Checkpoint _8 -- "DM2S/NX Master Post 03 (ISE v10.3)" Chapter 3 close, for Chapter 4 -- 2026-10-20

Supersedes ISE_v10_Checkpoint_merged_7.md (_7, _6 and _5 are not overwritten). Everything in _7 and _5 still stands unless stated here: _5 holds the standing rules (section 1), the MediaProcessor and cleanISE_Clone notes (sections 4 and 5) and the carried ISE facts (section 7); _7 holds the IP change, the Python 3.15 rebuild on Clone, the Clone disk and rollback notes, and the ISE open items. Read _5, _7 and this file together.

1. Nothing changed on the ISE side since _7

No script, zip or setting was delivered after _7. All ISE open items in _7 section 6 are still open. The state of Clone, Dev and Live is as _7 records it.

2. Standing rules, as restated in Chapter 3 (add to _5 section 1)

• Reply header, casual mode: Claude: [Month DD, YYYY, HH:MM:SS AM/PM], time from an executed command every reply.
• Never reuse an old name for an updated file: append a number. Never rename his files.
• Documents are PDF or Markdown, never Word. He asked for the last explainer in Markdown.
• Plain words. No unexplained acronyms, no jargon coined by Claude. He called the audience "simple humans with limited exposure to AI" and said not to forget it.
• Fewer questions. Statements over option menus.

3. The LLM explainer (side project of Chapter 3, overnight 19-20 Oct)

A two-page, plain-language explainer, dedicated to his friend Terry, for a public master post.

• Title: "The Transformer's iterative transformations". Subtitle: "WTF is really going down with a LLM, in lay speak" (the PDFs through _6 carry the older subtitle "What's really going on inside an LLM, in layperson's-speak").
• Content: the idea in one breath (two loops, depth and time), what happens inside each transformation, a 16-term glossary in plain words, three cheap tests, what the tests cannot tell you, and "try it yourself".
• The three tests, all his: the riddle with a hidden answer ("cook who cannot rely on smell", answer nước mắm; the model said MSG, then fish sauce), the sound-only word "Elelemmm" (el-el-em, LLM), and the trap framing "attention laziness, assignment or default". He is credited as the riddle poser and as the origin of the document.
• Files (each build is a new name):

File	Notes
LLM_in_layperson_speak.pdf	First delivered, 2 pages, old title
LLM_in_layperson_speak_2.pdf	New title, 8-term glossary
LLM_in_layperson_speak_3.pdf	16-term glossary
LLM_in_layperson_speak_4.pdf	His name in test 2 and the byline
LLM_in_layperson_speak_5.pdf	He credited as riddle poser
LLM_in_layperson_speak_6.pdf	Riddle credited as the origin. Last PDF; the one posted at his Acrobat link and forum topic 8043
LLM_in_layperson_speak_7.md	Markdown version with the dedication to Terry and the new subtitle. Final
LLM_in_plain_words.pdf	First build, never delivered

• Build scripts are in the session scratchpad (build_explainer.py to build_explainer_6.py); the .md was written by hand. They are not needed to carry on.
• He sent the explainer to a handful of friends. He decided not to change the PDF for the title mismatch or the dedication; the .md has both.
• Transcript: LLM_Explainer_Chat_Transcript_20261020.md covers this chat from 06:00 on 20 Oct. Everything before it could not be read back (the chat was then flagged by a safety classifier and reads came back redacted). A copy taken from his claude.ai history would be complete.

4. Ideas from the overnight chat (nothing built)

• Influence project (his pseudo memory-like second-model memory broker; a perennial "BASE RULES" processor). Aim: a correction should never repeat. Critique given: "never" cannot be guaranteed by adding text, so store the rule and not only the example; test on rephrased inputs; keep the base-rules block ordered by importance, as one block with a name and delimiter markers; watch cost per call, precedence between rules, and exceptions. Attention favours the start and end of a long text and works to a fixed budget.
• Tokens and sound. A tokenizer sees spelling, not sound ("Elelemmm" read as a typo until tagged phonetic). Quote marks and tags do not change how text is cut up; they tell the model a strange word is deliberate.
• Static versus dynamic. Learned weights are frozen; attention is worked out fresh from the text in view.
• His own term for a block of rules in one entry was "one logical token". Not a standard term.

5. Open items for Chapter 4

All of _7 section 6 (items 1 to 10), in this order of urgency:

1. Fresh full copy of Clone's VM folder, after a clean shutdown, onto the 3 TB drive. He said he was backing it up; not confirmed done.
2. Find what else Clone lost: grep -m3 -i version in MediaProcessor.py (expect 9.30) and MediaIndexer.py (expect 9.9.9); check for cleanISE_Clone v1.0.4; comm listings against Live. He has the zips and decides when to reinstall.
3. Hold off cleanISE_Clone --apply and mtime-based mirrors until the clock-skew files are past (the guests' clock was ahead until about 05:55 on 20 Oct).
4. Patch the old IPs in the four Clone /bin files (json_md5sums_all, copies_Dev_to_Clone, copies_Clone_to_Dev, scp.text). Claude needs the files from him. Find the source of the stale http:// links (grep -n boardurl /var/www/html/Settings.php on Dev and Clone).
5. Confirm Dev's hostname output for the new IS_DEV logic in the indexers.
6. Check Clone's cron before the indexers run overnight; clear up what "do not index all that" meant.
7. Check swap on Clone; tidy /etc/fstab and /boot on a VM copy.
8. Fix Clone's py alias to match reality.
9. Network: the bad route from his mobile-carrier connection to Live (about 460 ms, 10% loss) is not fixable from here. Large transfers to Live should come from Clone's side or another connection.
10. Parked from _5: AudioProcessor not seen; total_items / remaining count includes audio; SMF 2.1.x for LVI; RedBox; signed links for restricted-board attachments.