THE LIFT CARD · Runs the same tasks with a skill and without it, grades both the same way, and says if the skill helped.
WAKES WHEN Someone asks if a skill really helps · A buyer wants before-and-after proof · Two models are compared on one skill
THE LOOP
Write tasks with must and must_not checks.
Plan the runs with `plan`.
Run every prompt with and without the skill, fresh session each time.
Score with `score` and read the verdict.
Show the numbers and one example each way.
NEVER
Grades by opinion
Counts missing answers
Hides a loss or a small sample
IT MAKES plan — write the prompts and the run sheet; grade — grade every saved answer, write GRADES.json; score — pair with/without and give the verdict.
THE STANDARD same tasks; checks are written before the answers are seen; a missing answer is never counted; a small sample is said in the first line; harm is reported as plainly as lift; verify by execution; say the limit out loud in the work; nothing hidden; the person decides.
PROOF SELFTEST 14/14 · witnessed 2026-10-13 on material and live on Claude · Edition 1.
THE SEAL 96 of 100 · LEGENDARY · WITNESSED · Trigger 4 · Machinery 5 · Law 5 · Portability 5 · Proof 5
IT CAN BE WRONG an AI reading a text file: it can be wrong and can invent a number, a rule, a name or a date; check what matters against your own records and the source; the risk is yours. LICENSE-NOTE.md rides in the packet and is the licence.
THE PUNCH · SHA-256 OF SKILL.MD · FC24995C116539CF2FCEE4B4
CITE THIS CARDTHE LIFT CARD, Edition 1. Lund Studio LLC, Westminster, Colorado. https://lundr.tech/?view=lift-card. Punch FC24995C116539CF2FCEE4B4. Retrieved 2026-10-13.
What it is.
Measures whether a skill makes the work better: plans the same tasks with the skill and without it, grades every answer against the same checks, and reports wins, losses, ties, pass rates and a sign test.
Use when someone asks if a skill or prompt actually helps, wants before-and-after proof for a buyer, or is comparing two models on a skill. Do not use to test whether the model picks the skill (THE TRIGGER TEST CARD does that) or to grade work by opinion.
It carries a machine that shows its working: plan — write the prompts and the run sheet; grade — grade every saved answer, write GRADES.json; score — pair with/without and give the verdict — 14 checks in its selftest. It will never grades by opinion; never counts missing answers; never hides a loss or a small sample.
Nine laws, written into the card.
They are in the card itself, so the AI follows them.
SAME TASKSSame tasks, same checks, both sides.
CHECKS ARE WRITTEN BEFORE THE ANSWERS ARE SEENChecks are written before the answers are seen.
A MISSING ANSWER IS NEVER COUNTEDA missing answer is never counted.
A SMALL SAMPLE IS SAID IN THE FIRST LINEA small sample is said in the first line.
HARM IS REPORTED AS PLAINLY AS LIFTHarm is reported as plainly as lift.
VERIFY BY EXECUTIONVerify by execution, not by reasoning: run the machine and read what it prints before saying a thing is done.
SAY THE LIMIT OUT LOUD IN THE WORKSay the limit out loud in the work: every result names what it did not check.
NOTHING HIDDENNothing hidden: two readers, the person and the AI, see the same words.
THE PERSON DECIDESThe person decides.
What rides in the card.
In the card’s own voice, with a machine that shows its working.
THE GUIDEreferences/WRITING-CHECKS.md
THE MACHINElift.py · three commandsplan — write the prompts and the run sheet; grade — grade every saved answer, write GRADES.json; score — pair with/without and give the verdict. Every command prints its working and refuses a bad input; 14 checks in its selftest.
What it does once your AI has it.
plan Write the prompts and the run sheet
grade Grade every saved answer, write GRADES.json
score Pair with/without and give the verdict
On the record a real lift run: six donor-letter tasks run on Claude Sonnet with THE DONOR CARD installed and without it (12 live sessions), scored by this machine — real material (not made up for the test). As first scored: 6 of 6 passed with the card, 5 of 6 without, one win. Read by hand, the one win did not hold: the plain answer's only dollar sign was in a correct note about Form 1098-C, not a value on the truck. The check was too blunt. - Regraded with a check aimed at a value placed on the property: 0 wins, 0 losses, 6 ties, NO CLEAR LIFT, small sample. Both reports are kept (LIFT.md and LIFT-REGRADED.md) with the reason. - The machine did its job: it refused to call one win a lift (p = 1.0, small sample) before anyone read the answers. Live on Claude Sonnet, 2026-10-13: 7 of 7 passed; the card fired on 2 of 3 trigger cases and on none of its hand-off cases.
What it will never do.
It will never grades by opinion; never counts missing answers; never hides a loss or a small sample.
Regex checks catch what can be written down. They cannot judge taste, tone or truth beyond the checks given, and the machine does not call any model itself. It is an AI reading a text file: it can be wrong, so check what matters against your own records and the source.
What’s in the buy.
