Two thirds of what has been read has never been written up. That gap, not the reading, is what keeps RA out of the guide.
40 of those 921 are sitting only in the gemini lane and have never reached the main one, which is where the guide is built. The audit tool reads the main lane alone, so it reports 881.
Reading is nearly done and nothing has been written up, because no sweep has been defined for it yet. That is a decision, not a delay.
Plus 25,473 other posts in the same manifest, assigned to nobody.
1,152 of those readings carry one line of substance or none. They are real files, but they will yield almost nothing when swept.
The only corpus where reading and writing up are the same act, which is why the two bars match. Six articles left.
99 Muzik issues, 3 Mixmag and 1 DJ Mag, all catalogued and dense. A sweep of all 103 was written and then reverted for quality, so the written-up bar is back at zero.
Opened the day before yesterday and barely started.
research: sources register and ranked future reading list (Joel, 2026-09-22)
23 min ago judged by its last commit
6 uncommitted change(s) in the tree.
research: catalogue Mixmag features
12 min ago judged by its last commit
10 uncommitted change(s) in the tree.
research: DJhistory batch 56, the last four, and the pass closes at 99 of
3 min ago judged by its last commit
spark: RA sweep, nine overnight batches, 153 rows
11 min ago judged by file times, because this lane cannot commit for itself
Its git directory lives inside the parent repo, outside the folder the agent can reach, so it writes work it can never record. 35 change(s) are sitting unlanded right now.
research: sweep Mixmag features 3401-3480, 80 findings
15 hr ago judged by its last commit
1,525 articles have their page text sitting in the corpus and were never turned into a reading. 681 left only a stub saying unreadable and no page text at all, so they would have to be fetched again. 1,078 have nothing on disk whatever. Each needs a different decision and only the first is cheap.
All 103 catalogued issues were converted into 1,480 findings in one run, then reverted on quality grounds. A checking tool was written afterwards. Nobody has decided whether to run it again under that tool.
They are not features and were never part of anyone's assignment. Leaving them in the same file is what made this corpus look like a multi-year job.
Gemini writes findings into a worktree whose git directory it cannot reach. Somebody has to commit for it, every night, or the work is invisible. This will recur until the arrangement changes.
Last code change was 2026-09-22, version v0.2.324. 56 versions shipped in the seven days before that, so this is a stop, not a slow patch.
The grouping run that produces the "no research behind it" list was made 49 min ago over 8,364 findings. There are 10,967 now, so its answer predates roughly 2,603 rows. Rerun it before trusting the gap list.
The last five versions:
GLIDE controls live in the panel: adjustable graph back, TAP BEAT ONE button, works in the plain view
GLIDE turns TEMPO in place: rows flip by character morph, ruler reads the glide
GLIDE opens with no incoming tempo; target set inside GLIDE
GLIDE: tap beat one and ride a tempo change the room shouldn't hear (B)
the hold-to-peek slides smoothly both ways (B)
Every outside source, from research/SOURCES.md:
38 read, 4 partly read, 1 on disk and unread,
3 still to verify. ✓ read from the source · ◐ partly read ·
— held but not read · ? named in a bank, reading not yet confirmed. The corpora count as one row each;
their findings run into the thousands.
Books
Magazines (the printed publications)
Websites and online archives
Papers and documents held as files
research/READING-LIST.md, ranked by what version one of the guide needs.
Confirmed unavailable online, library route only: Bregman, Auditory Scene Analysis, Moore, Psychology of Hearing, Lawrence, Love Saves the Day, Fikentscher, You Better Work!, Veal, Dub, Rietveld.Tier 1: version one depends on these
Tier 2: strengthens version one
Tier 3: depth, and later versions
Requests to Mark at UQ, and where each stands
Magazines and blogs found and not yet read (research/ARCHIVES.md)
Future research projects raised and not picked up
Done on 21 September. The git history was packed, which returned 92 MB and changed nothing else. The hosted page now serves its typefaces as files instead of carrying them inlined, so a version bump no longer re-sends 422 KB; the single-file build still inlines everything, as it has to. The ledger now lists the documents it has never ingested, which previously vanished rather than reading as undecided. Those still need a sweep or a reason each, and they include the one the project instructions tell you to read before redesigning any surface:
The tool counted findings only in the main repo, so a hundred RA articles written up in the gemini lane were invisible. It now counts every worktree, and says plainly how much has not yet reached the lane that builds the guide.
It always said the index was behind. It holds 3,828 rows against 2,268 readable files, so it is 1,560 ahead of what has actually been read and 425 behind the file count. It now says which, and against what.
That sentence is a running total from batch 26. The pass stopped writing it and has since passed batch 54. A progress document appends, so a figure it stopped restating is history. The tool no longer treats one as a live claim.
The document carries a dated correction saying 96 was wrong and the manifest holds 99. It was being flagged for the error it had already fixed. A document that corrects itself is no longer counted as disagreeing.
That ledger lists 59 sources and not one of them is a corpus. It covers the documents and the banks only, 4,250 of the 10,967 findings in the database.
Every folder is a proposed field guide entry. It holds what its own sources agree on, where they disagree, and what rests on one source only, which is the frontier. Folders sort by how many sources feed them. The twelve teaching categories sit at the top; a category marked proposed is material that fitted none of them. Names are machine-made from the phrases a folder repeats, so some are rough.