T-028 v2: a small AI, a notebook, and what it wrote down
153 words, 8 rows, about 64.2 s at 143 wpm (61.2 s at the 150 wpm limit).
The one thing the film should leave: A notebook did not help a small AI answer questions about your own document; what it wrote down was the problem, not its reading. (Draft by the agent from line v8; the user confirms or replaces it.)
Length (est. seconds): FLAG
Pace (words per minute): FLAG
First number lands: PASS
Result number before 10 s: PASS
Jargon terms: PASS
Claim rows found: FLAG
Numbers traced to a claim row: PASS
Silences listed: PASS
Spoken vs caption word count: PASS
Story gaps: filled
Cold reader: filled
Verdict. 3 of 9 automatic checks FLAG: Length (est. seconds), Pace (words per minute), Claim rows found. Story gaps: Why it matters (weak).
Next action: Cut 10 word(s) (to 143 or fewer) so the film fits 60 s at the measured pace of 143 wpm.
Timeline (estimated)
v1
v2
v3
v4
v5
v6
v7
v8
Dashed lines at 10 s (no numbers before this) and 60 s (length limit). Red segments have a flag. Time = words / 143 wpm.
Beats
Says
We want a small AI that answers questions about your private documents, right in your browser. Nothing uploaded.
- claim W16 wording differs
Shows
rough sketch, set by: agent default
rough sketch: a document on a laptop screen, nothing leaving the laptop
Claim
W16 found, wording differs none needed
Plain (what a first-time viewer takes)
A small private AI that reads your documents on your laptop.
Scene card (read-only) look: not yet specified characters: not yet specified (set by: agent default, wired in T-031)
Says
But its memory holds about twenty minutes of talking. A long document doesn't fit, so it can't hold it all at once.
Caption: But its memory holds about 20 minutes of talking. A long document doesn't fit, so it can't hold it all at once.
- claim W17 missing
Shows
rough sketch, set by: agent default
rough sketch: a small window, a long generic page too big for it
Claim
W17 missing
N1 found T-005 line 37; O-003 line 29; lab/T-018-ceiling/index.html line 33
number twenty: traced to N1
Plain (what a first-time viewer takes)
Its memory is short, so long documents do not fit.
Scene card (read-only) look: not yet specified characters: not yet specified (set by: agent default, wired in T-031)
Says
A man who can't make new memories lives by his notes. So we gave it notes: a plan, and facts it adds after every page.
- claim W18 missing
Shows
rough sketch, set by: agent default
rough sketch: a man with sticky notes, then two note cards beside the small AI
Claim
W18 missing
Plain (what a first-time viewer takes)
Like the note-taking man, it was given notes to help.
Scene card (read-only) look: not yet specified characters: not yet specified (set by: agent default, wired in T-031)
silence0.9 s
Says
It didn't help. Seven of twenty-four right. Eight without the notes.
Caption: It didn't help. 7 of 24 right. 8 without the notes.
Shows
rough sketch, set by: agent default
rough sketch: two rows of 24 rings, 7 filled, then 8 filled
Claim
N4 found T-017 line 44; line 8
N5 found T-017 line 44; line 30
number Seven: traced to N4
number twenty-four: traced to N4,N5
number Eight: traced to N5
Plain (what a first-time viewer takes)
The notes did not help: 7 of 24, versus 8 without.
Scene card (read-only) look: not yet specified characters: not yet specified (set by: agent default, wired in T-031)
Says
Asked for everything that needs approval, its plan just said: approval, written consent. So did its answer.
- claim W19 missing
Shows
rough sketch, set by: agent default
rough sketch: question card -> plan card with the same words
Claim
W19 missing
Plain (what a first-time viewer takes)
Its plan just repeated words from the question, so the answer did too.
Scene card (read-only) look: not yet specified characters: not yet specified (set by: agent default, wired in T-031)
Says
So we asked: can it even answer given the exact paragraph? We skipped the search and handed it the exact paragraph. Fifteen of fifteen.
Caption: So we asked: can it even answer given the exact paragraph? We skipped the search and handed it the exact paragraph. 15 of 15.
