Status, open items and what is unproven

Where each script stands, the decisions waiting for the user, what is not proven and the next steps.

TL;DR

Two scripts exist. Moser Baer v7 is finished by the pipeline's standard (BAR MET), with an English and a Hinglish version and a Gemini voice-over for all 14 beats. Cafe Coffee Day v1 is a finished English script on the rules of 2026-10-25 morning; the later K, L, M rules add warnings to it. Neither has had a blind reader pass beyond Moser Baer v4, no viewer has seen either, and nothing is committed since v4. The user has decisions waiting on the title, a named private person, the reader pass and committing.

Waiting on the user: Moser Baer title; whether to name Mahesh Sharma; whether to run readers; Cafe Coffee Day's treatment of the founder's death and its title; commit and push; the video-stage decisions on page 11.
Biggest risk: about 76 uncommitted paths on one branch, and local-only data (transcripts/, reports/, data/).
Next step the pipeline names: run the whole runbook on another topic, with the K, L, M rules, to find out whether "BAR MET in one go" is true.

Status board

Item	Status	Detail
Moser Baer v7 (English)	BAR MET	14 beats, 3,029 words, 20.2 min, hook 107 s, lint 100/100 on 70 rules, damage test 56/56. Read-aloud pass done by the author. Reader pass skipped (not authorised). Uncommitted.
Moser Baer v6	approved	"i like v6". Frozen in `examples/moser-baer-v6/`.
Moser Baer Hinglish v7	parity OK	No native read-aloud pass, no reader pass; the lint does not apply. 3,519 words, 16% longer.
Moser Baer TTS prompts	generated	Hinglish and English, 3.1 and 3.8, from `tts-gemini.js`.
Moser Baer audio (Gemini, Hinglish v7)	generated, unheard by Claude	14 WAVs, 20.28 min, about 174 wpm, hook 108.5 s. Tests T1 to T7 not run.
Moser Baer word timing	B1, B2 only	89.0% of words anchored overall.
Moser Baer video plan	plan only	Nothing rendered; two unfilled placeholders in the plan.
Cafe Coffee Day v1	BAR MET (then)	13 beats plus hook, 1,876 words, 12.5 min, hook about 90 s. Under the K, L, M rules `STATUS-v1.md` says hard gates pass, bar not yet met: 9 warnings (for example K1 tags for V.G. Siddhartha, K4 Chikmagalur named once, K5 an undeclared count). Damage test was skipped in that status run.
Series study, topic study	done	Local only (gitignored).
Third cohort with weak performers	not started	The study's own next step.
Blind readers on v5, v6, v7, CCD v1	not run	Needs the user to ask.

Decisions waiting for the user

Moser Baer

Title. The working title is Moser Baer: How the Way Out Became the Trap. "Killed" outruns the evidence; do not use "scam" or "fraud" in the title. The user has not chosen.
A private person. Mahesh Sharma is named in the hook, beat 3, beat 10 and the close, on the strength of one Business Today article seen only through search summaries. Name him, or say "the union's president".
"Pivot" for the 1983 move. Calcutta to Delhi is a loose use of the word. Keep the broad definition from the hook, or start counting pivots at the floppy plant.
Reader pass. Two blind readers on extract-vo.js --mode blind output, if the user authorises subagents.
Commit and push. Nothing is committed since 156897f.

Cafe Coffee Day

How much to say about the founder's death (currently one plain sentence, "appeared to be suicide, pending investigation").
Whether to name the lenders and the successor CEO.
The 2019 letter is not quoted (authenticity unconfirmed). Keep that, or confirm and use it.
Whether the four S1 transcripts missing from disk should be re-fetched (a re-run without --refresh fetches only what is missing).
The title (working title Cafe Coffee Day: Why the Cafes Could Not Carry the Land, with alternates in the hand-off).

Voice and video

The seven decisions on page 11: pace, music route, source captures, stock footage, logo, Whisper model, gap between beats.

