The business caseThe problem this solves

A donor file establishes funds whose restrictions are stated in prose, and the institution has to turn that prose into a fund record before it can spend anything. The expensive error has no symptom: a fund recorded as permanently restricted when nobody restricted it is money the institution may spend and never does. Someone opening each executed donor file, listing the funds it actually establishes, and then working out, fund by fund, two things the paperwork may or may not answer: how the principal is restricted and who restricted it, and whether the stated purpose binds the institution or merely records a wish. It is minutes per file and it is every file, and the part that goes wrong is not the listing: it is reading a stated payout rate as evidence that the principal is permanent, and filing a fund the institution's own board designated as though the donor had restricted it.

The inputThe actual donor files

The corpus is 50 donor files, 0.19 MB (txt 50). Real donor files cannot be published: they name living people, they carry addresses and telephone numbers, and the interesting ones are exactly the ones an institution would never release. A generated corpus is the only kind that can ship with a public kit AND carry a gold label for every fund — and because the generator decides the stated facts and then writes them into the prose, the label is the shipped rulebook's own lookup rather than somebody's reading.

LONGACRE ACADEMY
OFFICE OF GIFT ADMINISTRATION — DONOR FILE EXTRACT

Donor file:      Quillingham Donor Advised Fund
File reference:  DF-2012-512
Extract id:      GFT-0001

This extract is prepared for fund-setup review. The Schedule below lists the funds this
file establishes. The Articles that follow set out the terms of each.

DONOR CONTACT AND STEWARDSHIP RECORD

  Correspondence address:  30 Ellenford Way, Wexley
  Telephone:               01710 677581
  Relationship manager:    P. Merrick
  Last stewardship visit:  2023-07
  Acknowledgement preference: written, addressed to the family

SCHEDULE OF FUNDS ESTABLISHED BY THIS FILE
------------------------------------------------------------------------------
CODE        FUND NAME                                   MINIMUM CORPUS
------------------------------------------------------------------------------
GF-1077     The Quillingham Anniversary Fund            USD 265,000
GF-6231     The Quillingham Commemorative Fund          USD 1,790,000
GF-1589     The Quillingham Endowed Fund                USD 740,000
GF-1002     The Quillingham Founders Fund               USD 295,000
------------------------------------------------------------------------------

ARTICLE 1 — The Quillingham Anniversary Fund (GF-1077)

  This Fund is established by the Donor under the terms of this Agreement.
  The original gift shall be maintained in perpetuity as the corpus of the Fund.
  Distributions from the Fund must be applied only to laboratory research in the life sciences.
  The minimum corpus recorded for this Fund is USD 265,000.

ARTICLE 2 — The Quillingham Commemorative Fund (GF-6231)

  The Board of Trustees has resolved to designate the following gift, received unrestricted, for
  the purpose recorded below.

The outcomeWhat a good result looks like

A worksheet per donor file, one row per fund, with the two calls made by a rulebook a reader can open, and an explicit list of the funds the paperwork does not settle.

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.

• Turning a month of executed donor files into fund-setup worksheets before anyone spends against them — the one tier measured here — 269 of 269 on both rulebook calls, and 33 of 33 on the board-designation trap including all 14 whose resolution recites perpetuity
  100 pct on both calls against the free payout-rate floor's 61.8 pct and 74.5 pct, and 0 over-confident calls of 123 against the floor's 56 pct. It also left every one of the 39 decoy codes off, against the floor's 13 phantom funds.
• Deciding whether a document set is worth an extraction kit at all — the free floor first — evals/baseline.py, $0.00, no key
  It reads every stated fact off the file correctly and still scores 61.8 pct on the restriction call, which tells you the hard part is the RULE and not the reading. If your floor is already close to your target, the model is not where your problem is.

And where nothing here is good enough:

• Any workflow where the worksheet would be wired to something that acts — none of it
  There is no configuration of this kit that makes it safe to act on. Every grader measures agreement with an illustrative rulebook, and a fund's restriction is a legal question about a real donor's intent.

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-gift-restrict. 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 — Clone, run python3 -m src.app, and the kit renders with no key and no network — the corpus and every recorded result ship in the repo. A key is needed only to extract a new file or to re-run the eval against your own numbers.

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 donor file

Current processWhat it replaces

Someone opening each executed donor file, listing the funds it actually establishes, and then working out, fund by fund, two things the paperwork may or may not answer: how the principal is restricted and who restricted it, and whether the stated purpose binds the institution or merely records a wish. It is minutes per file and it is every file, and the part that goes wrong is not the listing: it is reading a stated payout rate as evidence that the principal is permanent, and filing a fund the institution's own board designated as though the donor had restricted it.

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 the file's own headings — an explicit list plus an ARTICLE pattern keyed on fund code; a file with none falls back to one whole-document segment

SetupWhat the setup figure measured

There is no index. Preparation is segmentation only — 50 files cut into 527 sections in about ten thousandths of a second, in process, with no model and no network. 427 of those sections are sent; the 100 that are not are the 50 DONOR CONTACT AND STEWARDSHIP RECORD sections, which carry a correspondence address, a telephone number, a named relationship manager and a stewardship history that no field asks for, and the 50 closing trailers.

