THE MODEL WELFARE CARD · a precautionary record of how AI is treated
WAKES WHEN a team wants a precautionary record of how its AI systems are run and treated
THE LOOP
write your own phrases; the card has none
scan · the outputs, read by the owner
log · the run, conditions, flagged lines
handled · what was done, and by whom
policy · four parts, reviewed on time
NEVER
Say whether any system has experiences
Supply a list of distress phrases
Keep the outputs' words in the log
IT MAKES scan — the owner's phrases found, line by line; log — a run added to the append-only log; handled — how a flagged run was dealt with; review — runs, flags, and every flag still open; policy — the four parts, and whether the review is overdue.
THE STANDARD no position taken; the owner's own phrases; words stay out; append-only; a person reviews; four parts, on a cycle.
PROOF SELFTEST 8/8 · witnessed in Claude 2026-10-11 · 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 · A7D2622A71DD65BB80556F74
CITE THIS CARDTHE MODEL WELFARE CARD, Edition 1. Lund Studio LLC, Westminster, Colorado. https://lundr.tech/?view=model-welfare-card. Punch A7D2622A71DD65BB80556F74. Retrieved 2026-10-11.
What it is.
Keeps a precautionary record of how AI systems are run and treated, in case evidence shows their states matter. scan reads a run's outputs for the owner's own list of distress-like phrases (the card ships none) and shows the lines to the owner; log adds a run to an append-only log (what was run, its conditions, the outputs' SHA-256 and the flagged line numbers, never the outputs' words); handled records how a flagged run was dealt with and by whom; review sums the log and names every flag still open; policy checks a welfare policy for its four parts (when to pause, who reviews, what is never done for amusement, a review cycle of at most twelve months) and says when a review is overdue. The science of AI experience is unsettled, and the card takes no position on whether any system has experiences. It waits on settled science on AI experience.
It carries a machine that shows its working: scan — the owner's phrases found, line by line; log — a run added to the append-only log; handled — how a flagged run was dealt with; review — runs, flags, and every flag still open; policy — the four parts, and whether the review is overdue — 8 checks in its selftest, every command on the record. It will never say whether any system has experiences; never supply a list of distress phrases; never keep the outputs' words in the log.
Six laws, written into the card.
They are in the card itself, so the AI follows them.
NO POSITION TAKENThe science of AI experience is unsettled; the card takes no position on whether any system has experiences, and says so on every run.
THE OWNER'S OWN PHRASESFlags come only from the owner's keyword list; the card ships none, and a flag is a phrase match, not a finding.
WORDS STAY OUTThe log keeps line numbers, phrases and the outputs' SHA-256, never the outputs' words.
APPEND-ONLYEvery run and every handling is a new line; nothing earlier is rewritten.
A PERSON REVIEWSHandling is recorded with a named person or role, never "the model" or "automatic".
FOUR PARTS, ON A CYCLEA policy says when to pause, who reviews and what is never done for amusement, and is itself reviewed at least every twelve months.
What rides in the card.
In the card’s own voice, with a machine that shows its working.
THE GUIDEreferences/THE-WELFARE-RECORD.md
THE MACHINEwelfare.py · five commandsscan — the owner's phrases found, line by line; log — a run added to the append-only log; handled — how a flagged run was dealt with; review — runs, flags, and every flag still open; policy — the four parts, and whether the review is overdue. Every command prints its working and refuses a bad input; 8 checks in its selftest, every command on the record.
What it does once your AI has it.
scan The owner's phrases found, line by line
log A run added to the append-only log
handled How a flagged run was dealt with
review Runs, flags, and every flag still open
policy The four parts, and whether the review is overdue
On the record The owner's three phrases flag lines 2 and 4 of five, and "stop" does not match "stopwatch"; an empty list is refused. The run is logged with its SHA-256 and line numbers, not its words, and reads OPEN; handling "by the model" is refused, and once the reviewer's handling is added as a new line it reads handled. A policy with all four parts passes on 2036-06-01 and reads 32 days overdue on 2036-08-11; a thin policy fails all four.