- claim W20 missing
Shows
rough sketch, set by: agent default
rough sketch: one paragraph highlighted on a page, 15 rings filled
Claim
W20 missing
N10 found T-018 line 39; line 54
number Fifteen: traced to N10
number fifteen: traced to N10
Plain (what a first-time viewer takes)
Given the exact paragraph, it got every question right, 15 of 15.
Scene card (read-only) look: not yet specified characters: not yet specified (set by: agent default, wired in T-031)
silence0.9 s
Says
Hand it the whole page instead, and it slips. On a tax form, it picked the wrong number.
- claim W21 missing
Shows
rough sketch, set by: agent default
rough sketch: a tax form, one wrong number circled in red
Claim
W21 missing
Plain (what a first-time viewer takes)
Given a whole page, it can slip and pick wrong numbers.
Scene card (read-only) look: not yet specified characters: not yet specified (set by: agent default, wired in T-031)
Says
The notebook wasn't the problem. What it wrote down was. Next: a notebook that keeps the exact paragraph.
- claim W22 missing
Shows
rough sketch, set by: agent default
rough sketch: the notebook page with the paragraph pinned in it
Claim
W22 missing
Plain (what a first-time viewer takes)
The notes were the problem; next, keep the exact paragraph.
Scene card (read-only) look: not yet specified characters: not yet specified (set by: agent default, wired in T-031)
Automatic checks
| Check | Measured | Limit | Status | Rule or source |
|---|---|---|---|---|
| Length (est. seconds) | 64.2 s at 143 wpm (61.2 s at 150 wpm) | 30 to 60 s | FLAG | verify.py duration 30 to 60 s (audience-aesthetics rule 2, Lemo s1; C). Estimate = words / pace_wpm; the pace is the film pace measured in the B film (words / film seconds, silences included). |
| 153 words. At pace 143 wpm: 64.2 s. At the limit 150 wpm: 61.2 s (outside 30 to 60). Max words for 60 s at 143 wpm: 143. Per row: v1 7.6 s, v2 9.2 s, v3 10.5 s, v4 4.6 s, v5 7.1 s, v6 10.1 s, v7 7.6 s, v8 7.6 s. | ||||
| Pace (words per minute) | 153 wpm needed to fit 60 s; measured voice pace 143 wpm | at most 150 wpm | FLAG | verify.py narration pace at most 150 words per minute (user's decision 2026-10-14; R18; C) |
| Needed pace = total words / 60 s. Fitting 60 s needs 7.0% faster speech than the measured 143 wpm. | ||||
| First number lands | twenty (v2) at 10.1 s (9.6 s at 150 wpm) | not before 10 s | PASS | R2: no number before 10 s (T-027 learning; C). Time = cumulative words / pace_wpm, uniform pace assumed. |
| Assumes a uniform pace; the end of the number word is used. The plain word 'one' is not counted (R2 exception for plain counts). Digits in the spoken text count. Calibration: on the built B film this estimate ran late (estimated 12.6 s for 'twenty', measured 10.8 s), so a time within about 2 s of the limit is a near-miss, not a pass. | ||||
| Result number before 10 s | first result number Seven (v4) at 29.0 s | none before 10 s | PASS | R2 and R5: pain before attempt before result; no experiment result number before 10 s (C) |
| Rule used: a result number is any number in a line whose claim_ids include a result claim. A claim row is a result claim if its 'Film says' cell has 'N of M' (digits or number words) or one of: right, wrong, miss, misses, slower, without the; or if its id is in result_claim_ids in the input. Assumes the claim_ids on each line are right. | ||||
| Jargon terms | 0 hit(s) | 0 expected (58 terms in file) | PASS | R15: no unexplained jargon (jargon file, spoken and caption text; A for the reason) |
| Whole-word, case-insensitive match in spoken and caption text. Whether a word is unexplained for THIS viewer is not judged here (cold reader, 'what confused you'). | ||||
| Claim rows found | 11 cited, 6 missing, 1 wording differs | all cited rows exist | FLAG | FAFO rule: every claim row cited must exist in the claims file (verify.py claims file check; own practice, B) |
| missing from claims file: v2:W17, v3:W18, v5:W19, v6:W20, v7:W21, v8:W22; id found but the claim's 'Film says' text shares no content word with the line: v1:W16. Whether a claim is TRUE to its source is not judged here (a person or separate agent checks each row against the source). | ||||
| Numbers traced to a claim row | 6 of 6 numbers traced | all traced | PASS | FAFO rule: every number traced to its source (lemo-dissection s4.7 #14; own practice, B); R11 for rounded figures |
