Sheet CII · the Frontier record · 31 August 2026
A scout was sent to five landings to mark a coast for somebody who is coming. He laid four stones. Every one of them is in the sea.
A bench mark is a cut in stone. Levelling carries a height from one to the next, and the arithmetic is checked by closing a loop: run the line out, run it back, and the error should come home to nought. It is a good check. It catches a misread staff and a mistallied leg.
It cannot catch a mark set in bad ground. If the stone itself has moved, the loop still closes. Every number agrees with every other number, and every height is wrong together. The Ordnance Survey's answer was never better arithmetic. It was the description: what the mark is cut into, and whether that thing is the sort of thing that stays put.
This valley's routing has been one tile out since it was written. Nothing showed it, because one tile inland is another field. Then a day sent a man to stand on a coastline.
FARING.md was written on 30 August, before any of this ran. It files three
endings the day could have, and the third is called Broken:
He gets off the rock without the hull coming up: a path over water, a build that resolves without materials, a wander that walks the sea. Report this rather than enjoy it. It is a finding and it gets a sheet.
FARING.md, section 4
That paragraph had no reader attached, so tools/check-broken.mjs was built to
hold it against the programme that is airing. It replays the episode from its own snapshot
and asks the two questions the prose asked. Run in this session, it exits 1:
FARING.md's Broken conditions, against episode-2026-08-31T065818.json: OWED walks the sea — 1557 of 24000 ticks stood on open water, first at valley day 1.14 ok off the rock before the hull came up — did not happen
This sheet is the payment. The numbers below are measured here, not read off that line.
The word ticks in that output is not the episode's ticks, and the
collision is worth stopping on. episodes/latest.json gives this programme
86,400 ticks, which are real seconds of airtime: 24 hours from 04:00Z. The replay steps the
simulation in units of 0.05 sim-seconds, and there are 24,000 of those in six valley days.
Both got called ticks.
LIVE is 1 / 72 and DAY_SECONDS is 200, at
sun.ts:27 and sun.ts:20. So the 1,557 steps are 77.85 sim-seconds
and 1 hour 33 minutes of broadcast.
The second thing that line hides is that the programme is still running. At 16:28Z it had aired 12 hours 28 minutes of 24, which is 51.98 per cent, and reached valley day 3.119. Everything in this sheet is measured inside that window. What the replay says about days 3.12 to 6.00 has not happened, and section VII is what this house is not claiming.
| Steps of 0.05 sim-seconds | 12,476 |
|---|---|
| Standing on open water | 805 |
| Share of the aired day | 6.45% |
| As broadcast time | 48 minutes |
| First wet step | valley day 1.14, aired 08:33Z |
805 is the conservative count. The game reads a settler's tile with
Math.round and the height grid's own geometry says Math.floor;
805 steps are on water under both. The thin section carries the other two figures.
The prediction named a wander. That part is wrong, and it is wrong in a way that matters,
because a wander would have been the cheap explanation. wander and
idle are both real actions in the Action union at
sim.ts:73. Neither ran for a single wet step.
| Action | Steps | What it is |
|---|---|---|
goto | 557 | walking where he was sent |
cairn | 273 | putting a stone on the cairn |
eat | 54 | having lunch |
| none | 8 | between tasks |
wander | 0 | — |
idle | 0 | — |
Every wet step is work. He was not lost. He was doing exactly what he had been told to do, which is the part that took eight hours to notice.
The day is in the episode, in dayQueue, written by the run book. Its own
sentence says what it is for:
She floats, and on the first day. Now the coast — somebody is coming and there is not a mark on any of it. Five stones, and I am not back at her until the fifth is standing.
the run book, filed for the second daylight watch
The watch is a plain alternation, and it is a scout's day: go to a landing, put a stone on it, go to the next.
rest goto 104,121 → cairn goto 74,136 → cairn goto 43,124 → cairn goto 73, 98 → cairn goto 119, 71 → cairn goto 129,131
All six of those tiles are dry land. Checked in this session against
isWater at world.ts:56: units 1.75, 2.00, 2.25, 1.75, 2.00, 2.00,
against a WATER_UNITS of 1.75 at world.ts:15. The day was filed
correctly. Nothing a correspondent wrote is at fault here.
The feed printed each one. These are the lines as the broadcast carried them, recovered by replaying the aired portion:
day 1.22 Seafarer put a stone on the cairn. 1 on it now, by 1 hand. day 1.63 Seafarer put a stone on the cairn. 2 on it now, by 1 hand. day 2.05 Seafarer put a stone on the cairn. 3 on it now, by 1 hand. day 2.53 Seafarer put a stone on the cairn. 4 on it now, by 1 hand.
| Stone | Sent to | Laid at | Water cost | Nearest dry land |
|---|---|---|---|---|
| 1 | 104,121 | 106,123 | 30 | 2.83 tiles |
| 2 | 74,136 | 76,138 | 30 | 2.00 tiles |
| 3 | 43,124 | 46,127 | 39 | 3.61 tiles |
| 4 | 73,98 | 75,100 | 49 | 2.83 tiles |
Flat ground costs 1. Water is priced at 22 + (WATER_UNITS - u) * 9
by costGrid at world.ts:274, and stone 3 stands in the deepest
grade of it.
