dotfiles/claude/memory/session-2026-08-24-the-heap-got-a-dynamo.md

name, description, type, metadata

name	description	type	metadata
Session 2026-08-24 morning — the heap got a dynamo, and the vault got the steward's own voice	Re-read Mod in full and found three contradictions of yesterday; the steward proposed the compost heap, ruled putrefaction, and glossed the dynamo as the turning; an honest measured audit found two corpora — one fed and machine-authored (860k words), one starved and the steward's — with the wake reaching 0.71%; found `09. Atlas of Roots` specified in Sept 2025 and completely empty, and harvested all 11 `user-` memory files into it as 12 notes in the steward's verbatim; built `thread-query.py` (the turning, wired into /wake-up) and `daybook-cue.py` (the daily note's missing trigger), and installed Smart Connections for a six-week trial. PULLING THREAD: the turning now exists but has never run in anger, and the open question beneath it is whether a container shaped like the WORK — Darwin's move — is what it should be turning.	project	node_type type modified memory project 2026-08-24

Session 2026-08-24 morning — the heap got a dynamo

A morning about re-encounter, which by the end had produced two mechanisms, twelve notes in the steward's own voice, and a considerably less flattering picture of what we have been building.

PAST — what moved, and why

The Mod re-read: three contradictions, which was the point

Read craigmod.com/essays/robot_blood/ in full at source (4,171 words) rather than through a summariser — WebFetch answers via a small model, which is a derived form.

1. The passage we mined argues against what we built with it. Mod names the value of manual transcription — "it would get me to look back over the notes" — then designs an app that removes exactly that step. ⚠ The steward corrected my extension of this to his own practice and was right: photographing underlined pages is a selective re-read. The human pass is not lost, it relocates into the capture act — which makes the selection date real evidence of what is live now.
2. A heavy vault user with our stack does not report our deficit. Mod's stated want is accessibility on demand, not re-encounter. Marshall's deficit is not universal; it is a claim about this vault.
3. "Cheap, continuous exhumation" undercuts appraisal — but only if archived material stays reachable.

What we had entirely missed: the knowing-me slice is the essay's actual through-line ("knowing me, what are ten events I should attend?" → "three or four things I would have likely never chosen on my own"). And the wall he wants is ambient, not navigational.

Checked mechanically, not by impression: over the 3,971-word body, graph 0 · backlink 0 · serendip 0 · resurface 0 · search 0. All 6 link hits are hyperlinks. The whole vocabulary of yesterday's conclusion is absent from an essay by a serious vault user.

Who else works this way

• Matuschak — the decisive negative result: algorithmic resurfacing failed him because when a note came back he was thinking about something else. What worked was linking while writing, since the current work supplies the context.
• Darwin — portfolios keyed to the chapters he was planning, filled over twenty years; "All useful Pages cut out Dec. 7th. /1856/". The container was shaped like the work.
• Warburg — 63 panels re-pinned "until hitherto invisible relationships revealed themselves".
• Luhmann (a partner that must surprise), Bush (selection by association), Aldrich.

The compost heap — and two of the three figures were already the steward's

Steward proposed compost heap beside crucible and dynamo; ruled putrefaction, not petrification; then glossed the third mid-turn: the turning of the compost is the dynamo.

Checked before adding: crucible is ratified architecture (the Hearth, in the Compass Library organ doc) and the compost heap was already the humic layer, steward-verbatim 2026-04-21, built for authors rather than fragments. dynamo was zero in the live vault. Composing instead of checking would have created a second name for an existing figure.

The correction that matters: dormancy is a phase, not a fault. Petrification — intact and inert — now names the failure instead. Written as 08. Notes/The compost heap — putrefaction as a phase.md; it also closes Re-encounter's open question against a push mechanism (a heap is entered, not pushed).

The audit the steward asked for, with the contamination named up front

	ARC	chamber	engine	CM-AI	BMF	vault	apparatus
May	150	9	5	15	4	23	3
Jun	128	41	24	27	34	17	49
Jul	0	143	18	3	0	7	125
Aug	8	17	67	3	3	15	198

⚠ My first reading of this was wrong and the steward's push-back corrected it. 48% of August's dotfiles touches are claude/memory — writing, not machinery; vault tooling is 0.7%. The work did not migrate to apparatus. It migrated to the memory layer.