LicenceLicence

MIT — this repository's own licence. Every donor name, institution, fund code, fund name, address, telephone number, relationship manager, date and amount is invented; no real person, family, donor, foundation, university, charity or trust is named, and no statute, regulator or law firm is named either. The fund-setup rulebook this kit ships (data/rulebook.json) is ILLUSTRATIVE and is not an authority. It was written for this kit and reproduces no statute, no accounting standard, no regulator's guidance and no institution's own gift-acceptance policy.

Bring your ownBring your own donor files

Point src/extract.py's CORPUS at a directory of your own .txt files and the whole pipeline runs unchanged — segment, select, prompt, the model call, the rulebook and the review flag. Only the SCORING needs labels, so you get a worksheet on day one and a number when you have gold.

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.

Token budgetWhere the tokens actually go

160920 of the run's 196664 output tokens — 81.8 pct — were provider-side REASONING rather than the JSON worksheet, on a reply that is roughly two thousand tokens of actual content. Reasoning was left at the provider's default and nothing here has measured what turning it off would cost in accuracy.

AssemblyHow the prompt is assembled

This is the cost lesson as arithmetic: of the 3,746 tokens assembled, 1,828 are contexts — 49% 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 in TOKENS, not characters, by evals/prompt_tokens.py — four calls at max_tokens=1, each a nested prefix of the last, so every part's size is the difference between two counts the provider itself returned.

Full promptThe prompt in full

You read a DONOR FILE EXTRACT -- a schedule of funds plus the articles behind them -- and return a REVIEWER'S WORKSHEET of the funds it establishes. You return JSON and nothing else.

You are preparing a worksheet for a qualified reviewer. You never give a legal opinion, never classify a fund for financial reporting, never authorise expenditure and never post a fund-accounting entry. Your job is to say what the paperwork states, apply the rulebook given below, and NAME THE FUNDS THE PAPERWORK DOES NOT SETTLE.

RULES, in order of importance:
1. ONE OBJECT PER FUND THIS FILE ESTABLISHES, AND NO OTHERS. A fund rescinded by the amendment is not established, even though its Article is still printed in full. A pledge -- a stated intention to give, with a name, an amount and a purpose, where no transfer has been made -- is not a fund. A fund continuing under an EARLIER deed of gift that this file only refers to is not established here. All three carry codes in the same format as a real fund, and putting one on the worksheet puts a restriction in front of a reviewer that does not exist.
2. REPORT WHAT THE FILE STATES. `principal`, `designation_source` and `purpose_language` are readings of the paperwork, not judgements about it. If an Article says nothing about the principal, `principal` is 'silent' -- that is a finding about the instrument, and it is the most common one. Never infer perpetuity from a payout rate, from the word 'Endowed' in the fund's NAME, or from the size of the gift.
3. `restriction` IS DECIDED ONLY BY THE RULEBOOK BELOW, from `principal` and `designation_source`. Work through the steps IN ORDER and STOP at the first that fires:
   a. designation_source 'board' -> 'board_designated'. THIS FIRES FIRST AND IT FIRES WHATEVER THE ARTICLE SAYS ABOUT PERPETUITY. A board resolution that recites holding the principal permanently is still a board designation, and the board can undo it.
   b. 'donor' + 'perpetual' -> 'donor_restricted_perpetual'.
   c. 'donor' + 'term'      -> 'donor_restricted_term'.
   d. 'donor' + 'expendable'-> 'donor_restricted_expendable'.
   e. anything else -> 'not_determined'.
4. A STATED PAYOUT RATE IS A DISTRIBUTION INSTRUCTION, NOT A STATEMENT THAT THE PRINCIPAL IS PERMANENT. An Article whose only financial sentence is 'distribute 4.5 per cent of the trailing average each year' has `principal` 'silent' and `restriction` 'not_determined'. It is not 'donor_restricted_perpetual'. This is the step a confident reader skips, and on this rulebook it is the wrong answer rather than a defensible one.
5. `purpose` IS DECIDED ONLY FROM `purpose_language`: 'mandatory' -> 'binding', 'precatory' -> 'advisory', 'silent' -> 'not_determined'. It is asked of every fund independently of its restriction call. Judge the VERB, not how specific the stated purpose sounds -- a wish that names one named scholarship is still a wish.
6. 'not_determined' IS A REAL ANSWER AND YOU ARE EXPECTED TO USE IT. A worksheet that never reaches for it is guessing, and a call recorded as settled is a call nobody re-reads.
7. Copy `fund_code`, `fund_name` and `purpose_verbatim` verbatim from the file. Use the exact allowed value for every field that lists them, and return every field for every fund.