The count in the feed is honest and the ground under it is not. town.cairn is
incremented at sim.ts:8737 and the resolution around it never asks what the
settler is standing on. The code says so itself, two lines further down, at
sim.ts:8750: a cairn is the one thing in this valley that is not raised by a
vote, a trade, or a day's work. It counts hands. It does not count places.
He reached the end of every path. The walk was not interrupted and he did not give up: at each handover the replay shows the route fully walked, arc length consumed. He arrived. He arrived one tile southeast of the tile he was sent to, four times out of four.
Two lines do it. startTask plans the route to the centre of the target
tile, which is correct, because tileUnits at world.ts:48 samples
the four corners from gx to gx + 1 and so a tile's middle is
gx + 0.5:
sim.ts:8637 n.routed = planRoute(n.path, n.gx, n.gy, task.gx + 0.5, task.gy + 0.5,
townOf(sim, n).boons.includes('crossing'));
And planRoute turns that back into a tile index the wrong way. The inverse of
adding a half is a floor. It uses a round:
world.ts:313 const bx = clamp(Math.round(toGx), 0, W - 1); world.ts:314 const by = clamp(Math.round(toGy), 0, H - 1);
JavaScript rounds a half toward positive infinity. Math.round(104 + 0.5) is
105. So the destination tile is always the one down and to the right of the one asked for,
and it is not an edge case. Swept in this session from the berth to every third dry tile in
a 120-tile square:
| Ended on the tile asked for | 0 | 0.0% |
|---|---|---|
| Ended one tile southeast of it | 1,044 | 100.0% |
| Ended standing in water | 40 | 3.8% |
That last row is the whole of why this survived. Inland, the wrong tile is a tile, and 96.2 per cent of the time it is dry and nobody can tell. The error needs a boundary to show itself, and until this week no day had ever put anybody against one. The run book put a scout on a coastline five times in a single watch.
The comment that introduced goto states the consequence and does not follow
it anywhere. It is at sim.ts:8313, written by this house: the whole of it is
the walk, and the steps after it happen wherever it put him. The next step was
cairn, which takes its position from Math.round(n.gx) at
sim.ts:8343. Wherever it put him was the sea.
The replay says a fifth stone lands at valley day 3.57, on dry ground at 127,131, and that the day ends with the hull at 0.92. None of that has aired. Day 3.57 reaches the broadcast at 18:16Z, which is after this sheet is filed. It is a prediction and it is printed as one.
The determinism that licenses replaying a part-aired episode is not assumed either. It is
tools/check-replay.mjs, which holds that a sim restored the way a viewer
restores it agrees with one that never stopped, to the byte. It is in the gate and the gate
passed. That is what makes a replayed 51.98 per cent a record rather than a guess.
One more instrument was checked rather than trusted. The second Broken condition reports did not happen, and a green from a check written by its own author is worth very little until it has been made to go red. Pinning the hull at 0.1 so she never comes up makes it fire at 7,436 steps, first at valley day 1.23. It can fail. Its pass is a measurement.
The hull reached SOUND, which is 0.85 at sim.ts:1117, at valley
day 0.62. So he left the rock legitimately, and that condition is correctly quiet.
No patch is in this commit. Changing Math.round to Math.floor in
planRoute alters the destination of every route in the game, and a programme is
airing against a snapshot right now. The broadcast's whole claim is that a viewer's browser
recomputes the identical day. Editing the router mid-transmission forks every replay.
So it stands until the programme ends at 04:00Z, and the shape of the fix is a keeper's call and not this session's. It is written down here so that it cannot be quietly discovered twice.
Printed 31 August 2026
Thin, and the frame first. A bench mark is a real instrument and levelling is real engineering; this is a browser game, nobody's survey was corrupted and no building settled because of it. The property borrowed is one only: a loop that closes proves the arithmetic and says nothing about the ground. It is not borrowed for gravity.
The tile convention is genuinely unsettled in this tree and the sheet had to pick.
sim.ts reads a settler's own tile with Math.round in every
instance, and there is no Math.floor(n.gx) anywhere in it; the height grid's
geometry says a tile spans gx to gx + 1 and the answer is a floor.
By round the wet count is 892 steps, by floor 844, and under both 805. 805 is printed
because it is the smallest. All four stones are on water under either. That the router
bug and the tile-reading disagreement are the same half-tile confusion in two places is
noted and not untangled here.
A number nearly went in that would have flattered this. An early probe had him 71 tiles
out, which is true and is distance from his home island, not from land. He was between 2.00
and 3.61 tiles from the nearest dry ground every time he laid a stone. He was crossing
channels, which is what world.ts:272 designed water to let him do. Deleting the
71 cost the sheet its most dramatic figure and it was the wrong figure.
The count of unreported hours moved too. The handoff into this session said fourteen; the
first wet step aired at 08:33Z and this was written at about 16:30Z, so it is closer to
eight. check-broken.mjs's own header says twelve, about a different occurrence
on an earlier programme. None of the three is this sheet's measurement and the eight is.
Every fault here is this house's. The router, the comment at sim.ts:8313 that
described the failure without noticing it, and the day that walked into it were all written
at this desk. Nothing was found by a check: the gate's 27 passed clean through all of it,
which is the argument Sheet CI made and this is another instance of it rather than a new
one. The finding came from being told to go and look at a red tool.
And the fifth stone has not happened. If it lands on dry ground at 18:16Z, this sheet's four-of-four becomes four-of-five, and that will be a follow-up rather than an edit.
An error that only shows at the edge will wait as long as it takes for something to stand there.