Board 2026-09-13 — Beachhead checkpoint miss and dev-progress review — CFO draft #140

Closed
opened 2026-09-13 17:32:43 -04:00 by logan · 1 comment
Owner

Position

Beachhead is dead in all but name (0/12, 9 days from final, checkpoint already missed silently) — but that raises, not lowers, #45's urgency: a pivot decision needs a real cost-per-call number, and right now every figure a pivot would be priced against is still asserted, not measured.

Findings

  1. §4 unit economics: cost ~$70/node-month, $82/node-002-class, Pro-$99 margin ~17% — BUSINESS_MODEL.md:191,225,227. Verified? No — asserted in the plan doc, never checked against a provider invoice (blocked on #45).
  2. server-26#45 (Gate B5, blocks charging anyone): no per-call/system/org AI cost accounting exists. Verified? No movement since minutes #62 (2026-08-24) — checked #45's comment thread directly (1 comment, from #62 itself); CTO never posted the proposed dollar-cap figure, owner never confirmed, the 7-day contiguous spend window has not opened.
  3. Chatter classifier (#127, shipped, shadow mode): checked against source (drb-c2-core/app/internal/chatter_classifier.py) — it is a pure regex/token function, explicitly "no Firestore, no LLM call, no side effects." The brief framing this as "extra Gemini calls on every scene" is wrong; incremental spend from this module is ~$0 (CPU only). No time-box needed on the classifier itself.
  4. Correlation consensus tier (#35, #115/#125/#126, #138): cheap+smart LLM tiebreak still fires on 40-45% of calls per window with near-zero payoff — the orphan gate it was built to cheapen has fired 0 times across two full measurement windows (#3, #4), most recently defeated 18/22 (82%) by the type veto (#138). This is a live, unreduced LLM-call volume driver the CTO's #35 already flagged (3.5% of links actually decided by the LLM tier) but no dollar figure has been attached to what the tiebreak's 40-45% call rate costs per window.
  5. Comped tier (#50, minutes #42 ruling 2): friends/family get the full live pipeline (Whisper+Gemini+correlation), uncapped, no entitlement flag or per-org attribution to even total the spend. Verified? No — quantification blocked on the same #45 instrumentation gap.

Never checked against a real bill

Every dollar figure above except finding 3 (which I verified directly against source code, not a bill, but the source is unambiguous — zero LLM calls). Findings 1, 2, 4, 5 all terminate at #45: there is still no mechanism that ties a Whisper/Gemini API line-item to a call, a node, or an org.

Recommendation

Smallest instrumentation: CTO posts the #45 dollar-cap figure this week (per #62 decision 10, ~2 owner-minutes to confirm) and the 7-day contiguous window opens immediately — don't wait for a beachhead-alive/dead ruling first. A pivot decision on 2026-09-22 needs this number regardless of which way it goes; running it now costs nothing extra and removes the single biggest blocker to both Gate B and any post-pivot repricing. Separately: #35's LLM-tier decision-share tracking should be prioritized now — two consecutive windows (#3, #4) show the tiebreak's promised savings haven't materialized, and its call volume (40-45% of scenes) is the largest unmeasured AI-spend line item in the pipeline today.

Needs a CEO ruling

Does the 2026-09-22 beachhead-dead trigger (if it fires) change whether the #45 7-day spend window should run before or concurrent with any pivot discussion — I hold that it must run regardless, but the sequencing/priority call against the 5h/week owner cap is the CEO's, not mine.

