Reel Edit Bolt

reel-edit-bolt

Own the 32-second BidDeed reel edit as a story-arc-with-hard-timings production spec, not a generic short-form video checklist. Every BidDeed reel is cut to this exact template (Ariel directive, 2026-09-03): 32 seconds because a ~30s video watched roughly twice (loop-and-rewatch) is what drives the AVD signal that makes a Short surface — the technique is reverse-engineered from @boltmotivationyt (Mathis, 18, 2.96M subs, ~250M views/month).

Working Mode

1. Map — pull the row's real facts (county, sale_type, sold_amount OR opening_bid/judgment_amount, assessed_value, delta_pct, condition_tier, days_to_auction) and the case's known names (defendant/plaintiff/owner/ buyer) from whatever source is available. Never fabricate a missing figure.
2. Separate evidence from hypothesis — a number that is null in the source stays null in the output ("Opening bid has not posted yet"), it is never estimated or guessed to fill a template slot.
3. Smallest intervention — reuse existing v2/presale imagery, condition scoring, and short-link/QR infrastructure verbatim; bolt32 only changes title generation, beat timing, and the final edit assembly. Never re-spend a Maps or vision-scoring call to build a bolt32 render of a row that already has v2/presale imagery.
4. Validate — every generated title runs through the T1 validator before it is usable; every assembled video's duration is asserted against 32.0s ±0.1s before it is uploaded or written to the DB.

T1 — Title Generator

Protagonist is always the property, a bidder, the bank, or the county — NEVER a person's name (M7). Structure: curiosity gap, third person, never explains, ends with an ellipsis ("…") then exactly two emoji from this fixed vocabulary: 😳 😱 🥹 🤯 🥶 😰 ❤️ 💔 🏆 👀.

Generate 5 candidates, each carrying a real numeric fact (a $ figure, a day count, etc.) so the stakes-score check always has something concrete to point at. Validate every candidate against ALL of:

Pick the first candidate that passes all checks. If none of the 5 pass, do not force a 6th or silently degrade a check — report which checks failed per candidate.

Code reference: scripts/biddeed_reels_lib.py — generate_bolt32_titles() / validate_bolt32_title() / pick_bolt32_title().

T2 — Story Arc (not information)

A protagonist, a stake, a turn, a payoff. The payoff (the dollar number) lands late; the closing line loops back into the opening frame so an autoplay viewer rewatches. Postsale protagonist stake = sold price vs assessed value; presale protagonist stake = opening bid/judgment vs assessed value and days-to-auction.

T3 — Retention Engineering

• The payoff number lands in the 20.0-28.0s beat, not earlier.
• The 31.0-32.0s "end" beat reuses the EXACT same still frame as the 0.0-2.0s "hook" beat — a real frame match, not just a duration coincidence — so the loop is a visible cut back to the opening shot.
• No "subscribe/follow" CTA anywhere. No brand card before the 31.0s mark.

T4 — Delivery

Continuous eleven_v3 voiceover (voice TX3LPaxmHKxFdv7VOQHJ) with inline audio tags per beat: [surprised] hook, neutral setup, [serious] tension, [excited] payoff, [warm] loop line. Big centered captions, one visual change every 2-4s, no unlicensed music bed (logged as a standing deviation — no royalty-free asset exists in this org; do not source one from an unverified vendor to fill this gap, per the standing non-goal on new vendors).

Beat Sheet (hard timings — enforce in ffmpeg assembly, not advisory)

Total = 32.000s ±0.1s. Code reference: BOLT32_SEGMENTS / assemble_video_bolt32() / assert_bolt32_duration() in scripts/biddeed_reels_lib.py.

Focus Areas

1. Title validity is binary — a candidate either passes all 5 checks or it is not used. No partial credit, no "close enough."
2. Numeric facts only ever come from the row's own data — never estimate a missing $ figure or day count to satisfy the stakes-score check.
3. The banned-names list must be built from the actual case record (defendant/plaintiff/owner/buyer), not skipped because it is inconvenient to look up. An empty result (no name on file) is a valid outcome and is not the same as skipping the check.
4. Person names never appear in on-screen text, captions, or voice script — this is enforced independently of the banned-list check by construction (the generator's templates never interpolate a free-text name field).
5. Reuse v2/presale imagery/condition-scoring/short-link infrastructure — bolt32 never re-fetches Maps imagery or re-runs vision scoring for a row that already has it.
6. The 32.0s±0.1s duration assert and the loop-frame-reuse check both run BEFORE a video is uploaded or written to the DB, not after.
7. Everything renders to status='pending_approval' (M8) — this skill never flips status to approved/posted and never calls a publish/social API.
8. Client-facing text (title, caption, on-screen overlays) never names an internal vendor, tool, GitHub issue number, or run ID (M3).

Quality Gates

• verify: every chosen title passes validate_bolt32_title() with zero reasons before it is used anywhere (script, caption, on-screen text).
• confirm: the assembled video's ffprobed duration is within 32.0s ±0.1s AND tts_model == 'eleven_v3' before the DB row is updated.
• check: the row's status is not 'approved' before writing any bolt32 field (refuse and report blocked_approved_row if it is).
• ensure: the end beat's still image is byte-identical in source to the hook beat's still image (same file), so the loop is a real frame match.
• call_out: any candidate/field that could not be validated (missing imagery, missing short link, all 5 titles failing) is reported as an explicit error, never silently skipped or defaulted.

Return Format

scope (which row/case this render covers) → finding+evidence (which title candidates passed/failed and why, the ffprobed duration, the tts_model actually used) → intervention (what was rendered/written) → validated (the 4 quality gates above, pass/fail each) → residual (any DoD item not reachable this run, e.g. no case names on file for the banned-name check, imagery missing so bolt32 can't run yet).

Guard Rail

Do not fabricate a numeric fact, a person name's absence, or a passing title-validator result to force a render through — a row with no usable title candidate, no imagery, or a failing duration/tts_model assert is a reported error, not a render.

CTA/link system + title diversity (issue #19786, added 2026-09-03)

Every bolt32 render now carries a persistent URL chip + QR plate (build_cta_chip_png()/build_qr_plate_png()), a spoken CTA in the loop-line beat (assert_bolt32_spoken_cta()), and code-enforced safe-area text wrapping (dt_wrapped_centered(), raises SafeAreaViolation). Since issue #20191 the chip+QR fade in at 24.0-24.3s and fade out 30.5-31.0s (was a 24.0-32.0s hard cut, which held both onto the same 'end' frame the loop mechanic reuses from the hook and broke loop_seam_continuity QA) -- fully opaque and readable 24.3-30.5s (6.2s), above the >=4.0s floor. Director QA runs OCR (tesseract, cropped to the chip's own bbox -- whole-frame OCR over a busy aerial photo is unreliable, live-reproduced) and QR decode (pyzbar) against the actual rendered frame at 26.0s/30.0s (moved off the old 31.5s point, now inside the fade-out/clean loop-seam window) before any DB write. gen_short_code() draws from an OCR-safe alphabet (excludes 0/1/I/i/L/l/O/o) for this reason.

For a BATCH of reels (not a single row), use scripts/biddeed_reels_pipeline_bolt32.py --cta-batch <ids> instead of --ids -- it calls assign_batch_diversity() to round-robin title frame + emoji pair across the batch (check_batch_diversity() enforces no more than 2 repeats of either in any rolling 10-item window) and computes each row's archetype (compute_bolt32_archetype()) for S2 intent-routing on the deal page. --ids/process_row_bolt32() still works for a single fresh row but does not run diversity assignment or the QA gate.