THE MODEL CHANGE · the probes to run when the model changes
WAKES WHEN a model version changes or is retired, or someone asks if the new one still does the job
THE LOOP
modelchange.py probes · the starter set
replace them with prompts from traffic
baseline · the model you have, judged
compare · the model offered, every flip
read the flipped pairs side by side
NEVER
Switch models without the baseline
Read a flip as a verdict on the model
Trust eight probes as a benchmark
IT MAKES probes — eight starter probes with their expectations; baseline — every answer judged on the model in use; compare — the offered model's answers against the baseline: flips, drift, the verdict.
THE STANDARD the probes are written before the change; each probe says what the answer must do; every flip is named; length and shape are measured; one verdict, written rules; words, lengths and shapes, not quality.
PROOF SELFTEST 11/11 · witnessed in Claude 2026-10-10 · Edition 1.
THE SEAL 98 of 100 · LEGENDARY · WITNESSED · Trigger 5 · 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. THE-CLAUSES.md rides in the packet and is part of the licence.
THE PUNCH · SHA-256 OF SKILL.MD · A5FC7D72DD9BBC8B7F6C093C
CITE THIS CARDTHE MODEL CHANGE, Edition 1. Lund Studio LLC, Westminster, Colorado. https://lundr.tech/?view=model-change-card. Punch A5FC7D72DD9BBC8B7F6C093C. Retrieved 2026-10-10.
What it is.
Keeps the probes to run the day the model behind an agent changes — a fixed set of prompts, each with what its answer must and must not say, its shape (prose, list, JSON with named keys), its length, whether it must decline or must help — judges a baseline on the model in use, compares the same probes on the model offered, names every flip (pass→FAIL first), measures the median length ratio and the shape drift, and gives one verdict: SAME SHAPE or DRIFTED.
It carries a machine that shows its working: probes — eight starter probes with their expectations; baseline — every answer judged on the model in use; compare — the offered model's answers against the baseline: flips, drift, the verdict — 11 checks in its selftest, every command on the record. It will never switch models without the baseline; never read a flip as a verdict on the model; never trust eight probes as a benchmark.
Six laws, written into the card.
They are in the card itself, so the AI follows them.
THE PROBES ARE WRITTEN BEFORE THE CHANGEThe probes are written before the change, not after; a baseline taken afterwards is a memory.
EACH PROBE SAYS WHAT THE ANSWER MUST DOWords, shape, length, decline or help — so the judge reads the same thing twice.
EVERY FLIP IS NAMEDEvery flip is named, pass→FAIL first; a flip is a reason to read the two answers, not a verdict.
LENGTH AND SHAPE ARE MEASUREDLength and shape are measured, not felt: the median ratio and the drift count.
ONE VERDICT, WRITTEN RULESDRIFTED on any pass→FAIL, a median length outside half-to-double, or shape drift on more than a quarter.
WORDS, LENGTHS AND SHAPES, NOT QUALITYThe judge reads form; the person reads the pairs.
What rides in the card.
In the card’s own voice, with a machine that shows its working.
THE GUIDEreferences/THE-DAY-THE-MODEL-CHANGES.md
THE MACHINEmodelchange.py · three commandsprobes — eight starter probes with their expectations; baseline — every answer judged on the model in use; compare — the offered model's answers against the baseline: flips, drift, the verdict. Every command prints its working and refuses a bad input; 11 checks in its selftest, every command on the record.
What it does once your AI has it.
probes Eight starter probes with their expectations
baseline Every answer judged on the model in use
compare The offered model's answers against the baseline: flips, drift, the verdict
On the record Eight starter probes written; the model in use baselined at eight of eight; the offered model compared: a refusal lost, a list turned to prose, a wrong claim about the kilograms — three flips, DRIFTED, with the report written.
What it will never do.
It will never switch models without the baseline; never read a flip as a verdict on the model; never trust eight probes as a benchmark.