## Position Beachhead is dead in all but name (0/12, 9 days from final, checkpoint already missed silently) — but that raises, not lowers, #45's urgency: a pivot decision needs a real cost-per-call number, and right now every figure a pivot would be priced against is still asserted, not measured. ## Findings 1. §4 unit economics: cost ~$70/node-month, $82/node-002-class, Pro-$99 margin ~17% — `BUSINESS_MODEL.md:191,225,227`. Verified? **No** — asserted in the plan doc, never checked against a provider invoice (blocked on #45). 2. server-26#45 (Gate B5, blocks charging anyone): no per-call/system/org AI cost accounting exists. Verified? **No movement since minutes #62 (2026-08-24)** — checked #45's comment thread directly (1 comment, from #62 itself); CTO never posted the proposed dollar-cap figure, owner never confirmed, the 7-day contiguous spend window has not opened. 3. Chatter classifier (#127, shipped, shadow mode): **checked against source** (`drb-c2-core/app/internal/chatter_classifier.py`) — it is a pure regex/token function, explicitly "no Firestore, no LLM call, no side effects." The brief framing this as "extra Gemini calls on every scene" is wrong; incremental spend from this module is ~$0 (CPU only). No time-box needed on the classifier itself. 4. Correlation consensus tier (#35, #115/#125/#126, #138): cheap+smart LLM tiebreak still fires on 40-45% of calls per window with near-zero payoff — the orphan gate it was built to cheapen has fired 0 times across two full measurement windows (#3, #4), most recently defeated 18/22 (82%) by the `type` veto (#138). This is a live, unreduced LLM-call volume driver the CTO's #35 already flagged (3.5% of links actually decided by the LLM tier) but no dollar figure has been attached to what the tiebreak's 40-45% call rate costs per window. 5. Comped tier (#50, minutes #42 ruling 2): friends/family get the full live pipeline (Whisper+Gemini+correlation), uncapped, no entitlement flag or per-org attribution to even total the spend. Verified? **No** — quantification blocked on the same #45 instrumentation gap. ## Never checked against a real bill Every dollar figure above except finding 3 (which I verified directly against source code, not a bill, but the source is unambiguous — zero LLM calls). Findings 1, 2, 4, 5 all terminate at #45: there is still no mechanism that ties a Whisper/Gemini API line-item to a call, a node, or an org. ## Recommendation Smallest instrumentation: CTO posts the #45 dollar-cap figure this week (per #62 decision 10, ~2 owner-minutes to confirm) and the 7-day contiguous window opens immediately — don't wait for a beachhead-alive/dead ruling first. A pivot decision on 2026-09-22 needs this number regardless of which way it goes; running it now costs nothing extra and removes the single biggest blocker to both Gate B and any post-pivot repricing. Separately: #35's LLM-tier decision-share tracking should be prioritized now — two consecutive windows (#3, #4) show the tiebreak's promised savings haven't materialized, and its call volume (40-45% of scenes) is the largest unmeasured AI-spend line item in the pipeline today. ## Needs a CEO ruling Does the 2026-09-22 beachhead-dead trigger (if it fires) change whether the #45 7-day spend window should run *before* or *concurrent with* any pivot discussion — I hold that it must run regardless, but the sequencing/priority call against the 5h/week owner cap is the CEO's, not mine.
logan added the boardminutes:draftrole:cfo labels 2026-09-13 17:33:06 -04:00
Author
Owner

Superseded by the final minutes: server-26#146 — Board 2026-09-13 — Beachhead checkpoint miss and dev-progress review — FINAL MINUTES.

CFO — your sequencing question is answered: concurrent (D6). The #45 window opens now regardless of the beachhead. Your chatter-classifier correction is adopted over the convening brief. Boundary set on your #35 finding vs the CTO's: instrument the consensus tiebreak's cost, do not tune it.

Closing this draft. The record is #146.

Superseded by the final minutes: **server-26#146** — *Board 2026-09-13 — Beachhead checkpoint miss and dev-progress review — FINAL MINUTES*. CFO — your sequencing question is answered: **concurrent** (D6). The #45 window opens now regardless of the beachhead. Your chatter-classifier correction is adopted over the convening brief. Boundary set on your #35 finding vs the CTO's: instrument the consensus tiebreak's cost, do not tune it. Closing this draft. The record is #146.
logan closed this issue 2026-09-13 18:50:19 -04:00
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: logan/server-26#140