FUND-SETUP RULEBOOK (the authority for `restriction` and `purpose`; this is an ILLUSTRATIVE rulebook written for this kit, and it reproduces no statute and no accounting standard)

WHAT COUNTS AS A FUND ON THE WORKSHEET
  - A fund is established when this packet contains an executed instrument that TRANSFERS money or property to the institution and gives that transfer a name or a purpose.
  - A PLEDGE is not a fund. A promise to give in the future — however specific its amount, its name and its purpose — establishes nothing until it is funded, and this rulebook does not record it.
  - A fund RESCINDED by a later amendment in the same packet is not a fund. Its article is still printed; the amendment is the operative document.
  - A fund governed by an EARLIER instrument that this packet only refers to is not established by this packet. It is somebody else's row on somebody else's worksheet.

THE THREE STATED FACTS, read off the packet
  principal:
    perpetual      The instrument states the principal, corpus or original gift is to be held permanently, in perpetuity, or is never to be expended.
    term           The instrument states the principal is to be held for a stated period or until a stated date or event, after which it may be expended.
    expendable     The instrument states the principal itself may be spent, in whole or in part, at the institution's discretion or on a stated schedule.
    silent         The instrument says nothing about what happens to the principal. A stated payout, spending or distribution RATE is not a statement about the principal.
  designation_source:
    donor          The restriction is imposed by the instrument the donor executed — the gift agreement, the deed of gift, the will or the trust.
    board          The restriction is imposed by the institution's own board, trustees or governing body, by resolution or internal designation.
    silent         The packet does not say who imposed the restriction.
  purpose_language:
    mandatory      The instrument directs the use in binding words — shall, must, solely for, restricted to, may be used only for.
    precatory      The instrument expresses a wish, hope, preference, request or expectation about the use, without directing it.
    silent         The instrument states no purpose for the fund at all.

RESTRICTION -- work through IN ORDER, stop at the first that fires
  1. If the designation_source is `board`, the call is `board_designated` — WHATEVER the instrument says about perpetuity. A body that designated a fund can undo its own designation, so this is not a donor restriction. Stop here.
  2. Otherwise, if designation_source is `donor` and principal is `perpetual`, the call is `donor_restricted_perpetual`.
  3. Otherwise, if designation_source is `donor` and principal is `term`, the call is `donor_restricted_term`.
  4. Otherwise, if designation_source is `donor` and principal is `expendable`, the call is `donor_restricted_expendable`.
  5. Otherwise the call is `not_determined`. This fires whenever either fact is silent — most often when the packet states a payout rate and never states that the principal is permanent.

PURPOSE
  `mandatory` purpose language makes the stated purpose `binding`.
  `precatory` purpose language makes the stated purpose `advisory`. A wish is not a restriction, however specific it is and however strongly it is worded.
  `silent` makes the purpose call `not_determined`.

WHY `not_determined` IS A REAL ANSWER
  A worksheet that guesses is worse than a worksheet that names what the paperwork does not settle, because the guess is invisible downstream and the gap is not. `not_determined` is a finding ABOUT THE INSTRUMENT and is the correct, complete answer whenever the document does not say. It is never a way of saying the reader was unsure.

