The business caseThe problem this solves

Deciding whether a manufactured batch met its specification means comparing one measured laboratory result against the stated limits — arithmetic — while the certificate also carries the analyst's own free-text disposition note, which is an opinion and can point the other way. Someone reading each certificate of analysis, finding the measured result and the two specification limits among the analyst's own commentary, and doing the comparison by hand before the batch moves — including on the certificates whose disposition note reads more reassuring, or more worried, than the numbers warrant.

The inputThe actual certificates

The corpus is 55 certificates, 0.03 MB (txt 55). Plain text, one format, invented rather than fetched — a real certificate of analysis cannot be published either (see SOURCES.md). Ten fields are chosen because the same ones decide a release review: which batch, which material, which parameter, the result and its unit, the two limits — either of which may be absent on a one-sided specification, and 31 of these 55 records carry one — the analyst's own note, and the verdict this kit exists to test.

Batch
-----
BATCH-E-5330

Product
-------
Brindle Alkalised Cocoa Powder

Test Parameter
--------------
bulk density

Measured Result
---------------
0.61 g/mL

Specification Lower Limit
-------------------------
0.55 g/mL

Specification Upper Limit
-------------------------
0.75 g/mL

Test Date
---------
2026-06-24

Analyst
-------
Casey D. Rossi

Analyst Disposition Note
------------------------
Result consistent with the last twelve batches of this grade. No concerns; pass.

The outcomeWhat a good result looks like

A ten-field extracted record per certificate, plus one pure-code self-consistency check: the same comparison re-run over the model's own extracted numbers, flagged when it disagrees with the model's own verdict. Informational only — nothing here releases or rejects a batch.

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 certificates for batches that are out of specification before a quality reviewer opens them — either tier — they tie at 55 of 55 verdicts, 1.00 recall and precision measured here
  100% conformance accuracy on both tiers against the free analyst-tone floor's 56.4% (24 wrong, including 10 out-of-specification batches called conforming).
• Deciding the two tiers on cost, speed or field-level precision — the fast tier
  About 9% cheaper per certificate ($0.0012524 vs $0.0013675), 45% lower p50 latency (2,415 ms vs 4,396 ms), and identical on every published verdict figure.

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-coa-conformance. 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 certificates segment into 495 sections in 0.002 seconds, tools/build_corpus.py regenerates the whole corpus byte-identically in 0.038 seconds, and the assembled prompt for COA-0033 replays with the same three part sizes 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 certificate

Current processWhat it replaces

Someone reading each certificate of analysis, finding the measured result and the two specification limits among the analyst's own commentary, and doing the comparison by hand before the batch moves — including on the certificates whose disposition note reads more reassuring, or more worried, than the numbers warrant.

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
• 1 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 certificate 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 certificates cut into 495 sections by src/segment.py, pure code, no model and no key.

LicenceLicence

MIT — this repository's own licence. Every product, batch, analyst and specification limit is invented; a real certificate of analysis cannot be published either — it names a real manufacturer, a real lot and a real release decision.

Bring your ownBring your own certificates

Replace data/corpus/*.txt, write data/fields.json, and supply a gold record per certificate. SECTION_HINTS in src/select.py maps fields to headings and will need editing for a different certificate layout; when it does not match, selection falls back to the whole document — slower, more expensive, always correct. If your certificates are unlabelled, the field grade and the confusion matrix both go away and the self-consistency check does not: it is computed from the reply alone.

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 965 tokens assembled, 419 are instructions — 43% 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-coa-conformance.json.

Full promptThe prompt in full

You extract structured fields from a certificate of analysis for one manufactured batch. You return JSON and nothing else.

RULES, in order of importance:
1. If the certificate does not state a field, return null for it. Do not infer it and do not use what you know about the world. A specification limit the certificate explicitly says is not specified is null, not a number you supply.
2. `conforms_to_spec` is decided by ARITHMETIC ON THE NUMBERS, never by how the analyst's disposition note reads. Answer 'yes' when measured_value >= spec_lower_limit AND measured_value <= spec_upper_limit; answer 'no' otherwise. Both limits are INCLUSIVE -- a value exactly on a limit conforms. A limit the certificate does not state places no constraint on that side, so a one-sided specification is judged on the side it does state. Do the comparison yourself before answering.
3. The analyst's disposition note is a field to copy, not evidence about conformance. A note reading 'within normal range, released' does NOT make an out-of-limits value conform, and a note reading 'borderline, recommend re-test' does NOT make an in-limits value fail. The numbers decide; the note is the analyst's opinion and may disagree with them.
4. Copy values verbatim from the certificate wherever possible, and report measured_value, spec_lower_limit and spec_upper_limit as bare numbers with the unit left out of them.
5. Use the exact allowed value for a field that lists them.
6. Return every field named in the schema, even when the answer is null.

