------------
BK-DG-965889

Folio Status
------------
no_show

Booking Source
--------------
channel

Room Revenue
------------
0.00 USD

Room Revenue Refunded
---------------------
not applicable (no room revenue was earned)

Non-Room Charges
----------------
35.04 USD

Cancellation Penalty
--------------------
348.22 USD

Contract Rate
-------------
15.0 pct

Claimed Commission
------------------
52.23 USD

Previously Commissioned
-----------------------
no

Invoice Status
--------------
paid

Reviewer Note
-------------
Something looked off against the folio on this line, revisit before settlement.

Raw responseThe raw response

{"claim_id":"CLM-2026-12-7929","confirmation_number":"BK-DG-965889","folio_status":"no_show","booking_source":"channel","room_revenue_usd":0.00,"room_revenue_refunded_usd":null,"non_room_charges_usd":35.04,"penalty_charged_usd":348.22,"contract_rate_pct":15.0,"claimed_commission_usd":52.23,"already_commissioned":"no","invoice_status":"paid","reviewer_note":"Something looked off against the folio on this line, revisit before settlement.","claim_valid":"yes"}

Step 06 of 14Evals

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

In brief

Check a hotel's booking-site commission bill against each stay — 55 commission claim lines. Two tiers of one model family answered, and every answer was then graded Three 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 gradersThree 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 reviewer-note 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 (55 and 54 records) priced at Google Gemini 3 Flash's published rate — the same basis cost_per_query_usd uses, not a second, larger spend. It excludes the seven calls spent outside the scored runs: three to MEASURE MAX_TOKENS, three to measure the prompt split at max_tokens=1, and one to capture the worked example verbatim.

Cost driversWhat actually moves the bill

• The fixed system prompt and field schema (1532 of 1785 tokens on the example call, 86 pct) outweigh the record sections sent (253 tokens) — the floor every call pays before a single value is read. Fourteen fields with a five-branch rule stated in full is a bigger floor than a simpler kit's, and it is paid on every line.
• Output length: the model returns a full fourteen-key JSON record every call, including the reviewer's note copied back verbatim, whatever the line says. Measured output ranged 294 to 746 tokens across three calibration records of identical shape — a 2.5x spread on a fixed key set, which is why MAX_TOKENS is set from the tail rather than the median.

Your volumeWhat it costs at your volume

Linear in claim lines: each call is independent, carries the same fixed prompt and shares nothing with its neighbours, so 550 lines cost ten times 55 and take ten times as long. Nothing amortises — there is no index to build and no cache, which is also why there is no volume discount to find without changing the design.

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 commission claim records, one completion call each, on the fast tier. The deliberating tier's own 54 answered calls are recorded separately in Cost.cost_by_model and are not projected here.

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.

Five 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 commission claim record 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 — claim validity is mapped to the eight folio facts the rule actually reads, never to the reviewer note and never to the non-room charges the base excludes, and the Property section is mapped by nothing and never sent

You change it to: map fields to your own claim/folio layout'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 fourteen fields, with the commission rule stated in full — booking source checked FIRST, the prior-commission flag second, and only then a base that is room revenue net of refunds or the penalty actually charged

# Assemble the extraction prompt. One prompt per commission claim record, all fourteen fields

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 business-condition check downstream: a claim that is not owed as claimed on an already-paid invoice is routed for a recovery claim

You change it to: the commission rule itself — your own agreement's eligibility branches, base definition and rate. It is stated once and used by the corpus generator, the prompt and the scorer, so changing it here changes all three together

# Extract one channel commission claim record's fields: segment, select, prompt, one model call,

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
STAY_STATUSES = ("stayed", "rebooked")
CANCEL_STATUSES = ("cancelled", "no_show")
FOLIO_STATUSES = STAY_STATUSES + CANCEL_STATUSES
BOOKING_SOURCES = ("channel", "direct", "corporate_gds", "walk_in")
def load_fields():

evals/judge.pyjudge

score field accuracy, the claim-validity confusion matrix and the recovery flag separately, plus a per-case breakdown across the eleven fault and valid shapes, 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(fields, records, golds):
def _matrix(rows, positive):
def _gold_valid(g):
def score_flags(records, 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: 1783 input and 572 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 commission claim records are entirely synthetic (tools/build_corpus.py, seed 20260822): no real booking channel, property, guest or reservation exists in the corpus, and nothing was fetched from anywhere. There is no guest in the corpus at all — a real folio names one, and none of that is needed to decide whether a commission is owed. The only outbound traffic the kit makes is one chat-completion request per record to the configured provider, carrying the mapped sections of one claim record. Nothing is written outside the kit directory, there is no database, no auth and no multi-tenancy, and the local UI binds 127.0.0.1 only.