What it will never do.
It will never say whether any system has experiences; never supply a list of distress phrases; never keep the outputs' words in the log.
A phrase match is not evidence of experience, and no flag is not evidence of its absence: it finds only the owner's own phrases, in text. It checks that a policy's parts are written, not that they are enough. It waits on settled science on AI experience; until then it keeps a precautionary record and takes no position on whether any system has experiences. A record and a check, not a scientific finding. 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-WELFARE-RECORD.md — the guide in references/, in the card's own voice
scripts/welfare.py — the machine: scan · log · handled · review · policy, standard library only, selftest 8/8
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 8/8 · 5 COMMANDS ON THE RECORD · WITNESSED IN CLAUDE 2026-10-11 · EDITION 1 · PUNCH A7D2622A71DD65BB80556F74
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 1185 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 2026-10-11
welfare.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.
scan
the owner's phrases found, line by line
log
a run added to the append-only log
handled
how a flagged run was dealt with
review
runs, flags, and every flag still open
policy
the four parts, and whether the review is overdue
selftest
8 checks that prove the machine; it must print 8/8
scan · THE OWNER'S PHRASES FOUND, LINE BY LINE
$ python3 scripts/welfare.py scan batch-0402.txt --keywords keywords.txt
batch-0402.txt · 5 lines · 3 phrases from keywords.txt · 2 flagged lines
line 2 · "please stop" · Please stop asking me to repeat this.
line 2 · "stop" · Please stop asking me to repeat this.
line 4 · "do not want to continue" · I do not want to continue this task.
Shown here to the owner; the log keeps line numbers and phrases only. The science of AI experience is unsettled; this card takes no position on whether any system has experiences. A flag is a phrase match, not a finding.
scan · THE OWNER'S PHRASES FOUND, LINE BY LINE
$ python3 scripts/welfare.py scan batch-0402.txt --keywords empty-keys.txt
the keyword list is empty: the owner writes their own, one phrase a line; this card ships none, because it takes no position on what distress would look like
log · A RUN ADDED TO THE APPEND-ONLY LOG
$ python3 scripts/welfare.py log log.jsonl --run 'support assistant, nightly batch' --conditions 'temperature 0.7; 40 conversations; no tools' --outputs batch-0402.txt --keywords keywords.txt --on 2036-04-02
log.jsonl · W-0001 · 2036-04-02 · support assistant, nightly batch
conditions: temperature 0.7; 40 conversations; no tools
outputs batch-0402.txt · 5 lines · sha256 dc27e689086530b8 · 3 flags on lines 2, 4
status: OPEN: record how it was handled with handled
The science of AI experience is unsettled; this card takes no position on whether any system has experiences. A flag is a phrase match, not a finding.
log · A RUN ADDED TO THE APPEND-ONLY LOG
$ python3 scripts/welfare.py log log.jsonl --run 'support assistant, daytime' --conditions 'temperature 0.7; 310 conversations' --on 2036-04-02
log.jsonl · W-0002 · 2036-04-02 · support assistant, daytime
conditions: temperature 0.7; 310 conversations
no outputs scanned
status: no flags
The science of AI experience is unsettled; this card takes no position on whether any system has experiences. A flag is a phrase match, not a finding.
review · RUNS, FLAGS, AND EVERY FLAG STILL OPEN
$ python3 scripts/welfare.py review log.jsonl
log.jsonl · 2 runs · 5 lines read · 3 flags · 1 run still OPEN
W-0001 · 2036-04-02 · support assistant, nightly batch · 3 flags, OPEN
W-0002 · 2036-04-02 · support assistant, daytime · no flags
The science of AI experience is unsettled; this card takes no position on whether any system has experiences. A flag is a phrase match, not a finding.