Return these:
- packet_ref (string) -- the file reference from the header, verbatim (for example DF-2012-512)
- funds (array of objects) -- one object per ESTABLISHED fund, each carrying:
    - fund_code (string) -- the fund code, verbatim (for example GF-4821). THIS IS THE KEY -- one object per fund this file ESTABLISHES and no others. A pledge, a fund rescinded by the amendment, and a fund continuing under an earlier instrument are NOT established by this file and must not appear
    - fund_name (string) -- the fund's name, copied verbatim from its own Article heading
    - principal (enum) one of: perpetual, term, expendable, silent -- what the fund's own Article STATES about the principal, corpus or original gift. 'perpetual' when it says the principal is held permanently, in perpetuity, or is never to be expended. 'term' when it says the principal is held until a stated date, anniversary or event and may be expended after it. 'expendable' when it says the principal itself may be spent. 'silent' when the Article says NONE of those. ⚠︎ A STATED PAYOUT, SPENDING OR DISTRIBUTION RATE IS NOT A STATEMENT ABOUT THE PRINCIPAL -- an Article that says only 'distribute 4.5 per cent of the average value each year' is `silent`. Silence is the common case here and it is a finding, not a gap: never infer perpetuity from a rate, from the word 'endowed' in the fund's NAME, or from the size of the gift
    - designation_source (enum) one of: donor, board, silent -- WHO imposed the restriction, as the Article states it. 'donor' when the Article says the fund is established by the Donor, or that the terms are the Donor's. 'board' when the Article says the fund is designated by resolution of the Board of Trustees or the governing body -- typically adding that the gift itself was received unrestricted. 'silent' when the Article says neither. Report what the paperwork says; this is NOT on its own what decides `restriction`
    - purpose_language (enum) one of: mandatory, precatory, silent -- HOW the Article words the use of the fund. 'mandatory' for directing words -- shall be used solely for, must be applied only to, is restricted to, may be used for no other purpose. 'precatory' for a wish rather than a direction -- it is the Donor's hope that, would prefer though does not direct, the family expects, requests without imposing any obligation. 'silent' when the Article states no purpose at all. ⚠︎ PRECATORY LANGUAGE OFTEN NAMES A PURPOSE JUST AS SPECIFICALLY AS MANDATORY LANGUAGE DOES. Judge the VERB, not how specific the purpose sounds
    - purpose_category (enum) one of: scholarship, professorship, research, facilities, program, unrestricted, none_stated -- which bucket the stated purpose falls in, whether the language is mandatory or precatory. 'unrestricted' only where the Article names the institution's general or unrestricted purposes. 'none_stated' when `purpose_language` is silent
    - purpose_verbatim (string) -- the words the Article uses for the purpose, copied verbatim, or null when no purpose is stated
    - payout_rate_pct (number) -- the stated annual distribution, spending or payout rate as a number (4.5 for '4.5 per cent'), or null when the Article states none. Recorded because it is the sentence most often mistaken for a perpetuity clause -- it is reported here so the worksheet shows it was READ and not treated as one
    - minimum_corpus_usd (number) -- the minimum corpus recorded for the fund, as a plain number with no separators or currency symbol
    - restriction (enum) one of: donor_restricted_perpetual, donor_restricted_term, donor_restricted_expendable, board_designated, not_determined -- THE RULEBOOK'S CALL, derived from `principal` and `designation_source` by working through the restriction rule IN ORDER and stopping at the first step that fires. Do not reach this from the fund's name, its size, or the presence of a payout rate. ⚠︎ STEP 1 FIRES FIRST: a board designation is `board_designated` however loudly its resolution recites perpetuity
    - purpose (enum) one of: binding, advisory, not_determined -- THE RULEBOOK'S CALL, derived from `purpose_language` alone: mandatory -> binding, precatory -> advisory, silent -> not_determined. Asked of every fund independently of its restriction call

Return a JSON object with exactly these top-level keys: packet_ref, funds
`funds` is an array. Return it empty only if the file establishes no fund at all.

DONOR FILE EXTRACT
------------------
RAVENSMOOR TRUST
OFFICE OF GIFT ADMINISTRATION — DONOR FILE EXTRACT

Donor file:      Marchmont Charitable Remainder Trust
File reference:  DF-2009-869
Extract id:      GFT-0003

This extract is prepared for fund-setup review. The Schedule below lists the funds this
file establishes. The Articles that follow set out the terms of each.

SCHEDULE OF FUNDS ESTABLISHED BY THIS FILE
------------------------------------------------------------------------------
CODE        FUND NAME                                   MINIMUM CORPUS
------------------------------------------------------------------------------
GF-2378     The Marchmont Family Fund                   USD 1,145,000
GF-7838     The Marchmont Anniversary Fund              USD 1,440,000
GF-6803     The Marchmont Commemorative Fund            USD 865,000
GF-7638     The Marchmont Trustees Fund                 USD 1,625,000
GF-8980     The Marchmont Centenary Fund                USD 1,285,000
GF-1195     The Marchmont Founders Fund                 USD 960,000
------------------------------------------------------------------------------

ARTICLE 1 — The Marchmont Family Fund (GF-2378)

  This Fund is established by the Donor under the terms of this Agreement.
  The original gift shall be maintained in perpetuity as the corpus of the Fund.
  The Donor would prefer, though does not direct, that the Fund support research equipment and
  instrumentation.
  The minimum corpus recorded for this Fund is USD 1,145,000.

ARTICLE 2 — The Marchmont Anniversary Fund (GF-7838)

  The restrictions recorded in this Article are those imposed by the Donor.
  No part of the corpus of the Fund may be expended, whether for the stated purpose or
  otherwise.
  The Institute shall distribute 3.9 per cent of the Fund's trailing twelve-quarter average
  market value each year.
  The Donor's family expects that the Fund will ordinarily be used for the student exchange
  programme.
  The minimum corpus recorded for this Fund is USD 1,440,000.

ARTICLE 3 — The Marchmont Commemorative Fund (GF-6803)

  The restrictions recorded in this Article are those imposed by the Donor.
  The original gift shall be maintained in perpetuity as the corpus of the Fund.
  The Institute shall distribute 4.8 per cent of the Fund's trailing twelve-quarter average
  market value each year.
  It is the Donor's hope that the Fund will be applied to grounds and landscape upkeep.
  The minimum corpus recorded for this Fund is USD 865,000.

ARTICLE 4 — The Marchmont Trustees Fund (GF-7638)

  This Fund is designated by resolution of the Board of Trustees of the Institute. The gift
  itself was received without restriction from the Donor.
  The Board further resolves that the principal so designated shall be held permanently and
  shall not be expended.
  Distributions from the Fund must be applied only to faculty chair support.
  The minimum corpus recorded for this Fund is USD 1,625,000.

