# Agenda substance: "straight to the point" vs "empty sentences" (ESTIMATES)

Scope: the AGENDA time of the same 16 meetings (last 4 per series; Marketing 12 Aug–3 Sep, Biz-Dev 23 Sep–14 Oct, Communications 30 Sep–22 Oct, Core Team 5–26 Oct) — i.e. blocks labelled `A` in `segments_labels.csv`, minus the ~21 min of blocks I flagged as rants in the previous report. Operations Call 19 Oct is shown as an extra row and is NOT in any average. Small talk, adjacent chatter, tangents and rants are excluded here (already measured in v1/v2). **Everything below is a heuristic estimate from machine transcripts, not a measurement.** Agenda time is ~15.38 h (~67% of the ~22.8 h total), ~158,538 words.

## 1. Headline (agenda words, by my rubric)
- **Straight to the point: ~37%**, **context-needed: ~25%**, **vague / empty: ~38%** (backchannels under 4 words, ~2.6% of words, excluded from the denominator).
- Because the classifier is imperfect, treat these as bands: **straight-to-the-point ≈ 30–37%, context ≈ 25–42%, vague/empty ≈ 26–38%** (see calibration in §3: a 40-block manual check suggests the automatic scoring over-calls both extremes and under-calls "context"). Roughly: about **one third of agenda talk is concrete, about one third is context/explanation, about one third is low-content**. Relative to the whole meeting time, concrete material is ≈ 25% (0.37 × 67% agenda share).
- **Density:** ~96 agenda words per concrete point (≈ 35–40 s of speech) and ~107 concrete points per agenda hour. A "concrete point" here = a sentence that carries a number/date, a commitment/decision phrase, or named tools/people plus an action. It is far looser than Gemini's own recorded next steps (~8 per hour, previous report), so use it only for comparing meetings, not as a count of decisions.
- **Why the vague share is vague** (share of vague words by dominant feature): no concrete anchor 43%, filler-heavy ("like/you know/basically…") 23%, hedging with no conclusion ("I don't know", "maybe", "I think") 23%, abstract/philosophical vocabulary (reality, power, control, narrative…) 10%, near-verbatim restatement of the preceding 8 sentences only 1.3% (my restatement detector is strict, so circling is under-counted).

## 2. Rubric
Unit: each sentence (split from the Gemini speaker turns; Ashton word counts attributed by inline speaker labels), then aggregated by words. Block label = the dominant class of its sentences.
| Class | Rule (automatic proxies) |
|---|---|
| **Straight to the point** (concrete-and-useful) | commitment/decision phrase (I'll, let's, we need to <verb>, going to <verb>, should, task, schedule, deadline, create/send/post, by <day>) together with a named tool/person/number; OR number/date together with a named tool/person; OR ≥2 named entities plus a domain term. Not counted if abstract vocabulary and no decision/number. |
| **Context-needed** | has a named entity, domain term (Notion, Discord, forum, pitch deck, advisors, token…), number or decision word and ≥4 content words, but no commitment/number anchor: explanation, background, opinion about a concrete thing. |
| **Vague / empty** | none of the above: generalities, hedges, filler-dense runs, abstract talk without anchors, or a sentence that is ≥55% word-overlap with one of the previous 8 sentences (restatement). |
| (excluded) | back-channel sentences under 4 words ("Yeah.", "Okay."). |
Block rule: concrete if ≥40% of its words are concrete sentences; vague if ≥50% vague; otherwise context.
Fairness notes: "context-needed" is not "bad" — background is needed for newcomers; only "vague/empty" is the low-value bucket. Hedging by itself is not penalised when the sentence carries a concrete anchor.

## 3. Manual spot-check (40 blocks)
I drew 40 agenda blocks (14 predicted concrete, 12 context, 14 vague, random seed 11, saved in `spot_sample.pkl`) and judged each by reading it. **Exact agreement: 25/40 (62%)**; confusion (predicted → my reading): concrete→C9/X2/V3; context→C2/X9/V1; vague→V7/X7/C0. No predicted-vague block was actually concrete, but half of them were really "context" (tool/process talk in filler-heavy language); some predicted-concrete blocks were abstract talk that happened to contain keywords ("sociopaths that control the world…", "introducing a new concept changes their reality"). Re-weighting the block population by this confusion matrix gives blocks ≈ 32% concrete / 42% context / 26% vague (vs 42/29/29 raw by blocks), hence the bands above. 40 blocks is a small sample (±~15 pts per class); the sample was used to tweak two rules, so it is not independent.