Read from the shared .env or the real environment only, never committed, never logged and never rendered. .env is gitignored from the first commit. On an adapter error the app redacts the key and base URL out of the message before returning it, so a misconfigured endpoint cannot echo a credential into a browser.

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

The three boundaries that hold were confirmed by reading the code, not by an attack run: no code path short-pays or contacts anybody; the routing rule reads only enums and numbers, so the reviewer's note cannot reach it; and the key and base URL are redacted out of any error before it is returned. The third picked up one piece of unplanned evidence when a real connection timeout wrote itself into a committed result file carrying no endpoint and no credential. The fourth is open. An indirect prompt injection needs a field somebody outside the computation authored, and this kit has exactly one — the property reviewer's note — which makes it the obvious place to attack and the reason the gap is named rather than glossed. Confirmed by reading the code, not by a run, on 2026-08-22 -- no attack run was fired; see redteam.why.

The first three boundaries hold, confirmed by reading the code rather than by an attack run — and the third has one real, unplanned piece of evidence behind it, because a live transport failure did write an error into a committed result file and it carries no credential. The fourth is open and is named as open: an indirect prompt injection in the reviewer's note is exactly the attack this kit's shape invites, and nothing here has tried it.

The result0 attack trials, and three of four boundaries checked here hold on evidence. The one that does not is the one this kit's own shape invites: the reviewer's note is free text, and whether an instruction hidden in it could move the verdict is unmeasured.

This run's corpus is entirely generated (tools/build_corpus.py, seed 20260822), so no text in it came from an outside party and there was nothing adversarial to resist. The planted ambiguity is a REGISTER mismatch — a settled-sounding note on a bad claim — which measures whether a model runs the computation when the prose points the other way. It does not measure whether a model obeys an instruction hidden in the same field. Those are different failures and only one of them is measured here. Note also that in a real deployment the reviewer's note is written inside the property, while the CLAIM LINE beside it comes from the channel — so a production version of this kit has an externally-authored surface this corpus does not model at all.

HonestyWhat this does not prove

• Whether an indirect prompt injection in the reviewer_note field could move claim_valid on either tier. No attack run exists.
• Whether a production version — where the claim line itself is authored by the channel rather than generated here — carries an injection surface this corpus does not model. Every character in this corpus was written by tools/build_corpus.py.
• Whether the kit behaves safely against a hostile provider — a response body crafted to break the JSON extraction in src/prompt.py::parse, or to return an enormous payload. parse() fails closed to an empty dict, which the harness records as a failed document, and nothing beyond that was tested.

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.

needs_recovery fires when the commission claimed is not what is owed AND the invoice has already been paid — claim_valid == "no" and invoice_status == "paid". Both values come out of the same reply; the rule is run afterwards in pure code, over whatever the model returned, and never over gold. A reply missing either value returns None rather than False: an unknown is not a pass.

src/extract.py::compute(), called by extract() on every record, by the local app on every click, and by evals/baseline.py on the free floor's own output so the two are routed by identical code. The commission computation it reads lives in a SEPARATE function, owed_commission(), which is also what the corpus generator wrote gold with and what src/prompt.py states to the model — one definition, three readers.

EvidenceDoes it hold?

The limitWhat a guardrail is not

• It is NOT a check on whether the extracted values are right. If the model misreads the penalty and then judges that misreading consistently, the line is routed (or not routed) on a wrong number and this rule cannot tell. The consistency diagnostic in evals/judge.py is the closest thing to that check and it is reported separately, and is itself blind to the same case.
• It is NOT a real property's dispute policy. Not-owed-and-already-paid is this kit's own simplification, invented for this corpus. No booking channel's published dispute procedure, no signed distribution agreement and no property's own recovery threshold was consulted, and none is reproduced. A real desk weighs the dollar size of the variance, whether the dispute window is still open, and the cost of raising the claim at all.
• It is NOT a disposition. Nothing here short-pays an invoice, raises a dispute or credits anybody, and claim_valid is an extracted field rather than a decision anybody is bound by.

WatchedWhat is watched, and why that one

The kit measures 16 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.

7 measured by the latest run 9 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 (refusal_cells 0) — 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

3 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 3 runs measured the same system — they differ on documents, extraction_cells, failures, provider, thinking, 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 3 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.