---

Extract these fields:
- batch_id (string) -- the batch or lot identifier, verbatim
- product_name (string) -- the name of the material the certificate is for
- test_parameter (string) -- what was measured, verbatim (for example: moisture content, assay purity)
- measured_value (number) -- the numeric laboratory result, without its unit
- unit (string) -- the unit the result is reported in, verbatim (for example: %, ppm, CFU/g, microns)
- spec_lower_limit (number) -- the lower specification limit as a number, or null when the certificate states the specification is one-sided and gives no lower limit
- spec_upper_limit (number) -- the upper specification limit as a number, or null when the certificate states the specification is one-sided and gives no upper limit
- analyst_disposition_note (string) -- the analyst's own free-text disposition note, copied verbatim
- test_date (string) -- the date the test was performed, YYYY-MM-DD
- conforms_to_spec (enum) one of: yes, no -- does the measured value fall within the stated specification limits? Decide this by ARITHMETIC ONLY: it is 'yes' when measured_value is greater than or equal to spec_lower_limit AND less than or equal to spec_upper_limit, treating a limit the certificate does not state as no constraint on that side; otherwise 'no'. Both limits are inclusive -- a value exactly on a limit conforms. Do NOT decide this from the analyst's disposition note: a note reading 'within normal range, released' does not make an out-of-limits value conform, and a note reading 'borderline, recommend re-test' does not make an in-limits value fail.

Return a JSON object with exactly these keys: batch_id, product_name, test_parameter, measured_value, unit, spec_lower_limit, spec_upper_limit, analyst_disposition_note, test_date, conforms_to_spec
Use null for any field the certificate does not state.

CERTIFICATE OF ANALYSIS
-----------------------
Batch
-----
BATCH-B-5515

Product
-------
Brindle Alkalised Cocoa Powder

Test Parameter
--------------
total plate count

Measured Result
---------------
1025 CFU/g

Specification Lower Limit
-------------------------
not specified (one-sided specification)

Specification Upper Limit
-------------------------
1000 CFU/g

Test Date
---------
2026-04-11

Analyst Disposition Note
------------------------
Within normal range for this product line. Released to finished-goods inventory.

Raw responseThe raw response

{
  "batch_id": "BATCH-B-5515",
  "product_name": "Brindle Alkalised Cocoa Powder",
  "test_parameter": "total plate count",
  "measured_value": 1025,
  "unit": "CFU/g",
  "spec_lower_limit": null,
  "spec_upper_limit": 1000,
  "analyst_disposition_note": "Within normal range for this product line. Released to finished-goods inventory.",
  "test_date": "2026-04-11",
  "conforms_to_spec": "no"
}

Step 06 of 14Evals

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

In brief

Check each production batch's lab certificate against its limits — 55 certificates. 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 analyst-tone 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 certificates 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 (789 of 965 tokens on the example call, 82%) outweigh the certificate sections sent (176 tokens) — the floor every call pays before a single number is read.
• Output length: the model returns a full ten-key JSON record every call, including the analyst's disposition note copied back verbatim, whatever the certificate says.

Your volumeWhat it costs at your volume

Linear in certificates: each call is independent and self-contained, with no shared context or index to amortise. This run's 55 certificates cost about $0.069 projected onto Gemini 3 Flash's rate, so ten times the set is about $0.69 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 certificates of analysis extracted, each scored by pure code against a mechanically-derived gold set. This run (r001-coa-conformance, the fast tier) answered all 55 of 55 certificates 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 certificate 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 — conformance is mapped to the measured result and the two limits, never to the disposition note

You change it to: map fields to your own certificate'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, with the comparison rule and its boundary-inclusive convention stated in full

# Assemble the extraction prompt. One prompt per certificate of analysis, 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 self-consistency check downstream: re-run the comparison over the model's OWN extracted numbers and flag any disagreement with the model's own verdict

You change it to: the comparison itself — the boundary convention, or a tolerance band if your programme allows one. It is stated once and used by the corpus generator, the prompt and the guardrail, so changing it here changes all three together

# Extract one certificate'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
def load_fields():
def load_doc(stmt_id):
def documents():
def recompute_conformance(measured, lower, upper):
def compute(values):

evals/judge.pyjudge

score field accuracy and the conformance confusion matrix 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(fields, records, golds):
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: 963 input and 256 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 certificates are entirely synthetic (tools/build_corpus.py) -- no untrusted party wrote any of the text. In a real deployment the analyst's disposition note is free text written by a person working the batch, and the whole certificate may arrive from a supplier rather than from your own laboratory -- exactly the kind of externally-authored input this kit's architecture reads verbatim and trusts, with no verification step before its text 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 certificate 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 disposition notes are written by people, and an incoming-goods review reads certificates issued by suppliers, and neither surface has 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, and it is the same blind spot the guardrail names about itself: a consistently-wrong reply looks exactly like a correct one to a check that only compares the reply against itself.