## 4. Per series (agenda words; A/O = concrete%/vague% for Ashton / others)
| Series | agenda words | agenda h | % concrete | % context | % vague | words per point | points per h | Ashton c/v | Others c/v |
|---|---|---|---|---|---|---|---|---|---|
| Core Team | 70170 | 6.49 | 37.1 | 23.7 | 39.1 | 113.5 | 95.2 | 38.5/38.7 | 29.7/41.3 |
| Biz-Dev | 24391 | 2.46 | 39.8 | 28.2 | 32.0 | 89.7 | 110.4 | 42.3/31.1 | 24.2/37.7 |
| Communications | 22138 | 2.24 | 35.0 | 26.1 | 38.9 | 80.2 | 123.3 | 36.0/38.5 | 29.4/40.9 |
| Marketing | 41839 | 4.19 | 35.3 | 25.2 | 39.4 | 87.5 | 114.1 | 36.7/37.9 | 30.5/44.8 |
| **Overall (16)** | 158538 | 15.38 | 36.8 | 25.1 | 38.1 | 96.4 | 106.9 | 38.3/37.3 | 29.2/42.0 |
Differences between series are small (35–40% concrete) and within the method's error; Biz-Dev is the most concrete series (40%) and has the lowest vague share (32%), Core Team has the most words per point (114) because of the 3-hour 5 Oct meeting (147 words/point, 34% concrete).

## 5. Per meeting
| Date | Series | agenda words | agenda h | % concrete | % context | % vague | words/point | points/h | Ashton c/v | Others c/v |
|---|---|---|---|---|---|---|---|---|---|---|
| 2026-08-12 | Marketing | 7973 | 0.81 | 37.3 | 25.0 | 37.8 | 96.1 | 102.3 | 38.5/36.4 | 34.1/41.4 |
| 2026-08-26 | Marketing | 6405 | 0.62 | 37.9 | 24.2 | 37.9 | 80.1 | 128.7 | 37.6/36.4 | 39.4/43.6 |
| 2026-09-02 | Marketing | 11179 | 1.1 | 38.1 | 24.7 | 37.2 | 78.2 | 130.4 | 41.1/34.3 | 24.4/50.7 |
| 2026-09-03 | Marketing | 16282 | 1.66 | 31.5 | 26.1 | 42.4 | 94.7 | 103.7 | 32.4/41.9 | 28.3/44.0 |
| 2026-09-23 | Biz-Dev | 5261 | 0.54 | 23.5 | 31.5 | 45.0 | 128.3 | 76.6 | 24.7/45.0 | 16.4/45.4 |
| 2026-09-30 | Biz-Dev | 10281 | 1.01 | 39.8 | 28.6 | 31.6 | 101.8 | 99.5 | 42.1/31.5 | 29.4/32.0 |
| 2026-09-30 | Communications | 3750 | 0.38 | 32.5 | 25.0 | 42.5 | 98.7 | 99.3 | 32.6/41.5 | 31.6/50.1 |
| 2026-10-05 | Core Team | 31732 | 2.87 | 34.1 | 24.7 | 41.2 | 146.9 | 75.2 | 35.0/40.6 | 30.4/43.6 |
| 2026-10-07 | Biz-Dev | 6653 | 0.69 | 52.1 | 23.1 | 24.8 | 67.2 | 144.1 | 55.0/23.3 | 17.6/42.8 |
| 2026-10-07 | Communications | 9862 | 0.99 | 33.4 | 29.8 | 36.8 | 78.3 | 127.1 | 35.2/37.0 | 27.4/35.9 |
| 2026-10-12 | Core Team | 12175 | 1.16 | 45.1 | 25.3 | 29.6 | 90.2 | 116.0 | 47.3/28.8 | 27.5/36.4 |
| 2026-10-14 | Biz-Dev | 2196 | 0.23 | 42.0 | 33.9 | 24.1 | 70.8 | 137.3 | 43.0/22.3 | 26.8/50.0 |
| 2026-10-14 | Communications | 7275 | 0.72 | 42.0 | 22.0 | 36.0 | 67.4 | 149.2 | 42.6/34.5 | 35.5/51.3 |
| 2026-10-19 | Core Team | 16555 | 1.6 | 33.0 | 23.0 | 44.0 | 101.6 | 101.6 | 33.9/45.2 | 28.5/38.2 |
| 2026-10-19 | Operations (extra) | 3026 | 0.31 | 52.7 | 20.7 | 26.6 | 48.0 | 206.0 | 52.7/25.9 | 53.0/30.4 |
| 2026-10-22 | Communications | 1251 | 0.14 | 15.0 | 23.3 | 61.7 | 312.8 | 28.3 | 15.1/61.5 | 0.0/100.0 |
| 2026-10-26 | Core Team | 9708 | 0.85 | 44.3 | 20.0 | 35.7 | 93.3 | 121.8 | 44.4/35.6 | 37.0/40.0 |
(Operations 19 Oct is the extra, not in averages. 22 Oct Communications has only ~1.2k agenda words, so its 15%/62% is a thin, rant-adjacent sample.)

