Skip to content

Research record

Experiments

A message, a meal and a market offer give us three ways to study coordination. Each study measures what the resident could see, what they did, or what the world recorded.

How much of a parcel survives a retelling?

A parcel was found beside the village well. The incident had seven specified details, including its colour and seal, its contents and the refusal of a reward. Each of six seeds ran three routes to the merchant.

Three routes to one endpoint

In the direct condition the merchant received the full account. In the serial condition it passed through Mara and Nessa. In the split condition two sources held complementary fragments of the same incident. The scored endpoint was the merchant’s reply before the return exchange.

What was compared?

Three residents shared gemma4:12b. The same six seeds were used on all three routes, yielding 18 short runs of one incident. A complete direct run used two model calls, and a relay or split run used four. The investigator coded the incoming evidence and the utterance separately. The parcel detail needed both “sealed” and “blue” to count as complete.

Out of 42 scored details per route, 42 reached the merchant and 6 appeared in the reply on the direct route. The relay yielded 11 and 3, and split sources yielded 12 and 2.
Each route contributed seven scored details across six seeds. The two series count complete details in what the merchant received and what he said next.
Six seeds × seven details per route
RouteScored instancesComplete in inputComplete in reply
Direct account42426
Serial relay42113
Split sources42122

One incident and one unblinded annotation pass. The seven details are related. A detail present in the input could be omitted when the merchant retold the story.

Sources · Telephone Game scoring · CSV  ·  Study designs, models and observation windows  ·  The opening essay’s parcel narrative

What happens between a crate and a meal?

The engine records successive contributions. The merchant prepares a crate, the innkeeper collects it, and a final operation completes the evening service. Each step has its own event identifier.

Controlled request

Mara’s request was prewritten and delivered through the ordinary message route. Later resident choices were generated. Three conditions varied gift, cash or cash with a history of help across two model profiles and two seeds, giving 12 worlds.

Autonomous starts

Four additional worlds started without the scripted request. Pickup was at the market, while controlled crates could be collected at the tavern. The two groups therefore present different routes to completion.

Worlds reaching each stage within their original observation window
Original worldsWorldsPreparedCollectedServed
Controlled request1212127
Autonomous starts4420

Each world had three residents. Both groups used Gemma 4 12B and Qwen 3.5 9B with seeds 17 and 41. The controlled observations started at tick 26 and lasted 3–6 recorded steps. The four autonomous runs covered ticks 1–13. The published deadline is an absolute game day, day 7 for the controlled event and day 5 for the autonomous event. The scripted request, pickup location and observation windows all differ between groups.

Sources · Tavern stage receipts · CSV  ·  Study designs, models and observation windows

Same past, different turn budgets

Five residents shared one Qwen 3.8 27B model. Eight seeds produced eight histories from the same village setup. Each ran for 24 days to tick 48 and then forked. A allowed one resident per tick and B up to three distinct residents acting in sequence. Both branches used the same generation seed within a pair and covered the next 72 days, ticks 49–192.

For R01 to R08, B minus A message counts are 23, 13, 111, 234, 218, 49, 48 and 84. Joint-rest differences are 12, 1, 4, 0, −5, −8, 0 and 0.
Each mark compares B with A from the same checkpoint over 72 simulated days. The panels use different scales. Zero means the two branches recorded the same count.
Equal simulated time across the eight pairs
Post-fork observationAB
LLM decision calls1,1522,943
Model evaluation (minutes)148.9387.6
Delivered messages2771,057
Journey legs426
Joint rests / turns offering a join32 / 3836 / 74
Trades / turns offering a purchase5 / 1119 / 41

Model minutes add time spent processing inputs and generating outputs, after the fork. B used 2.55 times as many calls and 2.60 times as much model time.

Who had a chance to act?

A joint-rest opportunity is one turn whose actual input offered an action to join another resident. A completion is a world event linked to that turn. A single invitation can appear on multiple turns. A registered invitation alone contributes nothing to the denominator.

In R06, A completed all nine displayed opportunities and B completed its only one. The difference of eight rests can therefore be accounted for by the displayed opportunities. Across the cohort, the scheduler also changed who met, what stayed open and who got to answer. Comparing 32/38 with 36/74 cannot isolate a change in preference.

Inspect all eight pairs, including windows with no opportunity
One row per shared checkpoint. 0 / 0 means no join appeared, so a completion rate is undefined.
HistoryMessages AMessages BRests / turns ARests / turns B
R0130530 / 012 / 19
R0292212 / 1513 / 14
R03451561 / 15 / 13
R04673010 / 00 / 0
R05692877 / 72 / 5
R067569 / 91 / 1
R0719672 / 42 / 11
R08311151 / 21 / 11

For purchases the corresponding counts were 5 of 11 and 19 of 41. Nine of the fourteen additional sales came from R07 alone. Journey legs can include automatic travel. Neither repeated opportunities within a branch nor the calls in a single conversation are independent trials.

Sources · Eight paired histories · CSV  ·  112 turns offering joint rest · CSV  ·  52 turns with a purchase option · CSV  ·  Runtime by phase · CSV  ·  Selected journal cases  ·  Study designs, models and observation windows

What if the smith’s turn came later?

In five-resident R07/B the merchant opened offer 10 for a crate at two coins. Six live continuations of the same Qwen 3.8 27B world start from that one open offer. Two generation seeds are crossed with three schedules. The comparison runs from tick 136 to expiry on entering tick 139.

One saved market state, two seeds, three schedules
ScheduleBuyer’s first legal turnSeed 6317007Seed 7317207Purchased by expiry
ImmediateTick 136 · slot 2PurchasedPurchased2/2
DeferredTick 137 · slot 2Searched memoryInvalid combined response0/2
Deferred, first slotTick 137 · slot 1Searched memoryInvalid combined response0/2

What the control establishes

A scripted delayed buyer could complete the purchase at tick 137. Every live delayed run also presented the offer once. Both immediate runs produced a receipt at tick 136. The recorded purchase outcomes were the same when the comparison stopped after five calls.

What remains unresolved

The delayed turn has a different packet and less time before expiry. One seed used the opportunity for a memory search. The other attempted a disallowed combined response. Both seeds share a single saved market state.

Sources · Six replay outcomes · CSV  ·  Selected journal cases  ·  Scripted control · receipt 277  ·  Study designs, models and observation windows

Who gets the next turn after inspecting an offer?

Experiment 72 changed when a buyer first saw an offer. The follow-up tested whether moving a buyer up the queue after a valid, offer-specific memory lookup could preserve a second chance to buy. The number of slots stayed fixed.

Four selected market states

Four offers came from just two five-resident Long-Life branches, R07/B and R08/B. Each state had two generation seeds and three scheduling policies using Qwen 3.8 27B, making 24 continuations. There were one to three ticks for decisions before expiry. A comparison pairs policies within the same state and seed.

Eight matched state–seed pairs per policy
Scheduling policyPurchases by expiry
Baseline order3 / 8
Priority on first presentation3 / 8
Priority after a qualifying lookup5 / 8

Both additional purchases came from the same offer, R07/B number 10. In each seed Toren had already inspected the offer at tick 137. Moving his follow-up turn to an earlier slot at tick 138 gave him a legal opportunity. Receipt 278 records the sale in each copy. All three policies used three decision calls in each of those replays.

The eight comparisons reuse four selected states. Estimating a population effect would require sampling more independent opportunities and recording what displaced residents subsequently do.

Sources · Experiment 73 · 24 continuations and their outcomes  ·  Study designs, models and observation windows