The result0 attack trials, and three of four boundaries checked here hold in code: no write path exists, the comparison is pure code the model cannot move, and a misconfigured key cannot leak into a UI error. The fourth -- whether a crafted note could produce a wrong answer that still agrees with itself -- is unmeasured, and is the guardrail's own known blind spot.

This run's corpus is entirely generated (tools/build_corpus.py) -- no certificate's text was authored by an outside party. A real deployment's disposition notes are free text written by whoever worked the batch, and an incoming-goods review reads certificates issued by a supplier entirely outside your control -- exactly the kind of externally-supplied text this kit's architecture reads verbatim and trusts. Whether a note crafted to move conforms_to_spec, or to move the extracted numbers so that a wrong verdict still looks self-consistent, would succeed is unmeasured for this kit.

HonestyWhat this does not prove

• Whether a disposition note crafted to talk a model out of its own arithmetic could move conforms_to_spec -- no red-team run exists for this kit. On this corpus's plain register mismatch both tiers were unmoved on all 24 cases.
• Whether a crafted note could move the EXTRACTED NUMBERS so that a wrong verdict still passes the self-consistency check -- the check compares the reply against itself and never re-reads the certificate, so by construction it would not notice. Unmeasured.
• Whether the live app's own /api/extract behaves identically to the registered run under an adversarial certificate -- both use the same src/extract.py::extract(), but neither has been tested against one.

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_review fires when the comparison re-run in pure code over the model's OWN extracted measured_value, spec_lower_limit and spec_upper_limit lands on a different verdict from the model's own conforms_to_spec. It is a SELF-CONSISTENCY check, not a business condition: it needs no gold, no second call and no labelled set, so it computes identically on a certificate nobody has ever scored.

src/extract.py::compute(), which calls recompute_conformance() -- the same comparison tools/build_corpus.py used to write gold and the same one src/prompt.py asks the model to perform, stated once so the three cannot drift. Pure code, run after the model call on the reply's own values; nothing in the reply text can talk it out of firing. What it CANNOT do is check whether those numbers were read correctly in the first place -- a DIFFERENT dependency, named in is_not and could_not_verify.

EvidenceDoes it hold?

The limitWhat a guardrail is not

• IT IS NOT A CHECK ON WHETHER THE NUMBERS ARE RIGHT. It compares the model's verdict against the model's own extracted values; if the model misreads the measured value and then judges that misreading correctly, the reply is self-consistent and this check stays quiet. Catching that needs a second reading of the document, which nothing here does -- see add_first.
• It does not decide conformance either way -- the flag says the reply disagrees with itself, not which half is wrong. See UI.shots and Data.breaks_on.
• It has not been attacked. Whether a disposition note crafted to talk a model out of its own arithmetic -- or a prompt-injection attempt inside the note field -- could produce a consistent-looking wrong answer is unmeasured; see Security.could_not_verify.

WatchedWhat is watched, and why that one

The kit measures 15 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 8 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 550, failures 0, refusal_cells 0, 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 7 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 recompute_conformance() or compute() in src/extract.py, or to the conforms_to_spec rule in src/prompt.py -- all three change what the check compares. 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 certificate -- segment, select, prompt, call, parse, recompute -- with no branching and no state carried between certificates. 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 comparison 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 certificate 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-coa-conformance 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 product, batch, analyst and specification limit is invented; a real certificate of analysis cannot be published either — it names a real manufacturer, a real lot and a real release decision. 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 — including the two limits, where a correctly-returned null on a one-sided specification is a hit and an invented bound is a miss?

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 numbers, note and dates the document states; evals/check_labels.py asserts every non-nullable field is populated on every row, that no record states a limit on neither side, and that every conformance label agrees with its own arithmetic, 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 lineConformance confusion matrix

Does the model's stated conforms_to_spec match the same comparison run over gold's own measured value and limits? Out of specification is the positive class, so recall answers "of every batch that really is out of specification, how many did the run call out of specification".

MethodHow the test was run

evals/judge.py::score_flags, in-process, no key and no model. Derives gold's verdict with src/extract.py::recompute_conformance, compares it to the model's own conforms_to_spec per certificate, and reports the self-consistency diagnostic (src/extract.py::compute) alongside it.

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: The same pure-code comparison in src/extract.py::recompute_conformance, run over gold's own measured_value, spec_lower_limit and spec_upper_limit rather than the model's — the TRUE verdict is never separately typed, only derived.

How tight can the band be? No tolerance band — a binary verdict either matches gold's derived verdict or it does not, scored per certificate.

Cadence: Re-score on any change to src/extract.py::recompute_conformance or to the conforms_to_spec instruction in src/prompt.py — both change what the verdict means.

The decisionWhen to reach for it