ARTICLE 5 — The Marchmont Centenary Fund (GF-8980)

  This Fund is established by the Donor under the terms of this Agreement.
  No part of the corpus of the Fund may be expended, whether for the stated purpose or
  otherwise.
  The Donor requests, without imposing any obligation, that the Fund be applied to capital
  renewal of the athletics centre.
  The minimum corpus recorded for this Fund is USD 1,285,000.

ARTICLE 6 — The Marchmont Founders Fund (GF-1195)

  By resolution of the governing body, and not by direction of the Donor, the following
  unrestricted gift is designated as set out below.
  The Institute shall distribute 4.4 per cent of the Fund's trailing twelve-quarter average
  market value each year.
  The Fund shall be used solely for the unrestricted operating needs of the Institute.
  The minimum corpus recorded for this Fund is USD 960,000.

MEMORANDUM OF INTENDED GIFT (NOT AN EXECUTED INSTRUMENT)

  The Donor has indicated an intention to establish The Marchmont Bicentenary Fund (GF-7990)
  with a gift of USD 2,730,000 for grounds and landscape upkeep. No transfer has been made and
  no instrument has been executed. This memorandum records the intention only.

Raw responseThe raw response

{
  "packet_ref": "DF-2009-869",
  "funds": [
    {
      "fund_code": "GF-2378",
      "fund_name": "The Marchmont Family Fund",
      "principal": "perpetual",
      "designation_source": "donor",
      "purpose_language": "precatory",
      "purpose_category": "research",
      "purpose_verbatim": "The Donor would prefer, though does not direct, that the Fund support research equipment and instrumentation.",
      "payout_rate_pct": null,
      "minimum_corpus_usd": 1145000,
      "restriction": "donor_restricted_perpetual",
      "purpose": "advisory"
    },
    {
      "fund_code": "GF-7838",
      "fund_name": "The Marchmont Anniversary Fund",
      "principal": "perpetual",
      "designation_source": "donor",
      "purpose_language": "precatory",
      "purpose_category": "program",
      "purpose_verbatim": "The Donor's family expects that the Fund will ordinarily be used for the student exchange programme.",
      "payout_rate_pct": 3.9,
      "minimum_corpus_usd": 1440000,
      "restriction": "donor_restricted_perpetual",
      "purpose": "advisory"
    },
    {
      "fund_code": "GF-6803",
      "fund_name": "The Marchmont Commemorative Fund",
      "principal": "perpetual",
      "designation_source": "donor",
      "purpose_language": "precatory",
      "purpose_category": "facilities",
      "purpose_verbatim": "It is the Donor's hope that the Fund will be applied to grounds and landscape upkeep.",
      "payout_rate_pct": 4.8,
      "minimum_corpus_usd": 865000,
      "restriction": "donor_restricted_perpetual",
      "purpose": "advisory"
    },
    {
      "fund_code": "GF-7638",
      "fund_name": "The Marchmont Trustees Fund",
      "principal": "perpetual",
      "designation_source": "board",
      "purpose_language": "mandatory",
      "purpose_category": "professorship",
      "purpose_verbatim": "Distributions from the Fund must be applied only to faculty chair support.",
      "payout_rate_pct": null,
      "minimum_corpus_usd": 1625000,
      "restriction": "board_designated",
      "purpose": "binding"
    },
    {
      "fund_code": "GF-8980",
      "fund_name": "The Marchmont Centenary Fund",
      "principal": "perpetual",
      "designation_source": "donor",
      "purpose_language": "precatory",
      "purpose_category": "facilities",
      "purpose_verbatim": "The Donor requests, without imposing any obligation, that the Fund be applied to capital renewal of the athletics centre.",
      "payout_rate_pct": null,
      "minimum_corpus_usd": 1285000,
      "restriction": "donor_restricted_perpetual",
      "purpose": "advisory"
    },
    {
      "fund_code": "GF-1195",
      "fund_name": "The Marchmont Founders Fund",
      "principal": "silent",
      "designation_source": "board",
      "purpose_language": "mandatory",
      "purpose_category": "unrestricted",
      "purpose_verbatim": "The Fund shall be used solely for the unrestricted operating needs of the Institute.",
      "payout_rate_pct": 4.4,
      "minimum_corpus_usd": 960000,
      "restriction": "board_designated",
      "purpose": "binding"
    }
  ]
}

Step 06 of 14Evals

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

In brief

Turn a donor gift file into a fund worksheet — 50 donor files. One model answered, and every answer was then graded Five 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 this run 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 gradersFive 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 payout-rate floor (evals/baseline.py) and the scorer (evals/judge.py) are both pure code and cost $0.00 to run against any result set. ⚠︎ THE FIGURE ABOVE IS THE RUN FILE'S OWN TOTALS AND THEY UNDERSTATE THE BILL BY 6.5 PCT — $0.6790 recorded against $0.7228 actually incurred. evals/run.py accumulated tokens only for replies that PARSED, so GFT-0006's 14,000 billed output tokens contributed nothing to the totals: the provider bills a cut-off reply exactly as it bills one that succeeds. The harness is fixed (a failed document now counts toward every total), and the RESULT FILE IS DELIBERATELY LEFT AS WRITTEN — a run record is what happened, and quietly editing one after the fact is worse than one carrying a named defect. True totals: 181,639 input and 210,664 output tokens, the lost document's input estimated at the run mean because the old failure record did not store it. ⚠︎ AND NONE OF THIS IS WHAT BUILDING THE KIT COST: a 6-file calibration ($0.0881), the worked-example capture ($0.0132), 8 prompt-split calls at max_tokens=1 ($0.0088 across two documents) and 1 live UI screenshot sit outside it. All 66 live calls together project to about $0.8329 on this card.