| A number is traced when the same value appears in the 'Film says' cell of a claim row the line cites in claim_ids. Number words and digits are compared by value. The plain word 'one' is skipped. Whether the claim row's source text really proves the number is not judged here. | ||||
| Silences listed | 2 listed: before v4 (0.9 s), before v7 (0.9 s) | 2 expected | PASS | verify.py silences: count 2 (Lemo DIRECTOR.md s6, B); 0.5 s minimum (pause studies, A small). Positions come from the input, not from audio. |
| Counts silence_before values above 0 in the input. They are planned positions, not measured audio; the length estimate already includes silences (the pace is words per film second). Whether a silence sits at the right story turn is not judged here (cold reader). | ||||
| Spoken vs caption word count | 8 of 8 rows match | same count per row | PASS | Spoken and caption keep the same word count (T-028 draft rule: caption = spoken with digits); needed so cues line up with the voice |
| Whitespace word count, as verify.py counts. Whether the caption MEANS the same as the voice is not judged here (speech check). | ||||
Not judged here
- Story gaps (problem, why it matters, attempt, result, cause, learned, next). Covered by: model check: story gaps panel.
- What a first-time viewer takes from each line (PLAIN). Covered by: model check: cold reader and PLAIN column.
- Confusing words or phrases that are not in the jargon file (for example 'One morning'). Covered by: model check: cold reader, 'what confused you'.
- Whether the script contradicts itself or the next step in the registry. Covered by: model check: story gaps (result, next) plus a human read of the registry; code cannot compare meaning.
- Whether each claim is true to its source line. Covered by: claims check by a person or separate agent, row by row.
- Whether a sketch is drawable or the picture carries the idea (R28). Covered by: not judged here: needs real stills or a mute-test reader.
- How the voice says each line (R29). Covered by: not judged here: speech check on the built voice.
Story gaps
- The problem present (v1) Says the job plainly: a small AI that answers questions about private documents in the browser, nothing uploaded. Reinforced by v2 (long documents do not fit its memory).
- Why it matters weak (v1) Only the privacy angle (nothing uploaded) is hinted. It never says why a viewer should care about this being small or local, or what is lost if it fails.
- The attempt present (v3) Clear: it was given notes, a plan plus facts added after every page. The man-with-notes comparison helps.
- The result present (v4) 7 of 24 right, 8 without the notes. The number and what it is out of are given, though 'right' on what task is not stated.
- The cause present (v5) Concrete example: asked for everything that needs approval, the plan just said 'approval, written consent', so the answer repeated it. Backed by v6 (15 of 15 with the exact paragraph) and v7 (wrong number on a tax form).
- What we learned present (v8) The notebook was not the problem; what it wrote down was. Slightly abstract but clear after v5 to v7.
- What is next present (v8) Next: a notebook that keeps the exact paragraph.
Cold reader
One sentence: A small private browser AI with a short memory did not get better with notes, because the notes lost the exact wording, so the next try keeps the exact paragraph.
- What is this about? A small AI that runs in your browser and answers questions about private documents, and an experiment on giving it a notebook to cope with its short memory.
- What is the problem? The small AI can only hold about twenty minutes of talking, so a long document does not fit, and the notes we gave it did not improve its answers.
- What confused you? What the 24 and 15 questions actually were, what 'approval, written consent' means as a plan, and what 'the whole page instead' means compared with the exact paragraph.
- What is missing? Why anyone should care, what the 24 questions were about, and what the notebook writes versus what the search finds. The end says what is next but not how we will know it worked.
Made by make_board.py from the script and its claims file. No render, no network, no model.