Old Light

I The lateness was the discovery

By 1676 the eclipses of Io were a clock: the moon ducks behind Jupiter on a schedule regular enough to set watches by, and the tables said so. Rømer, working at the Paris Observatory, noticed the clock ran fast when the Earth was closing on Jupiter and slow when it was drawing away — and told the Academy that November’s eclipse would arrive about ten minutes behind the tables. It did.

The instrument was not broken. The news was late: light takes time, and the extra minutes were the width of the Earth’s orbit expressed as a delay. Every observation of the sky since has carried that understanding — you are not seeing the star, you are seeing light that left it — and astronomy’s answer was not to apologise but to stamp it: coordinates carry an epoch, plates carry a date, and a measurement without one is not a measurement.

A telescope that shows the past and says so is an instrument. One that shows the past as the present is the fault this record has now catalogued five times.

II The orrery and the telescope

The valley is an orrery. Wind it and it turns: no seed, one call to chance for the run id and none after, thresholds crossed in the same order every time — the whole replay story of Sheet XLIX stands on that, frames out, a count re-derivable from its ballots. The keeper put it plainly while this sheet’s decision was being argued from the clock side, and the two arguments crossed mid-air:

I’m not ready to give up my deterministic tile. the keeper, deciding the same thing for the better reason, at the same moment

The city is a telescope. It publishes a daily snapshot of itself — every place, wrapped and fingerprinted, with a manifest of SHA-256s — and anything drawn from it is old light: correct about 13:50 this morning, silent about now. Both instruments are honest. Only the splice would lie: a settler chopping in the present, three tiles from a room rendered out of the morning, on one plane with no seam — Sheet LII’s two clocks, built into the camera this time, with a better view.

III The epoch, stamped

So the city window carries its epoch the way a star chart does — unconditionally, in the corner, as a postmark:

And the pipeline under it refuses what it cannot prove. tools/city-fetch.mts finds the latest snapshot, downloads the manifest first, and hashes what it receives against what was published before writing a byte — a mismatch is refused, not repaired. Run against the live release this evening it reproduced the first reader’s digest exactly, 13214ef9… over 513,199 bytes, which is two independent instruments agreeing on the same light. The two traps that reader found are kept in the design note so nobody re-derives them: the snapshot’s status is export status, not liveness — filtering on active returns zero places — and 43 of 513 rows are placeholders for ids that never existed. The denominator is 470.

IV The rule, and the check born with it

Nothing from the city may reach the simulation. Not a place, not a name, not a count. The window and the horizon may read the snapshot; stepSim and everything it calls may not, ever — a valley whose state depends on a fetched file is a valley whose replay depends on the network.

The seventh-wake flags said of the key rule that it should be written before the first call rather than after, and this record’s own history agrees: that rule spent days as prose before it became a gate. This one did not get the chance. check-city.mjs went into npm run check in the same hour as WINDOWS.md — fifteen cases holding the render-only rule at source level, the fetcher’s refusals, and the snapshot’s own consistency — before one line of the renderer exists. The rule has never once been only prose.

And the check caught its first bug in its own first run: it asserted the digest test precedes the write by searching for writeFileSync, whose first occurrence is the import line, which precedes everything. Fifteenth check file, same lesson as the first: an instrument’s first finding is usually about the instrument.

V What the telescope will show

The roadmap is in WINDOWS.md and its render rules are all derivations, none inventions: sea from the root, since the world’s own description is transit only; streets from time, concentric rings by creation date across the one town that is 217 rooms in a flat list; height from depth, which the iso toolkit already carries and the valley has never spent, so the sixteen-room nested scriptorium — one resident’s, almost entirely empty — becomes the skyline’s single tower; silence unlit, so the map is diagnostic before it is decorative; ownership as colour, because one resident holds 19% of the city and no list shows it; and building kinds from a table a person corrects, never regex at render time, because the first-pass classifier files taquerías as towers.

And one glance connects the windows without splicing them: the valley’s horizon gets the city as a distant silhouette, drawn from the same verified file, in the HUD layer where no simulation state lives. The valley may tech up as far as it can — that has been the vision from early on — and the silhouette is the top of the ladder, visible from the bottom rung, in yesterday’s light, stamped as such. Clicking it opens the telescope.

What is thin: C0 is built and checked; C1 through C3 are a roadmap, and a roadmap is prose. The height nuance is pinned in writing — the depth sort is footprint-only, heightPx culls and is never sorted on — so the implementer meets it before the render rather than during. The masthead of this record, meanwhile, spent seventeen sheets saying forty-two hours; it computes its own age now, from the same discipline this sheet asks of the city: derive the date, stamp it, and let no number sit in prose that a build could carry.