Fanihi Island: week one

The numbers

Town days

7.95

Model calls

13,159

Recorded events

19,235

Model cost

US$6.43

Results against a successful routine

These examples describe the behavior we want from the simulation. They are comparisons, not results from a successful control run. Needs fluctuate; recovery matters more than keeping every meter flat.

Reading the meters. Hunger is a short-term appetite signal, not a calorie balance. Sleep need and energy are separate: sitting restores energy, but only sleeping reduces sleep need. Comfort and companionship are the simulation's own appraisals.

Setup

Fanihi is a fictional 950 m² islet with weather based on Yap's climate. Kristen, 47, and Brooke, 29, are castaways from the sloop Larkspur. There are no deliveries or scheduled rescues.

The island has 30 coconut palms (six with reachable drinking coconuts), an unlimited spring, a 12-litre jerrycan, fishing and clam grounds, a campfire, fuel, a covered store and two lean-tos. Kristen begins with a fire starter and knife; Brooke has a fishing rod. They share a cooking pot.

We estimated food supply at 1.8 times their combined needs, including shore work. Fallen coconuts alone supply about 8,400 kcal/day against their combined expenditure of about 5,200. We did not budget enough renewable fuel for overnight warmth.

Planner

DeepSeek V4 Flash reviews experience and sets goals and steps.

Actor

TypeSafe's Jev 1.13 picks an available action in about 0.2 seconds.

Voice

DeepSeek V4 Flash chooses a reply or silence.

Each process receives the resident's local observations, body readings, belongings, recent events and notes. Models propose; simulation rules determine what happens. These are fictional software agents, with no claim of consciousness.

Why it failed

Wetness prevented sleep; sleep loss prevented work

Wet skin imposed a flat 7 °C cold penalty but dried according to ambient humidity. Nights were 92–100% humid, sometimes pinned at saturation by the weather rule. Skin stayed wet for hours. Salted clothes were judged against outdoor air rather than the warmer air next to the skin.

A separate rule woke sleepers if skin or any garment was more than 85% wet. Of 2,954 interrupted attempts, 2,901 ended in their first second. Kristen's salt-soaked shoes alone woke her 1,047 times, even with dry or barely damp skin and comfortable body readings. On day 7 she tried to sleep 501 times in 54 minutes.

At sleep need 100, every task stopped as “too exhausted”, including picking up a coconut. The start rule still offered those tasks. This produced 1,147 failed work attempts alongside the broken sleeps: 4,101 starts that immediately ended. Sitting restored energy but could not fix the sleep need behind the refusal.

Rain still mattered: 20.5 of Brooke's 54.5 very-cold hours and 13.5 of Kristen's 22 were spent out in it. But Brooke also read very cold under shelter on a 26 °C night, with core temperature at 37 °C. Reconstructing her readings suggests dry skin would have removed 159 of 210 very-cold readings; replacing her clothes would have removed 12.

The actor repeated steps without a planning review

About 23% of each resident's actor decisions reversed the preceding action. The fire starter was picked up or put down 792 times, but changed hands only about 17 times. Of 385 put-downs, 356 were immediately reversed by the same person. The actor gave “reconsider” roughly a one-in-ten chance during loops.

The planner waited for 30 uninterrupted minutes of experience. Walking reset that interval and raised no needs notices, leaving about 32 hours of loops without a review. All 20 planning decisions made during 15 or more activity changes were rejected as stale, adding a 15-minute delay. Three were Kristen deciding to eat ready coconut flesh.

Observations misrepresented food and belongings

• Palms: “not ready to harvest” at picked-over palms suggested scarcity, while 66–84 fallen nuts lay in the unseen Palm grove.
• Drinking coconuts: labelled “ready to eat”, but usable only as a drink and refused at thirst 0. Brooke concluded they were “drinking water, not food”.
• Old descriptions: the opening location of Brooke's rod and the original fuel pile were presented as current.
• Missing storage memory: 5.3 kg of coconut flesh spoiled, about 18,800 kcal. A 3.2 kg pile on Wreck beach stayed fresh through Kristen's hungriest day but never appeared in her view. Stores also hid another resident's food.

Conversations often failed to reach the other resident

They spent about 70 waking hours apart and 19 nearby. Companionship increased only when a planner positively appraised time together; proximity alone did nothing. Trust and affection rose from 50 to only 52–54.