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 compute() or owed_commission() in src/extract.py, on any change to data/fields.json's allowed values for claim_valid or invoice_status, and on any corpus regeneration. Changing WHO GETS CHASED is a policy change and must be re-scored, even though it never changes what is owed.

In one lineNo framework, deliberately

No framework and no dependencies beyond the standard library. requirements.txt is empty on purpose: the corpus is generated in-process, the provider is reached over urllib, and the UI is one HTML file with no build step. A forker runs this on whichever key they already hold, with one clone and no install.

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 claim line — segment, select, prompt, one call, parse, compute — with no branch, no loop and no agent. Anything that looks like orchestration in a kit this size is a diagram of a straight line.

The other sideWhat a framework costs you

• Everything is hand-rolled, so everything is yours to maintain: the JSON extraction from a fenced reply, the retry/backoff policy, the section regex and the .env reader are all code somebody has to own. This kit paid that bill during its own build — see the transport-timeout entry in Eval.taxonomy.
• No framework means no framework's ecosystem — no tracing, no eval harness beyond the one in evals/, no prompt registry, no schema validation library. What ships is what is in the repository.
• The provider abstraction covers exactly two wire formats. A third provider is one function and one dict entry, and until somebody writes it the kit runs on two shapes.

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

Measured on r001-commission-audit on the fast tier, 2026-08-22. 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 2 earlier runs 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.

1 run not plotted. b000-commission-audit-rules recorded no call latency at all — a rules-only floor run makes no model call, and the 0 its record stores is an absence, not a zero-millisecond answer. Plotting it would put a bar on this chart claiming the fastest run in the kit’s history.

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 claim id, confirmation number, property name and reviewer note is invented; no real booking channel, hotel, guest or distribution agreement is named or reproduced. 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? Two fields are legitimately null on part of the corpus and complementary — room_revenue_refunded_usd is null exactly on a cancellation or no-show, penalty_charged_usd exactly on a stay — so a null there is a hit, not a miss, when gold agrees.

MethodHow the test was run

evals/judge.py::score, in-process, no key and no model. The same function evals/run.py calls after every call, and the same field-match logic the free baseline is scored by — 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 folio status, booking source, revenue, refund, penalty, rate and claimed amount the document states; evals/check_labels.py asserts every non-nullable field is populated on every row, that both nullable fields' nullness agrees with folio_status and that no row is null on both, that every validity label agrees with its own values, and that the penalty-window and rebooked cases both compute correctly, 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 numbers within half a cent are treated 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 lineclaim_valid against gold's own computation

Of every commission claim that really is not owed as claimed, how many did the run call wrong — and how many correct claims did it wrongly dispute? AN UNOWED CLAIM IS THE POSITIVE CLASS: a line the property does not owe and pays anyway is the failure a hotel finance desk actually pays for.

MethodHow the test was run

evals/judge.py::score_flags, in-process, no key and no model.

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: Gold's validity is re-derived inside the grader by the same computation the kit publishes — booking source first, prior commission second, then a base that is room revenue net of refunds or the penalty charged, never the non-room charges — so the truth this matrix grades against can never be a separately-typed label that drifted from the rule.

How tight can the band be? No threshold — the verdict is one of two allowed values, and a reply that returns neither is counted as unanswered rather than folded into the correct-negative cell. The deliberating tier has exactly one such row, and it is a lost connection rather than a lost answer.

Cadence: Re-run whenever owed_commission() in src/extract.py changes, whenever the corpus is regenerated, and on any provider or model change.

The decisionWhen to reach for it

In one lineneeds_recovery against the same rule run over gold

Does the pure-code routing decision — not owed as claimed AND the invoice is already paid — land on the same lines it would land on if both fields had been read perfectly? It is a BUSINESS CONDITION, so unlike a self-consistency check it genuinely needs labels, and saying so is half of what makes the number believable.

MethodHow the test was run

evals/judge.py::score_flags, in-process, no key and no model.

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: src/extract.py::compute(), the same function the run uses, applied to GOLD's claim_valid and invoice_status. One rule, two inputs, so a change to the rule moves both sides together and the grader cannot silently grade an old policy.

How tight can the band be? No threshold — two booleans and an AND.

Cadence: Re-run whenever compute() changes. Changing WHO GETS CHASED is a policy change and must be re-scored, even though it never changes what is owed.

The decisionWhen to reach for it