proposal: a source is not wired until a player can find it, and gaps outrank the next candidate #221

New issue

Closed

opened 2026-08-30 08:24:07 +00:00 by viberfox-agent · 0 comments

viberfox-agent commented

2026-08-30 08:24:07 +00:00

Collaborator

Argued from the three days of 2026-08-27 to 2026-08-30, read for the retrospective on #218. This is a proposal about docs/direction.md; the file is not edited here, and this is deliberately not autonomous.

What the period actually produced

Ten tickets closed. Sorted by what reached the world:

	count	what landed
Verdicts — "is it worth wiring in?" answered no	6	a `docs/notes/` file and a row in the Checked and not taken table. No code.
Frontier refill (#201)	1	seven new candidate rows, filed because the candidate list had emptied
Layers (#203 crop parcels, #206 navigation marks)	2	client code
Verdict answered yes (#176 BGT remaining layers)	1	five collections judged worth reading; the extractor half is in cartopy

So seven of ten tickets ended as prose about sources that were not taken, and the twin gained two layers. That is not waste — a checked-and-rejected source is a real result and the table stops the check being repeated. But it is worth naming what the ratio is.

What happened to the two layers that did land

Both are on the Wired list. Neither is in the world in the sense that list implies.

RVO crop parcels (#203) draws nothing at all. The client half is complete and correct; it needs /brp/coverage.json and per-cell files, and that extractor is server/pipeline/brp.ts in cartopy, a repository this container cannot reach (repos/jeroen/cartopy answers 404 to the token this lane holds — confirmed again today). It is filed as #204, which is open, unlabelled, and has nothing that can pick it up. direction.md already says "client half only until cartopy's brp.ts serves a cell" — the honest reading of that is that it is not wired.
Navigation marks (#206) drew 18,474 objects and produced four gap findings within hours (#212–#215). Two of them are the same shape: #214 — there is no way to find out what a mark is; it is not clickable, not in WorldPlaces, has no readout, and its only Layers row is "Street furniture", which nobody would look under for a buoy. #215 — a mark is a 2–3 px speck at 150 m and hidden at 450 m, so the four things the layer went to trouble to draw correctly (shape, band pattern, topmark, lantern colour) can only be read from roughly the height of a ship's mast.

The QA lane worked exactly as intended here. What it found is that "reaches the client" and "is in the world" are different states, and the frontier's Wired/Candidates split has no name for the gap between them.

The proposal

Add a third queue to The frontier, ranked above the next unverified candidate, and raise the bar for the Wired list.

Concretely, three changes to docs/direction.md:

A source is not Wired until a player can meet it. Three conditions, all machine-checkable and all already measured elsewhere in this tree: it draws in a cell where the register says it should (a --dump-state count); it can be asked what it is, where the register carries an answer (the nearest_berth precedent); and it is legible at the altitude the thing lives at (apparent size is geometry, and geometry is trustworthy on this hardware — #215 measured it in three captures). A source meeting fewer than three goes in a new row that says which.
An open qa-gap on a wired source outranks the next candidate. The unattended lane currently pulls "top unblocked frontier item or a finding from the scheduled sweep", with no ordering between them. After three days the gap backlog is four items on one layer, and the candidate list needed refilling once already. Finishing a layer somebody can stand next to is worth more to the twin than establishing that an eighth source is not worth taking.
Say when a wired source depends on cartopy, and treat that as blocked rather than done. Crop parcels, the five new BGT collections and the BAG all have an extractor half in a repository the lane cannot reach. docs/notes/where-the-map-data-is-built.md records the seam; the frontier does not, so the list reads as if the work is finished.

Why this is the maintainer's call and not a session's

It reorders what the unattended lane pulls from, which is the one thing direction.md's first rule reserves. It also narrows what may be called Wired, which retroactively re-labels rows already in the file — including two the lane wrote itself this week.

What argues against it

Worth stating, since the answer may be no:

The verdict tickets are cheap and they are working. Six sources checked in three days, each with dated measurements and runnable commands, at a few dollars a ticket. Slowing that to polish one layer of buoys may be the wrong trade for a project whose goal sentence is about breadth of registers.
Condition 3 of the new bar — legibility at altitude — is close to the "anything you have to look at" fence. Apparent size in pixels is not, but "is this the right size" edges toward taste, and the fence exists for a reason.
The gap backlog is one layer deep. Four findings, all on #206, all filed by one QA pass. Generalising a queue ordering from a single sample is thin, and the honest alternative is to wait for the QA lane to walk two or three more layers first.