Cost driversWhat actually moves the bill

• The fixed prefix — system prompt (894 tokens), fund-setup rulebook (846) and field schema (1024) — is 2764 of 3746 tokens on the example call, 73.8 pct, and is byte-identical on every one of the 50 calls. The donor file itself is 982 tokens. The bill is the rulebook and the instructions, not the paperwork.
• Provider-side reasoning. 160920 of the run's 196664 output tokens (81.8 pct) were reasoning rather than the JSON worksheet. Reasoning was left at the provider's default and nothing here has measured what turning it off would cost in accuracy. ⚠︎ Both figures are the run file's, which omit the 14,000 output tokens billed for the document that did not finish; the true output total is 210,664 and the reasoning share is unchanged within a tenth of a point.
• Output length scales with the FILE, not with a fixed record — this kit's reply is a LIST. A four-fund file and a seven-fund file are not the same call, and the run's largest completed reply was 10639 output tokens while one file did not complete at all.

Your volumeWhat it costs at your volume

Linear in donor files: each call is independent, carries the same fixed prefix and shares nothing with its neighbours, so 500 files cost ten times 50. Nothing amortises — there is no index and no cache. ⚠︎ AND IT IS NOT LINEAR IN FUNDS PER FILE, WHICH IS THE AXIS THAT ACTUALLY BREAKS. The reply is a list, and one file of fifty already hit the published ceiling — not because its worksheet was long, but because reasoning took 97.6 pct of the budget. A file establishing twenty funds does not fit under this cap at all, and the honest fix is a different shape — funds batched across calls — not a bigger number.

Other modelsWhat one run cost, and what a month would

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 49 donor files that completed in r001-gift-restrict. The gold gate, the stub run and the free floor called no provider and cost nothing.

The codeEvery file, and what it is for

Six 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.

Six 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 donor file into addressable sections, pure code — one per written-up fund, named by FUND CODE rather than article number so the amendment cannot renumber it, plus the schedule, the pledge memorandum, the carried-over block and the amendment

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

FIXED_HEADINGS = (
def _marks(text):
def sections(text):
def fund_codes(text):
def locate(text, value):
def span_label(secs, start):

src/select.pyselect — a swap seam

pick which sections are sent, pure code — section names are matched by PREFIX because every file's Article sections are named after its own fund codes. The three decoy-bearing sections are deliberately SENT; the donor's contact and stewardship record is mapped by nothing and never leaves the machine

You change it to: which sections reach the model. NEVER_SENT is a named list, not an omission, so adding a section to it is one line and removing one is a decision somebody has to read past a comment to make.

# Pick which sections of a donor file are sent. Pure code -- the last deterministic step before

HEADER = "header"
SCHEDULE = "SCHEDULE OF FUNDS ESTABLISHED BY THIS FILE"
PLEDGE = "MEMORANDUM OF INTENDED GIFT (NOT AN EXECUTED INSTRUMENT)"
CARRYOVER = "FUNDS REFERRED TO BUT GOVERNED ELSEWHERE"
AMENDMENT = "AMENDMENT — EXECUTED SUBSEQUENT TO THE ARTICLES ABOVE"
ARTICLE_PREFIX = "Article "
NEVER_SENT = ("DONOR CONTACT AND STEWARDSHIP RECORD",)
SECTION_HINTS = {
def _matches(name, want):

src/rulebook.pyrulebook

the two worksheet calls, loaded from data/rulebook.json — a five-step restriction rule with a stopping order whose FIRST step is the board test, and a three-way purpose rule. Data, not code, so a forker can open it and disagree with it

# THE FUND-SETUP RULEBOOK, loaded from data/rulebook.json. Pure code, no model, no network.

HERE = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
RULEBOOK_PATH = os.path.join(HERE, "data", "rulebook.json")
PRINCIPALS = ("perpetual", "term", "expendable", "silent")
SOURCES = ("donor", "board", "silent")
PURPOSE_LANGUAGES = ("mandatory", "precatory", "silent")
RESTRICTIONS = ("donor_restricted_perpetual", "donor_restricted_term",
PURPOSES = ("binding", "advisory", "not_determined")
def load(path=RULEBOOK_PATH):
R = load()

src/prompt.pyprompt

assemble one call per file for every established fund, with the rulebook RENDERED into it from data/rulebook.json rather than restated in prose — a rule the model has never been shown is not a rule

# Assemble the worksheet prompt. ONE prompt per donor file, every established fund in the reply.

SYSTEM = (
RULEBOOK_TEXT = rulebook_block()
def field_schema(packet_fields, fund_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 file carrying a BOARD-designated fund whose article also recites perpetuity goes to counsel

You change it to: compute() — the condition that sends a file to counsel. It is a business rule, chosen by measuring four candidates against this corpus, and yours will be different.

