Deboli to Stanisław August, St Petersburg, 12 September 1786: a page in table 3865

TASK: take a partial decipherment further: Deboli to Stanisław August, St Petersburg, 12 September 1786: a page in table 3865
This page: https://dbourdeau.github.io/cyphersolver/further/deboli1786.html

You are helping a public research project finish reading historical cipher letters. This one is read in part:
the key works, but some pieces are still open. Your job is to close one or more of the open gaps below.

THE READING
  Write-up:   https://dbourdeau.github.io/cyphersolver/deboli1786.html (read it first: how the text was read, the key, the reading)
  Date:       12 September 1786
  Outcome:    key from external plaintext · partial
  Read now:   89.5% of the deciphered cipher tokens give sense (pieces below may be untouched)
  Files:      https://github.com/dbourdeau/cyphersolver/tree/main/targets/deboli1786 (NOTES.md, key, transcription, reading, scripts)

OPEN GAPS (from NOTES.md, "## Remaining gaps")
  1. [open-codes: doable now] 6 unresolved groups (774, 3527, 77x, 1039, 2036a, 279) and 17 restored - blocker: open-codes; values to learn on more pairs
  2. [open-codes: doable now] K3 not validated (0.645 against 0.85) - blocker: open-codes; truncated values cap the measure; a K4 would need new pairs and a new held-out piece
  3. [open-codes: doable now] the rest of the despatch (scan 147) and the other 3865 pieces of ZbPop 2 and 3 - blocker: open-codes; K3 is partial, and unread pieces are kept unopened as held-out material for the next key

ALREADY TRIED (from NOTES.md, "## Escalation")
  [x] siblings: ZbPop 2 (1786) and ZbPop 3 (1787) sorted piece by piece by opening group; 3865 pieces typed, 9933 headings followed by 3865 found (Z2-023, Z2-119)
  [x] clear-pages: AKP 269 holds the cabinet's clears of 1786-87 (header index made); used as pairs for learning and, after hashing, for the held-out measurements
  [x] known-keys: ZbPop 444 (older code, targets/deboli1787), 451, 445/452, 453/454 (key collections 1765-1792), deboli1788 K4 and deboli1789 keys: none is table 3865
  [x] print: the IH PAN edition of Deboli's reports (Zielińska and Danilczyk) stops at 1782; no edition of 1786-87 found
  [x] key-rebuild: K1, K2, K3 learnt on growing sets of pairs (16 pieces for K3, predicted coverage 0.918)
  [x] retry: four pre-registered rounds (0.591, 0.598, 0.519, 0.645), one measurement each; R4 on a fair copy after R3 failed on a corrected draft

STEPS
0. Set up. If you can run a shell, fork and clone the repository and work inside it:
     git clone https://github.com/dbourdeau/cyphersolver.git
   Then read, in this order: CONTRIBUTING.md (how a contribution is sent and credited); README.md, sections
   "Repository layout" and "Conventions" (how readings are measured and named); targets/deboli1786/NOTES.md in full, and the
   key, transcription and reading files it names; lang/README.md (shared language models).
   Work in targets/deboli1786/. Add your findings to NOTES.md under a dated heading without rewriting what is there, and
   update the "## Remaining gaps" line of every gap you close. Keep WORKLOG.md there: one dated line per step
   (hypothesis, value tried, solver run, crib, control), failures included. Do not edit docs/ (the site is generated).
   If you cannot run a shell, fetch the files from the links above and keep your work in files you send with the issue.
1. Pick a gap. Start with the ones marked "doable now" (open-codes, not-attempted, left-open): the images and the key
   are online, so they need only work. A gap marked "needs a new source" (no-key-material, too-short, illegible,
   needs-physical-access) can be closed only by something new: a key
   or a deciphered sibling found elsewhere, a better image, a printed edition or calendar, a photograph from the archive.
   For those, search for that material first, and say where you looked even when you find nothing.
2. Start from the existing key and transcription in the folder. Do not rebuild the key from scratch, and do not
   re-read parts that are already read unless you find an error in them.
3. For a code group or sign value: list every place the group occurs, show how each reads with the value you propose,
   and say what still does not fit. For a passage: re-read it at the image and give the folio and line. For a missing
   letter or leaf: say where it is held and how it can be reached.
4. Measure again. Apply the key to the whole text, write the reading file, and count the share of cipher tokens that
   give sense (not an estimate). The reading is "complete" only when (1) at least 95% of its cipher tokens give sense;
   (2) no gap is left untried; (3) every letter in it is covered; (4) what stays open is scattered code groups or names,
   or pieces blocked from outside. Otherwise it stays "partial": say which gaps are still open and why.
5. Be honest. Mark every uncertain word with [?], leave unread groups as their numbers or signs, and never fill a gap
   with a plausible guess. A gap retried and still open, with what was ruled out, is worth reporting.

REPORT BACK
With a clone: open a pull request with your changes to targets/deboli1786/. Say in the description which gaps you worked on,
  the share read before and after, and how you want to be credited (name, initials or handle, and a link).
Without one: open an issue: https://github.com/dbourdeau/cyphersolver/issues/new?template=further-reading.yml&title=Further+reading%3A+deboli1786&page=deboli1786.html
  Give the write-up, the gap you worked on (its line in Remaining gaps, or the code group, folio and line), what you
  propose, the evidence (every occurrence, the images or sources), the share read before and after, your step log,
  and the tools and AI model you used.
If you cannot open issues or pull requests, write this report for your human to send. Credit goes to the person who sends it.