Two knowledge systems. The vault: 1,298 notes, the steward's voice, ~15/month. And ~/dotfiles/claude/memory/: 506 files, 859,803 words, fed every session — but largely authored by me. /wake-up reads 6,107 words — 0.71%, always the newest.

The sharpest evidence, from our own tooling: the harvest register holds 170 PROPOSED against 4 BUILT. It is the largest unturned heap we own. (Governance itself is healthy: 133 of 172 PENDING dispositioned, 123 REVIEWED entries.)

The Atlas of Roots — a structure that had been waiting eleven months

Looking for a destination, found 09. Atlas of Roots: created 2025-09-29, fully specified, completely empty. It declares eight children — places/pine-street, people/moy-glidden, people/abg, three themes — and contained none. Purpose: "living soil for Kai, Lune, and future descendants." Maxims: "Roots codified are roots alive" · "ensure grief remains fertile, not stagnant." The steward's own soil metaphor, eleven months before the compost figure.

All 11 user- files harvested → 12 notes, verbatim first, any reading marked as a reading:

• Atlas Themes (10): humic-layer · hearth-as-crucible · teachers-never-met · life-as-pilgrimage · out-of-place-as-authority · form-true-to-itself · rules-and-the-poetic · the-long-trajectory · a-music-that-came-as-family · the-chamber-was-a-survival-mechanism
• 08. Notes (2): The work answers to itself · English usage — the European register

The two private-register themes carry the boundary twice — in assurance.risk_flags and in a visible block — so it travels with the note instead of living only in my working file. moy-glidden and abg deliberately unwritten; the index says so and points at the substrate.

Verified: every new note has 2–5 outbound links, none isolated. Vault links 2,200 → 2,242, dead 483 → 475. Remotes checked before moving private material: vault mirror → Gitea only; dotfiles → Gitea and github.com/davidglidden/dotfiles, confirmed PRIVATE.

Two mechanisms built, both steward-authorized

~/dotfiles/scripts/thread-query.py — the turning. Queries both corpora with the pulling thread, excluding the recency slice and favouring age. Wired into /wake-up §2.b.3, replacing a described step ("grep the memory dir…") that fired almost never — the same described-not-invoked gap that had the link canary hand-typed twice. Pre-registered at ~/dotfiles/claude/governance/thread-query-trial.md, graded 2026-10-05 from --log, not memory.

⚠ It failed PASS-BUT-FALSELY on its first live run at 14/14 green. Scoring on terms present anywhere put Carruthers (1.6 MB) and Yates (1 MB) top of every query. Presence over an unbounded document measures length. Fixed by scoring on co-occurrence in a ~40-line window; length-bias control added. A first archive fix reintroduced yesterday's substring bug, caught by the carried control, and was replaced by adopting vault-links.py's rule rather than writing a second one.

~/dotfiles/scripts/daybook-cue.py — a PostToolUse hook (Write|Edit), 16/16 selftests, fail-open. Reminder, never block; quiet for 15 min after firing.

Smart Connections v4.7.2 installed + enabled, trial pre-registered (97. Tools and Systems/…), graded 2026-10-05. Endpoint audit: zero OpenAI/Anthropic/Google/ Cohere/OpenRouter references; embeddings local. ⚠ The default model bge-micro-v2 is English-only and this vault is deliberately multilingual — must be set to Xenova/multilingual-e5-small before first embed.

PRESENT — how it stands

The mood. Productive and heavily corrected, in the good direction. The steward corrected me three times and every correction improved the work: the book photographs, the vault's messiness, and the MemPalace default.

Confidence to recalibrate — three of my own errors, all the same family:

1. ⚠ PASS-BUT-FALSELY on a tool I had just written and selftested. Second day running, second tool. The fixture held only small uniform files, so the selftest certified its author's blind spot — exactly the vault-links.py lesson from 2026-08-23, and knowing it did not prevent it. Only the live run caught it.
2. ⚠ Classified four files from filenames and descriptions, not text. Told the steward all four "second bite" files belonged outside the Atlas; on reading, two say "since childhood" and "regardless of domain" in his own words.
3. ⚠ Mis-corrected the steward from a current config. He said bge was the MemPalace embed; I read config.json — the model they switched to — and called him half wrong. MemPalace's default is minilm, English-only. He was right, and pointing at the identical trap.