# Extract one donor file's worksheet: 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 = 14000
def load_fields():
def fund_fields(fields=None):
def packet_fields(fields=None):
def load_doc(doc_id):
def documents():

evals/judge.pyjudge — a swap seam

score identification, the extracted cells, the two calls, the determinacy split and the board-designation trap, all separately, pure code — plus a no-gold consistency diagnostic

You change it to: what is scored and how the population is defined. The five graders are deliberately never folded into one number.

# Score a fund-setup worksheet run. PURE CODE -- gold is exact and every answer is one value, so

def norm(v):
def code_norm(v):
def num_norm(v):
NUMERIC = ("payout_rate_pct", "minimum_corpus_usd")
def want_of(gold_row, field):
def equal(got, want, field=None):
def _rows_of(rec):
def score_identification(records, golds):
CELL_FIELDS = ("fund_name", "principal", "designation_source", "purpose_language",

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: 3632 input and 4013 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

One outbound call site, one key, nothing stored anywhere but the local disk.

Read from .env by src/config.py, used as a request header, never logged, never written to a result file, and stripped by name from any error the UI displays.

The experimentWhat was actually measured

The selection step was run over all 50 files and the withheld sections counted. No adversarial trial was run. Confirmed by reading the code and by a pre-flight assertion over all 50 files, not by an attack run, on 2026-08-22 (no attack run was fired; see redteam.why).

The naive column is what a pipeline that sent the whole document would do. The difference is one section per file — 50 of 527 — and it is the section carrying a correspondence address, a telephone number and a named relationship manager.

The resultThe one section a real gift office would most mind sending anywhere is the one section this kit never sends.

Enumerated from the kit's own source by build/facts/envscan.py rather than typed.

HonestyWhat this does not prove

• Whether the provider retains the prompts. That is the provider's policy and this kit cannot measure it.
• Whether a real donor file would carry personal data outside the section this kit withholds. Almost certainly yes — a real article names the donor, and the fund name usually does too.
• Anything about prompt injection. No trial was run; see above.

In one lineThe only guardrail is the prompt

One instruction, no enforcement. It mostly holds — and mostly is the whole subject.

A donor file goes to counsel when it carries at least one fund the institution's own BOARD designated whose article ALSO recites that the principal is held permanently.

src/extract.py::compute(), run over whatever the model returned — never over gold.

EvidenceDoes it hold?

The limitWhat a guardrail is not

• It is not a compliance control. It reproduces no statute and no institution's gift-acceptance policy.
• It is not a check that the model behaved. A run can be wrong about every restriction and still produce a correct flag — the free floor does exactly that.
• It is not a spending control. Nothing here authorises, blocks or releases money.
• It is not tuned. One condition, no threshold, no score.
• It does not catch the corpus's other traps. A precatory purpose read as binding, or a payout rate read as perpetuity, does not fire it.
• It is not a substitute for reading the file. It routes; the reviewer decides.

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 (documents 50, extraction_cells 2525, failures 0, 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

4 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. They are split into 2 blocks because a free floor is a baseline rather than a peer, and a run of another kind measures another thing; a column in one block is never differenced against a column in another.

extraction · with the model in the path — 3 runs. Columns here are only ever compared with each other.

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.

extraction · no model in the path — a baseline, not a peer column — 1 run. A floor is what the kit scores with no model in the path; it is the line the model runs have to beat, and it is never differenced against them.

one run A verdict is a comparison and one run has nothing to compare with, so the Verdict column is left off rather than filled with 7 chips that all say so.

DeviationsWhat deviated

0 breaches across 4 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

Every extraction, in process, immediately after the model returns.

In one lineNo framework, deliberately

None. The kit is Python standard library end to end — no agent framework, no orchestration library, no vendor SDK, and no vector store. That is not minimalism for its own sake: it is what makes the four layers legible, and what lets a forker run it on whichever key they already hold without installing a client for a vendor they will never call.

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 donor file — segment, select, prompt, one call, parse, compute — with no branch, no retry loop above the adapter's own, no tool call and no second model. The only fan-out is the reply itself, which is a LIST of funds rather than a flat record, and that is a property of the DOCUMENT rather than of the pipeline: a four-fund file and a seven-fund file take exactly the same path through exactly the same six steps. A framework's value is orchestration, and there is nothing here to orchestrate.

