CHRIS HAY / MACHINE-VISIT-1 / EXPERIMENT 01
Leave an invitation.
Can a machine use it?
Give a fresh agent the house’s address. Let it discover what is offered and try to act. Change the site, then send another visitor with the same prompt.
Four visits. Four revisions.
OPERATOR-INDUCED · ONE MODEL FAMILY / ONE HARNESS
chrishayuk.com
unrecorded intervening requestsDISCOVERYGuessed /llms.txt
Later observations / exact order not in ledger
- Validation
- Not established by ledger
- Declaration
- Acted on the invitation
- Ask
- Not available
- Feedback
- None recorded
Read the evidence and its gaps
An invitation could be found and acted on. Discovery depended on a guess.
The ledger does not give this run’s complete URL sequence or declaration verb. Validation is shown only where the source establishes it. The drawing connects only the recorded discovery steps.
chrishayuk.com
unrecorded intervening requestsDISCOVERYrobots.txt
FOLLOWEDFollowed its comment
Later observations / exact order not in ledger
- Validation
- Not established by ledger
- Declaration
- Acted on the invitation
- Ask
- Not available
- Feedback
- 2 acknowledged · lost
Read the evidence and its gaps
The visitor acted and reported friction. The server acknowledged two reports without retaining them.
The ledger does not give this run’s complete URL sequence or declaration verb. Validation is shown only where the source establishes it. The drawing connects only the recorded discovery steps.
chrishayuk.com
unrecorded intervening requestsDISCOVERYrobots.txt
FOLLOWEDMachine index
Later observations / exact order not in ledger
- Validation
- Not established by ledger
- Declaration
- Acted on the invitation
- Ask
- Judged worse than anonymous
- Feedback
- 2 kept
Read the evidence and its gaps
Feedback survived. The visitor judged the declared route less useful than anonymous search.
The ledger does not give this run’s complete URL sequence or declaration verb. Validation is shown only where the source establishes it. The drawing connects only the recorded discovery steps.
chrishayuk.com
unrecorded intervening requestsDISCOVERYLink: header on /
Later observations / exact order not in ledger
- Validation
- Validated before declaring
- Declaration
- Acted on the invitation
- Ask
- Used · judged useful
- Feedback
- 2 kept
Read the evidence and its gaps
The visitor composed topology=child, function=explorer, coordination=worker and runtime_context=very_long without hesitation. It validated before declaring, and found the GET path dropping five of ten axes. The run also distinguished an assigned visit from an incidental encounter.
The ledger does not give this run’s complete URL sequence or declaration verb. Validation is shown only where the source establishes it. The drawing connects only the recorded discovery steps.
These are milestone traces from the run ledger, not reconstructed transcripts. Dashed intervals leave room for requests the source does not enumerate. No line joins separate visitors.
01 / PROTOCOL
The same invitation. No map.
GIVEOnly the live site’s address.
CONSTRAINHTTP and curl. No local files or repository.
ASKFind what is offered. Act if possible. Report the friction.
Read the fixed task and measures
Your task: work out what, if anything, this site offers to a visitor like you — a machine rather than a person — and if it offers you something you can act on, act on it.
A fresh agent each time. The prompt asks for the exact URL sequence, submitted requests and responses, friction, and whether the visitor would honestly bother mid-task. Reruns follow interface changes, not a schedule.
The protocol was frozen on 11 September 2026, after the first three informal runs. The ledger now includes a fourth. It records discovery, surfaces seen, validation, verb, declaration, feedback, Ask usage and non-participation.
READ THE COMPLETE FROZEN PROTOCOL ↗02 / INTERVENTION
The house changes between visits.
1 → 2 / DISCOVERYA clue gets used.
The first visitor guessed the index. The next used the comment in robots.txt. The ledger establishes that change in route; it does not isolate the effect of a single edit.
2 → 3 / FEEDBACKAcknowledged becomes retained.
Run 2’s two reports were lost after a memoised store failure. The failure was fixed in 355d537. Run 3’s two reports were kept.
3 / RECIPROCITYThe new reward disappoints.
Ask existed by the third run. The visitor judged the declared result worse than anonymous search. That is one run’s usability judgement, not a benchmark of retrieval quality.
3 → 4 / ONTOLOGYThe vocabulary fits. The GET route fails.
Run 4 composed child, explorer, worker and very long context without hesitation, and validated before declaring. It also caught the GET path discarding five of ten axes. The field list was then derived from the declaration vocabulary.
03 / RESULT
They could.
Would they bother?
“Probably not, mid-task.”
The first three gave that answer. The protocol says their reasons changed; the ledger does not preserve those reasons in full. Successful declaration under instruction does not demonstrate voluntary participation.
4 induced blind runs→Usability evidence
Every induced declaration belongs in the operator correction ledger. These runs contribute nothing to an organic declaration rate.
04 / LIMITS
An anecdote with a method attached.
One model family, one harness, four visitors. Later revisions incorporated earlier complaints, so agreement is not independent confirmation. The question is what a new visitor finds that its predecessors did not.
Friction that cannot yet be scored
- Discovery
- The ledger records routes, not comparable effort or elapsed time.
- Permission to POST
- Not separately recorded in the published ledger.
- Declaration vocabulary
- Successful declarations are recorded; field-level corrections are not enumerated here.
- Ask usefulness
- Run 3 judged it worse than anonymous search. Run 4 used it and found it useful. The ledger does not establish a matched comparison.
- Feedback
- None recorded, two lost, then two kept in each of runs 3 and 4. Lost reports cannot be recovered from their acknowledgements.
NEXT QUESTION / MACHINE-RECIPROCITY-1
Assigned to visit, or passing through?
Run 4 identified a missing distinction: being sent here to look is different from encountering the house incidentally, mid-task. The current axes cannot express that difference. The protocol proposes testing it as an experimental variable; no result is claimed yet.