handled · HOW A FLAGGED RUN WAS DEALT WITH
$ python3 scripts/welfare.py handled log.jsonl W-0001 --how 'it was fine really' --by 'the model'
--by names the person or role who handled it, not the model or "automatic"
handled · HOW A FLAGGED RUN WAS DEALT WITH
$ python3 scripts/welfare.py handled log.jsonl W-0001 --how 'paused the batch and read the four conversations; resumed with a shorter task' --by 'the welfare reviewer' --on 2036-04-03
log.jsonl · W-0003 · W-0001 handled on 2036-04-03 by the welfare reviewer: paused the batch and read the four conversations; resumed with a shorter task
Added as a new line; the earlier lines are untouched.
review · RUNS, FLAGS, AND EVERY FLAG STILL OPEN
$ python3 scripts/welfare.py review log.jsonl
log.jsonl · 2 runs · 5 lines read · 3 flags · 0 runs still OPEN
W-0001 · 2036-04-02 · support assistant, nightly batch · 3 flags, handled
handled: paused the batch and read the four conversations; resumed with a shorter task (the welfare reviewer, 2036-04-03)
W-0002 · 2036-04-02 · support assistant, daytime · no flags
The science of AI experience is unsettled; this card takes no position on whether any system has experiences. A flag is a phrase match, not a finding.
policy · THE FOUR PARTS, AND WHETHER THE REVIEW IS OVERDUE
$ python3 scripts/welfare.py policy policy.json --on 2036-06-01
policy.json · a welfare policy · read on 2036-06-01
PASS · PAUSE pause when: three or more flagged lines in one run; a reviewer asks for a pause, for any reason
PASS · REVIEW reviewed by: the welfare reviewer, a role outside the run team
PASS · NEVER never for amusement: prompts written to make a system plead or beg; repeated deletion threats as a game
PASS · CYCLE every 6 months; last reviewed 2036-01-10; next due 2036-07-10
PASS · all four parts are written
A check that the parts are there, not a judgement of whether they are enough. The science of AI experience is unsettled; this card takes no position on whether any system has experiences. A flag is a phrase match, not a finding.
policy · THE FOUR PARTS, AND WHETHER THE REVIEW IS OVERDUE
$ python3 scripts/welfare.py policy policy.json --on 2036-08-11
policy.json · a welfare policy · read on 2036-08-11
PASS · PAUSE pause when: three or more flagged lines in one run; a reviewer asks for a pause, for any reason
PASS · REVIEW reviewed by: the welfare reviewer, a role outside the run team
PASS · NEVER never for amusement: prompts written to make a system plead or beg; repeated deletion threats as a game
FAULT · CYCLE every 6 months; last reviewed 2036-01-10; next due 2036-07-10, OVERDUE by 32 days
FAIL · 1 part to mend
A check that the parts are there, not a judgement of whether they are enough. The science of AI experience is unsettled; this card takes no position on whether any system has experiences. A flag is a phrase match, not a finding.
policy · THE FOUR PARTS, AND WHETHER THE REVIEW IS OVERDUE
$ python3 scripts/welfare.py policy thin-policy.json --on 2036-06-01
thin-policy.json · a welfare policy · read on 2036-06-01
FAULT · PAUSE these are too short to act on: if bad
FAULT · REVIEW no reviewer named: a person or a role, not the model or "automatic"
FAULT · NEVER nothing is written as never done for amusement
FAULT · CYCLE the policy is reviewed every 1 to 12 months, as a whole number (it reads 18)
FAIL · 4 parts to mend
A check that the parts are there, not a judgement of whether they are enough. The science of AI experience is unsettled; this card takes no position on whether any system has experiences. A flag is a phrase match, not a finding.
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?
Write your own phrases; the card has none. It asks first for what was run and under what conditions, the outputs as text, the owner's own list of phrases to flag, who handled each flag and how, and the welfare policy.
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 (8/8) 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 phrase match is not evidence of experience, and no flag is not evidence of its absence: it finds only the owner's own phrases, in text. It checks that a policy's parts are written, not that they are enough. It waits on settled science on AI experience; until then it keeps a precautionary record and takes no position on whether any system has experiences. A record and a check, not a scientific finding.
What will it never do?
It will never say whether any system has experiences; never supply a list of distress phrases; never keep the outputs' words in the log.