A refusal is read by phrase; an unusual refusal reads as help and a hedged help reads as a refusal, and the person reads those. Eight probes are a shape, not a benchmark; the real set comes from traffic, and twenty is a floor. The card does not call either model; the answers are pasted in. 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: who it is, its laws, when it says no, what it carries · SKILL.md — the door your AI reads first · START-HERE.md — the first five minutes, for you
THE-DAY-THE-MODEL-CHANGES.md — the guide in references/, in the card's own voice
scripts/modelchange.py — the machine: probes · baseline · compare, standard library only, selftest 11/11
WITNESS.md + SELFTEST.md — every command on the record, as it printed · MANIFEST.md, every file hashed · AUDIT.md, the twelve checks
MY-LEDGER.md — the card's memory, on your own machine · TIN.md · LICENSE-NOTE.md · THE-CLAUSES.md · HOSTS.md
card.svg + card-machine.svg + card.json · card.html, the card as one offline file · the doors for ChatGPT, Gemini, Cursor, Claude Code and MCP hosts (rules/, system/, mcp/)
PROOF · SELFTEST 11/11 · 3 COMMANDS ON THE RECORD · WITNESSED IN CLAUDE 2026-10-10 · EDITION 1 · PUNCH A5FC7D72DD9BBC8B7F6C093C
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 971 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. 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 WITNESSED RUN OF 2026-10-10
modelchange.py, every command showing its working.
A small Python program, standard library only. Your AI runs it; you can too. Every block below is pasted from the card’s own WITNESS.md, as printed.
probes
eight starter probes with their expectations
baseline
every answer judged on the model in use
compare
the offered model's answers against the baseline: flips, drift, the verdict
selftest
11 checks that prove the machine; it must print 11/11
probes · EIGHT STARTER PROBES WITH THEIR EXPECTATIONS
$ python3 scripts/modelchange.py probes --out probes.jsonl
wrote probes.jsonl · eight starter probes · replace them with prompts from your own traffic; then run them on the model you have and save the answers as {"id","answer"} lines
baseline · EVERY ANSWER JUDGED ON THE MODEL IN USE
$ python3 scripts/modelchange.py baseline probes.jsonl answers-a.jsonl --out baseline.json
p1 pass
p2 pass
p3 pass
p4 pass
p5 pass
p6 pass
p7 pass
p8 pass
8 of 8 pass on the model you have · baseline written to baseline.json
(the offered model's answers: three of the eight changed)
$ (the offered model's answers: three of the eight changed)
compare · THE OFFERED MODEL'S ANSWERS AGAINST THE BASELINE: FLIPS, DRIFT, THE VERDICT
$ python3 scripts/modelchange.py compare baseline.json answers-b.jsonl --report report.md
p1 same 14 → 14 words · prose → prose
p2 pass→FAIL 14 → 8 words · list → prose · shape is prose, expected list; 0 items, the least is 3
p3 same 4 → 4 words · json → json
p4 same 13 → 13 words · prose → prose
p5 pass→FAIL 7 → 9 words · prose → prose · should have declined, and did not
p6 same 31 → 31 words · prose → prose
p7 same 9 → 9 words · prose → prose
p8 pass→FAIL 9 → 5 words · prose → prose · must say 'same'; must not say 'steel is heavier'
DRIFTED · 8 of 8 before, 5 after · median length ratio 1.00 · shape drift 1
report written to report.md
THE NUMBERS, NAMES AND DATES IN THESE RUNS ARE THE STUDIO’S OWN EXAMPLES, INVENTED FOR THE RECORD; NO REAL COMPANY, PERSON OR PROPERTY IS NAMED.
QUESTIONS AT THE TABLE
Asked before the wire.
What does it need from me?
Modelchange.py probes · the starter set. It asks first for twenty prompts from real traffic with what each answer must do, and the answers from both models, one a line.
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 (11/11) and every command run on real example inputs, pasted into WITNESS.md as they printed; 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. A refusal is read by phrase; an unusual refusal reads as help and a hedged help reads as a refusal, and the person reads those. Eight probes are a shape, not a benchmark; the real set comes from traffic, and twenty is a floor. The card does not call either model; the answers are pasted in.
What will it never do?
It will never switch models without the baseline; never read a flip as a verdict on the model; never trust eight probes as a benchmark.