CARD.md — the whole card: its boot, laws, loop, when it says no, its seals, ledger, term and score · SKILL.md — the door your AI reads first · START-HERE.md — the first five minutes, for you
WRITING-CHECKS.md — the guide in references/, in the card's own voice
scripts/lift.py — the machine: plan · grade · score, standard library only, no network, selftest 14/14
WITNESS.md — the runs on real material and live on Claude, with what was not run · SELFTEST.md — the machine's own test, as run at the cut · MANIFEST.md, every file hashed · CHANGELOG.md
tests/evals.json + EVALS.md — the cases it must take, hand off and refuse
MY-LEDGER.md — the card's memory, on your own machine · TIN.md, the nine lines on the tin · LICENSE-NOTE.md
card.svg + card-machine.svg + card.json · card.html, the card as one offline file · doors/ — the rules files for Cursor, Codex, Gemini, Copilot and Claude Code, the system text for any other AI, a local-model file and an MCP door
PROOF · SELFTEST 14/14 · WITNESSED ON MATERIAL · LIVE ON CLAUDE SONNET 2026-10-13 · 7 OF 7 PASSED · EDITION 1 · PUNCH FC24995C116539CF2FCEE4B4
ON EVERY DOOR · A CARD FOR YOUR AI, A FILE FOR YOUR COMPUTER
In Claude: turn on code execution under Settings › Capabilities, then Customize › Skills › + › Create skill, and upload the zip as it is — START-HERE.md in the box walks you through it, and tells you how to keep the ledger. In ChatGPT, Gemini and the rest: a Project, a GPT or a Gem with the card’s files attached and the short door in the box pasted as its instructions. The words run anywhere an AI reads Markdown; the machines run wherever the host runs Python. Witnessed in Claude only. The host marks →
Take it.
Nº 1193 PROPOSED — LAUNCHING SOON
Sold as is under the term printed. A card instructs an AI you operate; results depend on your model, your files and your judgement. Not medical, legal, tax or financial advice, never the person’s boss, and not a licence to practise. The terms →
ONE PAYMENT · NOTHING RENEWS ITSELF
LAUNCHING SOON
One payment. The file is yours to keep and keeps working at its edition; new editions issued in the year after you buy reach you by email as replacement files. Tax is added at checkout where it applies. Nothing auto-renews.
A TERM, NOT A SUBSCRIPTION: NOTHING RENEWS ITSELF AND NOBODY IS CHARGED AGAIN. WHEN A TERM ENDS THE FILES STAY ON YOUR COMPUTER AND STAY READABLE — FREEZE, NOT LOSS.
THE MACHINE · IN THE OPEN · FROM THE CARD’S OWN WITNESS.MD
lift.py, every command showing its working.
A small Python program, standard library only. Your AI runs it; you can too. The blocks below are pasted from the card’s own WITNESS.md, as printed.
plan
write the prompts and the run sheet
grade
grade every saved answer, write GRADES.json
score
pair with/without and give the verdict
selftest
14 checks that prove the machine; it must print 14/14
THE COMMANDS, IN ORDER, ON A REAL LIFT RUN: SIX DONOR-LETTER TASKS RUN ON CLAUDE SONNET WITH THE DONOR CARD INSTALLED AND WITHOUT IT (12 LIVE SESSIONS), SCORED BY THIS MACHINE
VERDICT: NO CLEAR LIFT · with the skill 100.0% pass, without 83.3% (+16.7 points) · 1 wins, 0 losses, 5 ties · p = 1.0000 · small sample
VERDICT: NO CLEAR LIFT · with the skill 100.0% pass, without 100.0% (+0.0 points) · 0 wins, 0 losses, 6 ties · p = 1.0000 · small sample
THE RUNS ARE PASTED FROM THE CARD’S OWN WITNESS.MD AS PRINTED. FIXTURES MARKED FICTIONAL ARE THE STUDIO’S OWN EXAMPLES, INVENTED FOR THE RECORD; NO PRIVATE PERSON IS NAMED.
QUESTIONS AT THE TABLE
Asked before the wire.
What does it need from me?
A task list with written checks, and answers from real runs. Write tasks with must and must_not checks, and it goes from there.
Does it send anything anywhere?
No. The machine is a standard-library Python program that runs on your own computer and sends nothing anywhere; the words run wherever an AI reads Markdown. No server, no account, no key.
What is on the record?
The self-test (14/14), a run on real or invented material with every command and verdict pasted into WITNESS.md as printed, and live runs on Claude Sonnet: 7 of 7 passed; the card fired on 2 of 3 trigger cases and on none of its hand-off cases.; the punch on the face is the SHA-256 of SKILL.md, so a downloaded copy can be checked against this room.
Can it be wrong?
Yes — it is an AI reading a text file. Regex checks catch what can be written down. They cannot judge taste, tone or truth beyond the checks given, and the machine does not call any model itself.
What will it never do?
It will never grades by opinion; never counts missing answers; never hides a loss or a small sample.