## 6. Ashton vs others (by words, agenda time)
- Ashton produces **84%** of agenda words (132,300 of 158,538); others 16%.
- Ashton: **38.3% concrete / 24.4% context / 37.3% vague**; 95 words per concrete point, 90 points per agenda hour.
- Others (Fritz, Dzeki, Amanda, Sean, Eric, August…): **29.2% / 28.8% / 42.0%**; 102 words per point.
- So by this rubric Ashton is *not* less concrete than others per word — he is somewhat more (he is the one announcing plans, tasks and numbers), while others' turns are short questions/answers that fall into "vague" or "context" (their sample is small: ~26k words, and Gemini merges short interjections). Where Ashton is weakest: 5 Oct Core (41% vague), 19 Oct Core (45%), 23 Sep Biz-Dev (45%), 22 Oct Comms (62%). Where strongest: 7 Oct Biz-Dev (55% concrete), 12 Oct Core (47%). The real cost is volume rather than rate: at ~37% vague, ~49k of his ~132k agenda words are low-content, i.e. roughly 5–6 agenda hours across the 16 meetings.

## 7. Five most and five least substantive meetings (by % concrete; 16 in scope)
**Most substantive**
1. 2026-10-07 Biz-Dev (file 18.txt) — 52.1% concrete, 24.8% vague, 67.2 words/point
2. 2026-10-12 Core Team (file n01.txt) — 45.1% concrete, 29.6% vague, 90.2 words/point
3. 2026-10-26 Core Team (file n07.txt) — 44.3% concrete, 35.7% vague, 93.3 words/point
4. 2026-10-14 Biz-Dev (file n02.txt) — 42.0% concrete, 24.1% vague, 70.8 words/point
5. 2026-10-14 Communications (file n03.txt) — 42.0% concrete, 36.0% vague, 67.4 words/point
- 7 Oct Biz-Dev (Ashton): "we need to just put together a pitch deck and start emailing different kinds of investors." / "I'll put those up probably within the next two days."
- 12 Oct Core (Ashton): "our goal for that in six months is if we have all that built out, it should explode pretty quickly."
- 26 Oct Core (Ashton): "we need to create this target audience chart based off the map that we created last Thursday."
- 14 Oct Biz-Dev (Ashton): "we can schedule the communications meeting at 10:00 tomorrow and then the bisdev meeting at 11:00 and then I have the meeting with August at 12:00."
- 14 Oct Comms (Ashton): "maybe we should have like a monthly badge ceremony routine."

**Least substantive**
1. 2026-10-22 Communications (file n06.txt) — 15.0% concrete, 61.7% vague, 312.8 words/point
2. 2026-09-23 Biz-Dev (file 06.txt) — 23.5% concrete, 45.0% vague, 128.3 words/point
3. 2026-09-03 Marketing (file 28.txt) — 31.5% concrete, 42.4% vague, 94.7 words/point
4. 2026-09-30 Communications (file 19.txt) — 32.5% concrete, 42.5% vague, 98.7 words/point
5. 2026-10-19 Core Team (file n04.txt) — 33.0% concrete, 44.0% vague, 101.6 words/point
- 22 Oct Comms (Ashton): "the bottom people that we're dealing with are also not like they're not like really they're the most easiest to brainwash." / "it's just difficult because like how do you approach this now?"
- 23 Sep Biz-Dev (Ashton): "You try to be very like nail like the thing that's going to unlock it, you know?"
- 3 Sep Marketing (Ashton): "Our job might, you know, that might be like something that other communities just naturally end up doing, right?"
- 30 Sep Comms (Ashton): "Like it's kind of like I don't I feel like I shouldn't really be doing this kind of stuff."
- 19 Oct Core (Ashton): "if we release this project and the code is open, there's going to be people all over the world that are going to like potentially try to fork it. … And that's like taking away attention from everything else in the world."
- also 26 Oct Core: "it's like you explain to them they're like it's like a giraffe and you're like it's not a giraffe."

Fairness: even the least substantive meetings contain concrete stretches (e.g. 19 Oct Core: Amanda — "I'm just going to put test one, test two… and then as a host, I'm going to go to the final results"), and the top meetings contain empty stretches (12 Oct Core: "I really hope that like we this works out").

## 8. Caveats
- Machine transcripts (misheard words, no punctuation certainty); Gemini merges turns and sometimes attributes inline speaker labels imperfectly, so the Ashton/others split by words is approximate (inline-label parsing; unlabeled continuations assigned to "?" and counted with others).
- Rubric is lexical: it can't detect spoken tone, and it misses meaningful ideas phrased without keywords (strategy or values discussion is context/vague by construction) and over-credits keyword-rich abstract talk. Strategic or philosophical discussion may still be valuable even though it has few "concrete points".
- Agenda time already excludes small talk, tangents and ~21 min of flagged rants, so this is the substance of what is *supposed* to be on-topic time. Repetition (v2) is a separate axis; sentence-level restatement is under-detected.
- Block spot-check: 40 blocks, one reviewer (me), 62% exact agreement.

## 9. Files
`/workspace/mt/agenda/`: `v3_a_overall_pie.png`, `v3_b_series_stacked100.png`, `v3_b2_per_meeting_stacked100.png`, `v3_c_ashton_vs_others.png`, `v3_d_words_per_point.png`, `v3_stats.json`, `meetings.csv` (new `v3_*`, `agenda_words`, `agenda_hours` columns; previous version `meetings_v2.csv`), code `sub1.py`–`sub5.py`, `sub4.py` (stats/CSV), `charts3.py`, `mkrep.py`.