The loop — where the work is right now

One pass round this loop turns a failing answer into merged code. Every box shows how much work is waiting in it — never a running total, which would only climb. The numbers are the reading order of the picture, not a single chain: after grading the loop splits, so 4 and 5 are the two things that can happen next.

Generated August 15, 2026 at 12:09 PM ET · work units as of 0 min ago · every figure derived from the simulator rack, the work orders on file, the grades, the work ledger and the live process table.
Independently checked 4 min ago: 18 figures recomputed from the underlying files and processes and agreed, 10 moved between the two readings — 3 DISAGREED: page merge count vs the scan: page 150, scan 202 — 52 apart, worktrees appear and vanish between the two; blocked items: page says 140, independent count is 142 · counted straight from the open-alert .
NR means no counter exists for that stage (11 of them, drawn with a dashed edge) — it never means idle. Where a stage is measured and simply quiet, the box shows a real 0. Point at any box to see what is actually in it.

This page rebuilds itself every 15 minutes. Pressing this asks for a rebuild straight away — it usually takes two to four minutes, because the work-unit census reads the ledger, git and every grade file before anything is drawn.

What is blocked, and why

142 things are blocked and waiting on a decision · the oldest has been stuck 19 days · 5 causes

NOT CURRENT — nothing here has been cleared since August 1 at 9:28 AM ET (14 days ago). This is what has piled up, not a live reading: the part of the machine that would re-check these and close the ones that fixed themselves is switched off.

How manyWhat is stuckWhat stopped itSinceThe one thing that releases it
37 Fixes that failed their check and nobody has looked at since Nothing picked them up after they failed July 27 at 8:22 PM ET
19 days		 Send each one back to a coder, or retire it with the reason written down
the supervisor thread
33 Parts of the machine that reported themselves broken The part went red and nothing has re-checked it since July 29 at 9:34 AM ET
17 days		 Look at what went red, then either fix it or record why it is allowed to stay red
the supervisor thread
9 Work that is finished and ready to go in, still sitting Nobody has merged it or dropped it July 29 at 11:35 AM ET
17 days		 Merge it now, or retire it with the reason written down
the supervisor thread
55 Finished fixes that could not be collected from the coder who wrote them The step that picks up a coder’s work hit an error and stopped August 9 at 6:56 PM ET
6 days		 Read the error on the row, repair the hand-off, and run the collection again
the supervisor thread
8 Test jobs a gate refused because they could not be run A gate refused them and said why August 10 at 7:55 PM ET
5 days		 Repair what the gate named and put the job back — or fix whatever wrote a job that cannot run
the supervisor thread

Grouped by cause, oldest first — point at a row to see the individual items. Nothing on this page releases anything: a block is cleared by a person or a thread recording what they did about it, which is why the list can be trusted to be complete.

Colour says who does the work:deterministic codea temporary AI threada permanent AI threadthe orchestrator today — meant to become code two colours on one box = part code, part thread dashed edge = no counter exists (NR), not idle
1 Tests waiting to run 164 set up, waiting for a simulator the full set: NR · bundle tests: NR 2 Test runs NR simulators running a corpus now NR — no simulator heartbeat here 3 Grading NR being graded right now waiting to be graded: NR 4 Waiting to merge 153 finished, not in the trunk oldest finished 4 days ago507 carry an unmerged report in all778 cannot be compared at all108 merged in the last 24 hours · 1289 in the trunk in all25 carry a clean proof, 177 of 212 could not be checked 5 Cluster into fix units 5 the last run could not explain counted 56 min agono corpus run has finished since127 older rows may still be openacross 2 corpora23 already grouped into 11 6 The rack 4 work orders ready to take app 4 · tools 016 drafts awaiting approval11 clustered, no order yet165 in the rack, 570 dispatched 7 Coders 7 coding right now on the app 0 · on the tools 0 · 7 not stated 8 Row check 66 written, awaiting check checking now: 122 more sit only on retired branches11 already pass on trunk, 177 of 212 unknowable 9 Cheat & drift review NR waiting for review reviewing now: NR 10 Batch into one corpus NR batches being assembled fixes in them: NR 11 Tool check NR waiting to be checked 12 Wire into the machine NR checked, waiting to be wired the bundle’s fixes passed new failing rows the bundle’s fixes failed — the batch goes back whole tool jobs into the machine new features & requests queued as one corpus run the app is rebuilt, and the next full set is queued

The gap between clustering and the rack. The work ledger lists 85 fix units with nothing built for them. 11 of those have no work order anywhere — clustered, then never written up, which is the real gap and it is counted on the rack above. For the remaining 74 this page cannot tell you where they stand: the only thing linking a ledger unit to a work order is the unit's name appearing in the order's text, and a sample showed most of those matches are units merely mentioned in some other order's body. That is a missing link between two records, not a stage of the process, so it is written here rather than drawn as a box.

Two ways back. A fix whose rows do not really pass, or that is caught bending the requirement, goes back to the same coder with its evidence rather than to a new thread that has to learn the problem again. A batch that fails its shared corpus run goes back whole, never split apart.

Three things on this page were wrong in the last two days and are worth knowing about: the rack counted every work order on file rather than the ones a coder may actually take; waiting to merge called a fingerprinted proof record “proven”, which reads as row-checked and is not; and cluster into fix units read only the newest clustering file, which dropped four product areas and made a flat number look like progress. All three now say what the record actually is. Stages are named rather than numbered here on purpose, so a renumbering cannot rot the note. The 11 NR figures are one kind of gap — the machine records what has piled up, not what is in motion. What each needs, and what changing it risks, is in docs/reset/For Other Codex/LOOP-SCOREBOARD-INSTRUMENTATION-20260809.md.