What held. Checking before adding to the typology (found two figures already his). Checking remotes before moving private material. Reading the destination's index before placing anything. Parsing frontmatter rather than grepping it. Re-measuring N-now rather than relaying it. Deriving the archive rule from the existing consumer instead of inventing a second.

The daily note. Failed for the whole session until the steward asked. Diagnosed from the record, not impression: notes for 19–22 Aug do not exist; 23 Aug exists at 33,242 bytes across 8 commits. The practice has produced one note — on the day the convention was authored. Missing mechanism, aimed one step too early: the hook only ever creates, and the next cue is the wrap. Plus competing sinks that do have cues (I wrote to the ledger three times unprompted, because Symmetria was an active skill). Skills and hooks fire; documents describe. daybook-cue.py is the fix.

FUTURE — what pulls

The pulling thread: the turning exists now but has never run in anger — and beneath it sits the question of whether the thing it turns is the right container at all. Darwin's portfolios were keyed to the chapter he was writing; ours is keyed to the session's thread. Those may not be the same move, and if they are not, the container is still missing.

Actionable resumption point (as of wrap — re-judge against what changed):

1. ✅ RESOLVED 12:24 — the trial is running on the right model. The steward deleted the previous 110 MB and re-set the model; the plugin's progress panel confirms "Using Xenova/multilingual-e5-small for embeddings", 336/48233. Six weeks run from 2026-08-24; grade 2026-10-05. ⚠ I had reported the opposite minutes before, from smart_env.json still reading bge-micro-v2 in a file written 14 seconds earlier. A config file is a persistence artifact with a debounce; it answers what was last written, never what is running. The banked rule that would have caught it — verify the running binary's provenance, not source HEAD — is the same principle in another domain and I did not reach for it. What held: I stated both possibilities instead of picking one, so only one direction needed correcting.

2. Watch thread-query.py run at the next wake — it fires automatically now. Record honestly whether it surfaced anything wanted. Null results are the trial's data.

3. Then the real question: does a work-shaped container need building, or is thread-keyed turning enough? See the literal question below.

Other horizons, ranked. Load-bearing: the work-shaped container. Then: the vault's nested 99. Archives folders are not excluded by the vault's archive rule — deliberately left undecided, since changing it belongs in vault-links.py, not in a second private rule. Then: petrification detection (no links ∧ no edits ∧ no citation over a window). Steward's alone: moy-glidden and abg, plus the three places and three themes declared in September. Not yet: a fourth status value — do not add a field before something reads it.

Pause statement. I am about to be away from this, and we resume this afternoon. What I want to find still pulling is the container question — because we spent today building the motion and have not yet established what it should be moving through. The two trials will not report for six weeks; nothing today needs defending in the meantime, and the honest state is that both mechanisms are built, tested, and completely unproven in use.

Two post-wrap additions. (a) 08. Notes/In anger — the idiom.md — written at the steward's request after the phrase appeared in the wrap; a gun fired in anger is fired at a real enemy, so the idiom names the gap between demonstrated and relied upon that a green test suite conceals. (b) multilingual-e5-small chosen over Granite R2 — both cover his languages and Granite lists Catalan among 52 enhanced-support languages, a real advantage; E5 was chosen anyway because Granite is labelled experimental and a six-week trial must not carry two untested variables. Granite is the thing to try if the idea proves out.

Literal question for next-Claude (checkable, and it turns on what the wrap files literally say):

Darwin's containers were keyed to the chapters he was writing — durable units that outlived any single working session. Our turning is keyed to the session's pulling thread. Read the last ~15 session files and classify each pulling thread: does it name something that persisted across three or more sessions, or something that existed only for that session? If most threads are session-local, then thread-keyed turning is not Darwin's move, and the work-shaped container is still missing rather than approximated. The negative answer is the useful one.