Two tiers of one population. The top10 label says which ten come next and
rank/top10.json says in what order; the on-deck label is the shelf a slot is
filled from, and it is deliberately unranked. Every other thing on these cards — the title, whether
it is still open, its milestone, how far its sub-issues have got — is read from GitHub rather than
written down anywhere.
Ranked order exists from 2026-08-20 12:39 — 6 commits across one day, a window of 9 hours 3 minutes. Membership reaches further back, to 2026-08-18 21:32, from the label events on the issues themselves. There is nothing before that: neither record existed.
Every re-rank there has ever been, drawn to scale between the first commit and the last. The rail stops at the last commit because the record does.
The other record: 63 label changes on the top10 and on-deck labels, grouped into 10 acts of labelling. It starts earlier than the order does and carries no ranking — a label cannot hold one.
Rank 4 is a slot open: #267 — "the judge can see the constraints it is judging against" is closed (2026-08-21). Something shipped. Choosing the replacement belongs to 🎩🤖 obot-prime and is not decided here or on this page.
Ten, in order, each with the one line rank/top10.json gives for why it sits there.
Where a row carries a review note, that is the ranking flagging its own uncertainty; it is shown rather than hidden.
gates the session that gates scheduling, and a planning session with him will finalise it, which removes the write-from-memory risk
promise honoured - this row said it went to rank 1 the moment he put a date on the machine move, and on 2026-08-20 he said the migration is this weekend
his word, critical, and obot.agent/main is where the merge policy lives
now one choice away — D0022 is live and the tooling is merged; it needs him to pick A, B or C and nothing else
the file with the most authority over every agent has no owner and no way to change it safely - and it currently names an active plan that does not exist on this machine
two workers have now collided on the same fix in one week - the Navigator's and mine, four hours apart on the deploy incident - and each collision spends fleet capacity twice on work that was already done
he asked for it tonight and it compounds - every new surface starts by copying whichever one was nearest, and the number of surfaces he reads has roughly doubled this week
agent commits still go out under his name, and no token fixes that lane
prime moved this to the bench on 2026-08-20 arguing the new machine fixes it at the root; three agent commits landed under his name on the hub that same evening, so it is back
nobody can write the board, so every requirement's stage is wrong and every surface built on it inherits that
one bullet left and it is his: he has never read a config card and said whether it is better
D0021's premise had already expired by the time he read it; a decision argued from a fact that stopped being true wastes the one thing only he can spend
SafetyCensus rebuilt on metrics and reports, and the design needs his answer before any build starts
The ten carrying `top10` are ranked and are what gets picked up next; the bench carrying `on-deck` is what a slot is filled from, unranked until one is promoted; everything else is open work with no claim on either. These cards carry no rank on purpose: the bench is unranked, and ordering it on screen would invent a decision nobody has made.
Every change to the order is a commit against one file, so the file's history is the whole record of how the ranking has moved. The player below replays it. It is a step per commit — not a time axis; the rail at the top of this page is the one thing drawn to scale.
▲ rose ▼ fell new entered the head back returned after leaving
An order that only ever moved one way would be a ranking nobody was arguing with. These are the items that left the head and came back:
a0c4d51
(2026-08-20 19:41, “The laptop IQ to rank 1, a promise kept, and the stylesheet onto the head”) and came back at rank
6 at 39fb645 (2026-08-20 21:41,
“#260 back on the head — prime was wrong, and the evidence arrived the same night”) — 2 hours later.The same record, without needing the player to read it.
97ed7e32d95ca687653d74e17909a0c4d5139fb645The label carries membership and the file carries order — two mechanisms, so they can disagree, and every disagreement is reported rather than quietly resolved. This is what the labels themselves record, read from GitHub's issue events. Each row is one act of labelling: changes made within five minutes of each other are one decision, not eleven.
top10 · 0 on on-decktop10 · 10 on on-decktop10 · 10 on on-decktop10 · 10 on on-decktop10 · 5 on on-decktop10 · 5 on on-decktop10 · 2 on on-decktop10 · 10 on on-decktop10 · 11 on on-decktop10 · 11 on on-deckrank/top10.json in
jwildfire/obot.agent, the single declared store.
It holds an order and a reason and nothing else, by test.top10 and on-deck labels on
jwildfire/obot.roadmap, one API call each.rank/top10.json, with the store's
bytes at that commit asked for directly rather than replayed from a patch. A commit whose store
will not parse produces a frame that prints its reason; nothing is filled in for it.labeled and unlabeled issue events
for the two labels, 4 pages of them.
This is the only part of this page that reaches back past the store's first commit.