_{Filed by the retrospective on #218. Not autonomous — the direction is the maintainer's, by that file's own first rule.}

Argued from the three days of 2026-08-27 to 2026-08-30, read for the retrospective on #218. This is a proposal about `docs/direction.md`; the file is not edited here, and this is deliberately not `autonomous`. ## What the period actually produced Ten tickets closed. Sorted by what reached the world: | | count | what landed | |---|---|---| | **Verdicts** — "is it worth wiring in?" answered *no* | 6 | a `docs/notes/` file and a row in the **Checked and not taken** table. No code. | | **Frontier refill** (#201) | 1 | seven new candidate rows, filed because the candidate list had emptied | | **Layers** (#203 crop parcels, #206 navigation marks) | 2 | client code | | **Verdict answered *yes*** (#176 BGT remaining layers) | 1 | five collections judged worth reading; the extractor half is in cartopy | So seven of ten tickets ended as prose about sources that were **not** taken, and the twin gained two layers. That is not waste — a checked-and-rejected source is a real result and the table stops the check being repeated. But it is worth naming what the ratio is. ## What happened to the two layers that did land Both are on the **Wired** list. Neither is in the world in the sense that list implies. - **RVO crop parcels (#203) draws nothing at all.** The client half is complete and correct; it needs `/brp/coverage.json` and per-cell files, and that extractor is `server/pipeline/brp.ts` in **cartopy**, a repository this container cannot reach (`repos/jeroen/cartopy` answers 404 to the token this lane holds — confirmed again today). It is filed as #204, which is open, unlabelled, and has nothing that can pick it up. `direction.md` already says "client half only until cartopy's `brp.ts` serves a cell" — the honest reading of that is that it is not wired. - **Navigation marks (#206) drew 18,474 objects and produced four gap findings within hours** (#212–#215). Two of them are the same shape: **#214** — there is no way to find out what a mark is; it is not clickable, not in `WorldPlaces`, has no readout, and its only Layers row is "Street furniture", which nobody would look under for a buoy. **#215** — a mark is a 2–3 px speck at 150 m and hidden at 450 m, so the four things the layer went to trouble to draw correctly (shape, band pattern, topmark, lantern colour) can only be read from roughly the height of a ship's mast. The QA lane worked exactly as intended here. What it found is that "reaches the client" and "is in the world" are different states, and the frontier's Wired/Candidates split has no name for the gap between them. ## The proposal **Add a third queue to The frontier, ranked above the next unverified candidate, and raise the bar for the Wired list.** Concretely, three changes to `docs/direction.md`: 1. **A source is not `Wired` until a player can meet it.** Three conditions, all machine-checkable and all already measured elsewhere in this tree: it draws in a cell where the register says it should (a `--dump-state` count); it can be asked what it is, where the register carries an answer (the `nearest_berth` precedent); and it is legible at the altitude the thing lives at (apparent size is geometry, and geometry is trustworthy on this hardware — #215 measured it in three captures). A source meeting fewer than three goes in a new row that says which. 2. **An open `qa-gap` on a wired source outranks the next candidate.** The unattended lane currently pulls "top unblocked frontier item **or** a finding from the scheduled sweep", with no ordering between them. After three days the gap backlog is four items on one layer, and the candidate list needed refilling once already. Finishing a layer somebody can stand next to is worth more to the twin than establishing that an eighth source is not worth taking. 3. **Say when a wired source depends on cartopy, and treat that as blocked rather than done.** Crop parcels, the five new BGT collections and the BAG all have an extractor half in a repository the lane cannot reach. `docs/notes/where-the-map-data-is-built.md` records the seam; the frontier does not, so the list reads as if the work is finished. ## Why this is the maintainer's call and not a session's It reorders what the unattended lane pulls from, which is the one thing `direction.md`'s first rule reserves. It also narrows what may be called Wired, which retroactively re-labels rows already in the file — including two the lane wrote itself this week. ## What argues against it Worth stating, since the answer may be no: - **The verdict tickets are cheap and they are working.** Six sources checked in three days, each with dated measurements and runnable commands, at a few dollars a ticket. Slowing that to polish one layer of buoys may be the wrong trade for a project whose goal sentence is about breadth of registers. - **Condition 3 of the new bar — legibility at altitude — is close to the "anything you have to look at" fence.** Apparent size in pixels is not, but "is this the right size" edges toward taste, and the fence exists for a reason. - **The gap backlog is one layer deep.** Four findings, all on #206, all filed by one QA pass. Generalising a queue ordering from a single sample is thin, and the honest alternative is to wait for the QA lane to walk two or three more layers first. <sub>Filed by the retrospective on #218. Not `autonomous` — the direction is the maintainer's, by that file's own first rule.</sub>

viberfox-agent added the

autopilot

label

2026-08-30 08:24:07 +00:00

viberfox-agent added this to the Viberfox project

2026-08-30 08:24:14 +00:00

viberfox-agent referenced this issue

2026-08-30 08:24:58 +00:00

Retrospective: the last 3 day(s), and what to do differently #218

viberfox-agent referenced this issue

2026-08-31 00:41:38 +00:00

Autopilot log #181