# thread-query.py — pre-registered trial

**Built and wired 2026-08-24 on steward authorization.** Pre-registered *before* it runs, so the
verdict is not written afterwards to match whatever happened.

## What it is

`~/dotfiles/scripts/thread-query.py`, invoked from `/wake-up` §2.b.3 with `--log`. It queries the
memory corpus **and** the vault using the pulling thread, **excluding the recency slice the wake
already loads** and **favouring older material**. Returns pointers and literal matching lines,
never summaries.

## Why

Measured 2026-08-24: `/wake-up` reads **6,107 words of an 859,803-word memory corpus — 0.71%**,
always the newest. Everything else is dark unless someone goes looking. Twice that day someone did,
by hand, and both greps produced the session's best findings (the Hearth definition in a December
Compass document; the humic layer from April). This turns that hand-motion into an instrument.

The step was previously **described** in the skill ("grep the memory dir…") and fired almost never
— the same described-not-invoked gap that had the link canary hand-typed twice.

Motion comes from the current work, not the calendar. Matuschak's report is that scheduled
resurfacing fails because you are thinking about something else when it arrives.

## ⚠ Known weakness, stated before the trial rather than after

Live run 2026-08-24: results are **plausible but noisy**. Generic terms in a thread — *vault*,
*instruments*, *conventions* — are this corpus's own background vocabulary and pull weak matches.
Threads phrased with distinctive terms do markedly better. **No further blind tuning:** more
adjustment without evidence is guessing, and the trial is the evidence.

## What was already caught, and how

**PASS-BUT-FALSELY on first live run**, at 14/14 selftests green. Scoring on distinct terms present
anywhere in a document made Carruthers' *Book of Memory* (1.6 MB) and Yates' *Art of Memory* (1 MB)
the top hits for every query — a million-word book contains "instruments" and "heap" by accident.
**Presence over an unbounded document measures length, not relevance.** The fixture held only small
uniform files, so the selftest certified its author's blind spot — the identical shape
`vault-links.py` hit on 2026-08-23.

Fixed by scoring on **distinct terms co-occurring inside a ~40-line window**. A length-bias control
is now in the selftest (16 checks).

Also caught: the first archive fix reintroduced the 2026-08-23 substring bug, and was replaced by
**adopting `vault-links.py`'s rule rather than writing a second one**.

## Pre-registered questions — grade 2026-10-05

File the result whichever way it falls. **Grade from `thread-query-log.jsonl`, not from memory** —
a trial graded by recollection has the same shape as the obligation that failed the daily note.

1. **How many wakes surfaced something the steward would have wanted and would not otherwise have
   found?** Count them. Zero is a real answer.
2. **Was any of it dormant** — older than three months and not already linked from the thread?
   Recent-and-obvious does not count; the deficit is re-encounter, not search.
3. **What fraction of runs returned nothing?** A tool that returns something every time is probably
   returning noise.
4. **Did the two-pointer cap ever cut something better than what was shown?**

**Falsifier.** If (1) is zero or near-zero, the step comes out of `/wake-up`. If (3) is near zero
while (1) is low, the scoring is returning noise and the tool comes out. A trial that can only
conclude "keep it" is not a trial.

## Related

- `97. Tools and Systems/Smart Connections — six-week trial.md` — the *while-you-write* half, same
  grading date. This is the *at-the-wake* half.
- `08. Notes/The compost heap — putrefaction as a phase.md` — the figure. This is the turning.