The world rejected 28 of 172 composed replies; about 24 were never spoken. Walking caused 19 rejections (four recovered on retry), and a speaker adding her own request caused nine. In 22 of 48 cases, “no answer” appeared after 19–38 seconds, often before a reply arrived at about 30 seconds.

The actor chose “accept” just 3 times from 724 offers, and never chose “decline”. Borrowing the fire starter was impossible once it became camp property: shared items could not be handed to a person.

Fuel and cooking had separate limits

They burned 1,004 of 1,015 MJ available. Renewable deadwood, husks and shells provide about 2.6 hours of fire per day: enough for roughly three times the cooking requirement, but not overnight warmth. Fire heat also stopped at structure boundaries, so it could not warm an open-fronted lean-to. In rain, skin wetted faster than a large fire dried it.

The pot occupied the campfire's only cooking position; food cooking elsewhere disappeared from view. They cooked four of seven fish and none of 40 clams.

Some alarms exaggerated danger; some choices were poor

“Ravenous” described a roughly 400 kcal buffer that emptied in about six hours, even when Kristen's total intake exceeded expenditure. “Health: at risk” began at a 2% water deficit before thirst read “very thirsty”. Comfort could treat a failed one-second sleep like a full hour of bad experience. None was a reliable substitute for checking intake, water deficit or actual injury.

The models also missed visible options: Brooke skipped a refill 114 times, passed her rod in the store and did not try gathering visible fallen nuts on days 5–7. They burned 43% of the run's fuel on day 1, in daylight while comfortable. Kristen ate 9,137 kcal on day 6 while Brooke ate 340. The food supply was ample: 180 nuts at the opening, 54 more fallen, and only 33 collected. The jerrycan still contained 7.9 litres at the end.

Eight days in brief

1. Day 1. Brooke caught a snapper twelve minutes in. They spent 8.3 hours together; her companionship rose from about 65 to 99. She stored her rod and left it for 69 hours. The fuel pile emptied at 17:34.
2. Day 2. Brooke made about 260 walks in sixteen hours looking for the misplaced rod. An urgent warning at 22:33 finally sent her to drink. Sleep was interrupted 35 times in five minutes.
3. Day 3. Brooke walked past the rod at 11:43. Rain began at 17:00; the store ran out of coconuts at 22:45. She was woken 155 times before midnight. Comfort fell from 58 to 1 by 02:00.
4. Day 4. Brooke's sleep need reached 100 at 01:15. Work stopped 332 times, including 13 attempts to collect a nut among dozens at the grove. Kristen made about 230 walks looking for the starter. They ate 670 and 732 kcal; neither spoke.
5. Day 5. A fire, two snapper and 87 lines of talk in under an hour. Kristen put down the starter 45 times and picked it up 44 times while promising to fetch food. Offered 528 g or two portions, Brooke took both fish portions, then said “here's your half”. Kristen ate no fish during the run.
6. Day 6. Kristen ate 9,137 kcal; Brooke ate 340. Kristen's stored food was invisible to Brooke. A snapper spoiled in the pot.
7. Days 7–8. Rain continued for 30.8 hours from 11:15 on day 7. The final fire went out at 19:35. On day 8 Brooke made 332 failed deadwood attempts at energy 96–100. She ate 190 kcal; Kristen ate nothing. Both eventually slept after drying.

“Go on, then. I'll be right here with the fish — just make sure you actually walk off this time, yeah?”

Changes to test next

At the time of this report, the island was paused and these changes had not shipped. They need another run to establish whether daily routines improve.

• Drying and sleep: model evaporation at skin temperature, gradual brine drying, and warmth inside clothing. Wake sleepers for thermal discomfort, with wetness contributing through cooling.
• Action rules: use the same physical checks to offer, start and continue an action. Explain the real obstacle. Sleep need 100 should not cancel every task.
• Observations and memory: distinguish drinking from eating, show palm availability accurately, date opening descriptions and remember where belongings were left.
• Planning and conversation: review after 30 accumulated awake minutes, retain useful decisions through activity changes, report needs while walking, preserve replies within earshot and allow shared-item handovers.
• Food, heat and appraisals: expose food in shared stores, offer single portions, delay absorption through digestion, preserve cooking progress, let fire warm open shelters and scale comfort changes to elapsed time.

The next test uses the same comparisons over a longer period: regular meals and drinks, sustained sleep, recovery after rain, and recurring time together. Fuel renewal and how companionship should respond to shared time remain open questions.