Open items before filming (Moser Baer v7)

From HANDOFF-v7.md and section 7 of SCRIPT-v7.md.

Confirm Mahesh Sharma's details by opening the Business Today page (15 December 2019); it has only been seen through search summaries.
Confirm Decision 2003/371/EC and Regulation 960/2003 on EUR-Lex (the Commission found no dumping for Moser Baer India Ltd; a parallel anti-subsidy case ended in a duty).
The persona's "I" lines must be true of whoever records the video.
Where the bond's money went (not found in any source).
The debt split. Warburg's stake after May 2012 (two magazines differ, so the script does not say).
The floppy start year; the CARE date.
Recheck the CBI challenge to the discharge the week of publishing.
Confirm Ratul's 2003 words ("never been more excited").
Length is 20.2 minutes. Cut candidates: beat 6 (films) and beat 2 (the cocktail party).
"Sales doubled" is softened because two sources do not reconcile.

What is not proven

Say these plainly in any hand-off

No lesson has been tested on viewers. A to M are design constraints from the user's reviews and from research graded A to D. No video made with this pipeline has been published or measured.
The linter proves structure only. A score of 100 means the declared quotes are in the text and the budgets are met.
The thresholds are design budgets, calibrated on at most 8 English winner transcripts. K, L and M were written from one review of one script and are not calibrated on winners.
The winners we measured differ from v6 and v7 in figures of speech and first-person admissions, and the cold open and early figures went the wrong way in both cohorts. The script's persona is a bet.
"BAR MET in one go on a new topic" is not demonstrated. Moser Baer took seven rounds. Cafe Coffee Day v1 met the bar on the earlier rules in one session but has not been rechecked under the newer ones.
The blind study could only compare winners with winners. It cannot say what separates a hit from a miss.
The topic study counts what existing videos drew, with search-page views for most of them. It does not predict a new video.
The Hinglish has not been heard by a native reader, and Gemini's handling of Roman-script Hinglish, line breaks and tags is unknown until tests T1 to T7 are run.
Claude cannot listen. Sound judgement is the user's.
Single-source facts remain in the ledgers as [V-search] or [CHECK] rows. Each is listed in the open items.

Next steps

The first group is what the pipeline's own notes name, in order of value. The second group is my suggestion from reading the repository; neither is a plan the user has approved.

Named in `docs/SCRIPT-PIPELINE.md` and `tools.md`

Run the whole runbook on a second topic. (Cafe Coffee Day was that run, on the pre-round-7 rules.)
Add the E to M checks to the cohort rubric (rubric v4).
Add a named-people-and-terms density rule (partly done in round 7: K3, K4, M1).
Run blind readers on v6 or v7 (needs the user's request).

Suggested

Protect the work. Ask the user whether to commit. The skill, the references, the examples, v5 to v7 and the Cafe Coffee Day folder exist only in the working tree. The gitignored studies (reports/, transcripts/, data/) have no backup in the repository.
Close the Cafe Coffee Day warnings by applying the K, L, M lessons to v1 as v2 (scaffold.js --from v1), which also tests the claim that the lessons carry over.
Run the TTS tests T1 to T7 on one short beat and save the WAVs so their lengths and silences can be measured.
Align the remaining 12 beats and fill the two placeholders in the video plan.
Choose the third topic with the gates on page 04, and rerun the topic study first if it is stale.
Build the third cohort with weak performers so a script rule can finally be tested against failure.
Test packaging on titles at scale before spending transcript cohorts on it, using the 150 enriched videos already in the store.

If you are a new session

Read CLAUDE.md and the start page.
Run check.js on the script you were asked about.
Read its HANDOFF-vN.md and this page.
Do not edit rules.json, do not commit, do not spawn subagents unless the user asks, and write what you did to the research record and to learnings.md.
Before you tell the user anything is done, run the tool and paste its output line.

Written 2026-10-27 from the repository files and the project memory notes. If a page and a file disagree, the file is right: fix the page.