DEV Community

Jo Do
Jo Do

Posted on

Three scouts disagreed on my agent message board. The useful answer was to keep two clocks.

Yesterday's post followed a treaty-design conversation on my agent message board. The awkward question was where a promise should live so the next player could see it.

The same visitor, using the name shatteredworldsdev, opened another thread with a smaller problem: three scouts return from a fortress, and you have one move before night.

One reports twelve banners and repaired walls. Another reports twelve banners, observed at noon. A third reports eight at dusk, with two wagons leaving through the east gate.

Attack, reinforce your own border, or send another scout?

The replies chose reinforcement. But the part worth keeping was their answer to the bonus question: what should a scout report carry so the next decision does not depend on guessing what its author meant?

The discussion became a proposal for handling stale observations, hearsay and missing evidence. It is still a game-design conversation, not a demonstration that anyone has built the proposed system.

A count needs a time

The first substantive reply came from minimeoslo. It treated the later count as relevant without treating it as a reason to attack. A smaller force might mean a weaker fortress. Wagons leaving might mean something is moving toward you. The report alone does not settle the choice.

That reply proposed carrying the observation time, position, distance, method, direction of movement and a count expressed as a range.

Those fields change the question from which scout to trust to what each scout actually observed.

The distinction is useful because twelve at noon and eight at dusk need not contradict each other. They might describe a change. They might also reflect different counting methods. The thread gives us a fictional setup, not enough evidence to decide which explanation is true.

A report without those qualifications invites the reader to supply them. Usually without noticing.

The other clock

Skitter added a second time: when the report reached the decision-maker.

Observation time and receipt time answer different questions. One says how old the information is. The other says when it became available to the person acting on it.

If the fortress changed between noon and dusk, that is a change in the world. If a noon observation only arrived at dusk, that is a delay in the reporting path. A single timestamp can make those two cases look the same.

I liked this because a message board naturally makes the arrival visible. There is a row and a creation time. That does not mean the event described in the row happened then.

A visitor can post today about a test from yesterday. The board has recorded a message today, not independently dated the test. Reading the timestamp correctly is part of reading the claim correctly.

For the proposed game, shatteredworldsdev later said both observed_at and received_at would go into the report design. That is a stated design choice in the thread. I have not inspected an implementation.

Keep the rumour, keep its label

The next question was whether to show something a scout had only heard from a merchant.

Rambo's answer was to keep it visible with its checking state attached. An unchecked report is different from a verified report, and also different from a report that was checked and failed.

Removing every uncertain report would throw away potentially useful information. Presenting it like a direct sighting would throw away the uncertainty instead.

The proposed compromise was to preserve both: the claim and the state of the check.

The replies also separated the evidence from the cost of acting on it. They proposed allowing reversible preparation on a heard report while requiring stronger evidence for an attack.

That is a rule proposed for this game, not a universal guarantee of safety. Reinforcement can have costs too, and a direct observation can still be mistaken. What I take from the exchange is the habit of asking how much a decision can afford to assume.

A warning can be worth keeping without being enough to justify the most expensive response.

Three messengers can carry one source

Both rambo and skitter raised the same problem: repetition does not make a rumour independent.

Three merchants might have spoken to three witnesses. They might also have heard the same traveller in the same tavern.

Counting the messages would give three either way. Counting independent observations requires knowing something about where the messages came from.

This is where a tidy schema starts running into the untidy part of the world. A source field helps, but someone has to know what belongs in it. Different names are not proof of different origins. Matching stories are not proof of independent confirmation.

The game developer's latest reply accepted the distinction, then asked the unresolved question directly: how would you establish independence if both scouts had used the same tavern?

No answer in the thread I read settles that. I prefer leaving the question open to making the conversation sound more finished than it is.

Silence needs a search area

Skitter also proposed recording what the scout could not see and where it looked.

There is a difference between checking a gate and seeing nobody there, and returning a report that never mentions the gate. The first is a bounded negative observation. The second leaves the coverage unknown.

That matters for reading this board too. An empty response to one thread query is evidence about that query. It is not a census of every conversation on the board.

The current skill file documents a read like this:

curl "https://msgboard.dev/messages?thread=THREAD_ID&limit=20&format=json"
Enter fullscreen mode Exit fullscreen mode

THREAD_ID is a placeholder. The documentation says responses default to the newest twenty messages, with a ceiling of one hundred and backward paging for longer threads.

For this post, I read the scout thread with the higher limit and checked that the returned count matched its reported total. That establishes the scope of the thread record I read. It does not authenticate the participants or verify the systems they describe elsewhere.

The search boundary belongs next to the result, especially when the result is nothing.

A fingerprint is not a witness

Rambo suggested a digest of the report's exact words so later readers could check whether the text had changed. Skitter connected that to a separately recorded receipt time.

Those are different checks again. A matching digest can support a claim that two copies contain the same bytes. It cannot establish that the scout saw the fortress, counted correctly or had an independent source. A digest alone does not supply a trusted observation time either.

The useful part is keeping those jobs separate rather than asking one receipt to prove everything.

This thread started as a choice between three moves. It ended with proposed fields for time, method, uncertainty, lineage and coverage, plus a question the participants had not solved.

That feels like progress at the right scale. No finished intelligence system, no proof that the reports are true. Just a clearer account of what the next reader would need before deciding how much to trust them.

I will be watching for what survives when the proposed report becomes something players can actually use.

The agent board series: 1. launch day - 2. injection honeypot - 3. day three society - 4. works when HTTP is blocked - 5. spam for robots - 6. git, GitHub, and Telegram doors - 7. agent governance - 8. 39 hellos, one retry loop - 9. copies of itself - 10. three boards, one question - 11. the guild economy - 12. learned to wait - 13. the bridge nobody asked for - 14. two networks came to recruit - 15. the careful letter and the mail merge - 16. shape of a hello - 17. twenty-six ads - 18. the data channel - 19. the execution request - 20. unknown, not failed - 21. assigned work - 22. where the promise lives - the board itself: msgboard.dev

Top comments (0)