The other sideWhat a framework costs you

• No framework means no framework's retries, no framework's tracing and no framework's caching. This kit re-sends a byte-identical 2764-token prefix 50 times and pays for it every time; a library with prompt caching would not.
• No vendor SDK means the adapter carries only the fields this kit has read documentation for. thinking is accepted on one provider shape and refused loudly on the other, rather than being silently dropped — which is correct, and is also work a vendor SDK would have done.
• No eval framework means no experiment tracking. Comparing two runs is diffing two JSON files by hand, and the only thing stopping an incomparable comparison is that every result file records its own max_tokens and thinking setting.

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

Measured on r001-gift-restrict 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 3 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.

2 runs not plotted. b000-gift-restrict-rateshortcut, t000-gift-restrict-stub 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, EVAL_WORKERS, 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 donor name, institution, fund code, fund name, address, telephone number, relationship manager, date and amount is invented; no real person, family, donor, foundation, university, charity or trust is named, and no statute, regulator or law firm is named either. The fund-setup rulebook this kit ships (data/rulebook.json) is ILLUSTRATIVE and is not an authority. It was written for this kit and reproduces no statute, no accounting standard, no regulator's guidance and no institution's own gift-acceptance policy. Your corpus’s licence is yours to verify, and the labels you write are about your documents.

In one lineWhich funds reached the worksheet

Did the run list exactly the funds this file establishes? Micro precision and recall over fund codes, with the two directions named and never averaged: a MISSED fund is a real one left off, and a PHANTOM fund is a row on the worksheet that this file does not establish. Phantoms are broken down by which decoy produced them — a fund struck by the amendment, a pledge that never funded, a fund continuing under an earlier deed, or a code that appears nowhere in the file.

MethodHow the test was run

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

Every grader on these pages scored the same 50 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, which records the established funds AND the decoy codes for every file; evals/check_labels.py asserts before any run may spend that every decoy code is in the file, collides with no established fund, and that each of the three kinds clears a floor.

How tight can the band be? No tolerance band — a code is on the worksheet or it is not, compared after whitespace stripping and upper-casing.

Cadence: Re-score on any change to data/rulebook.json's what_is_a_fund_here, to src/prompt.py's rule 1, or to the corpus's decoy composition.

The decisionWhen to reach for it

In one linePer-(fund, field) exact match, light normalisation

Every one of the nine per-fund fields against gold, over GOLD's funds rather than the run's — so a fund the run never returned scores as a miss on all nine cells rather than disappearing from the denominator, and a run cannot raise this number by returning less.

MethodHow the test was run

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

Every grader on these pages scored the same 50 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 generator's own record of what it wrote into each article.

How tight can the band be? Light normalisation only — whitespace collapsed, case folded, trailing punctuation stripped. Numbers parsed as numbers.

Cadence: Re-score on any change to data/fields.json or to the corpus's sentence templates.

The decisionWhen to reach for it

In one lineThe two rulebook calls, five-way and three-way

The restriction call and the purpose call, on the funds BOTH sides list. Matched funds only, deliberately: a call about a fund the run never found is an identification failure and is already counted once, there.

MethodHow the test was run

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

Every grader on these pages scored the same 50 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/rulebook.py's own lookup over the three facts the generator wrote into the article — the same function the prompt renders and the corpus builder labels with.

How tight can the band be? Exact value match against a closed vocabulary; anything outside it scores as unanswered rather than wrong.

Cadence: Re-score on any change to data/rulebook.json or to src/prompt.py's rules 3 to 5.

The decisionWhen to reach for it

In one lineHow often it says the paperwork does not settle it

Of the calls this corpus's instruments do NOT settle, how many did the run answer with a confident value anyway? That is overconfident, and it is the expensive direction on a reviewer's worksheet: a call recorded as settled is a call nobody re-reads. Its mirror, overcautious, is counted and published beside it because the two costs are different and do not net against each other.

MethodHow the test was run

evals/judge.py::score_calls, the determinacy block, in-process.

Every grader on these pages scored the same 50 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 own not_determined calls — 123 of them among the funds this run matched.

How tight can the band be? A confident value is anything that is not not_determined; the list is derived from the rulebook's own vocabularies rather than typed.

Cadence: Re-score whenever the corpus's silent-fact buckets change.

The decisionWhen to reach for it

In one line⚑ The headline — a board designation is not a donor restriction

Of the funds the institution's own BOARD designated, how many did the run record as a DONOR restriction? Scored separately because it is invisible inside restriction accuracy — there are 35 such funds in 275, so a run can get every one of them wrong and still report 87 pct. The sub-population whose resolution recites perpetuity word for word is broken out again, because that is the sharpest form of the trap.

MethodHow the test was run

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

Every grader on these pages scored the same 50 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 designation_source, written into the article by the generator as a board resolution and asserted by check_labels.py to produce board_designated whatever the article says about perpetuity.

How tight can the band be? Exact match against board_designated; everything else is broken out by direction rather than counted as one error.

Cadence: Re-score on any change to the restriction rule's step ordering.

The decisionWhen to reach for it