DIRECTOR BRIEF — Pre-Existing Content Ingestion & Brenda Guardrails
Date: 2026-07-25
Prepared for: The Director
Subject: Letting influencers upload content made OUTSIDE WhamWorld (any provider, any tool) while keeping the Kingdom standard enforced — images AND video.
1. THE PROBLEM
Influencer content in the wild is produced across a huge range of external tools (Canva, CapCut, Midjourney, phone editors, other AI providers). We cannot replicate that diversity in-platform, and we shouldn't try. The winning move is: accept anything, gate everything.
The gate must respect the existing opt-in doctrine:
- Opted IN to Brenda moderation (brenda_corrective_opt_in, default true) → Brenda scans, and where the violation is correctable, she fixes it herself and the content proceeds.
- Opted OUT → no corrective touch. Any violation = straight to the Rejected category. No negotiation, no rewrite.
2. WHAT ALREADY EXISTS (we are not starting from zero)
- Caption/text gate (gateCaptionText, shared textGateCore): LIVE — clean / corrected / rejected verdicts
- Brenda Redraw for static images: LIVE in postcardManager submit_for_review — reads text INSIDE the image, rewords it, regenerates the image, auto-publishes
- Amber vision scan (photo pipeline): LIVE for social posts/stories
- TranscribeAudio integration (Whisper): available, unused in this pipeline
- brenda_corrective_opt_in on WhamWorldUser: exists — currently only wired to the social text gate
- Rejected List tab in Influencer Dashboard: exists
GAP: none of this is wired for bulk pre-existing uploads, and nothing scans VIDEO in the influencer pipeline.
3. PROPOSED PIPELINE — STATIC IMAGES (LOW complexity)
This is essentially generalizing Brenda Redraw from "one postcard at submit" to "an upload inbox."
1) Influencer uploads image(s) into a new Content Library area of the dashboard (multi-file upload).
2) Each image enters status "scanning".
3) Brenda runs a TWO-layer pass per image:
- Visual layer (Amber-style): nudity/modesty/violence/occult imagery — pass/fail with context discernment.
- Text-in-image layer (existing Redraw scan): profanity, hostile attacks, slander, crude/divisive language.
4) Verdict routing:
- clean → status Ready (usable in postcards/campaigns).
- correctable + opted in → Brenda regenerates the image with corrected wording (existing Redraw machinery) → Ready, flagged "Brenda-corrected" with before/after visible to the author.
- correctable + opted out → Rejected (with reason). This is the doctrine the CEO specified.
- rejected (unredeemable) → Rejected regardless of opt-in.
5) Every action logged to MAP (action_log) for the audit trail.
Estimated build: 1-2 sessions. Reuses ~80% existing machinery.
4. PROPOSED PIPELINE — VIDEO (the hard part, MEDIUM-HIGH complexity)
We cannot "redraw" a video. The honest scope is: video gets scanned and judged, but never corrected. Correctable-quality violations in video still route to Rejected (with a precise, timestamped reason so the influencer can fix it in their own editor and re-upload). This keeps Brenda's corrective touch honest — she only corrects what she can actually author.
Scan architecture (all with existing integrations — no new vendors):
Layer 1 — Frame sampling (visual + on-screen text):
- Client-side (browser video + canvas, same technique as our existing reel thumbnail capture): extract N frames — first, last, and one every ~3 seconds (a 30s reel = ~12 frames; 60s video = ~22 frames).
- Frames upload as a batch; Brenda's vision pass runs on ALL frames in a single AI call — one verdict covering visual content AND any on-screen text/captions/overlays, with the offending frame index (→ timestamp) identified.
Layer 2 — Audio (spoken content):
- Extract/upload the audio track (or the video file itself — Whisper accepts mp4/webm ≤25MB, and our reels/videos are already compressed under that).
- TranscribeAudio → transcript → run through the EXISTING text gate. Profanity or attacks in speech = violation with the transcript excerpt quoted in the rejection reason.
Layer 3 — Caption/title: existing text gate, unchanged (this part IS correctable for opted-in users, since Brenda authors the caption text herself).
Verdict routing for video:
- All layers clean → Ready.
- Any violation, opted in or out → Rejected, but the rejection reason is surgical: "Frame at 0:14 shows X" / "Audio at ~0:22 says '…'". Opted-in users get Brenda's suggested rewording of the offending line as guidance for their re-edit.
- Hard-exclude content (hate, obscenity) → Rejected + MAP log + Papa G escalation per existing thresholds.
Estimated build: 2-3 sessions (frame extractor UI, scan orchestration, timestamped verdict UI).
5. COST & ABUSE GUARDRAILS (Director decision points)
1) Scan budget: a video scan = 1 vision call (multi-frame) + 1 transcription + 1 text-gate call. Recommend a per-tier daily upload cap mirroring the AI-generation daily cap formula (e.g., ceil(slots × 3) uploads/day) so a bad actor can't burn integration credits.
2) File limits: keep the existing 30s reel / 60s video caps and ≤25MB post-compression — this also keeps us inside Whisper's limit for free.
3) Trust-tier sampling: reuse the Phase 3 Trust Trend machinery — 850+ rising accounts could get relaxed frame density (every 5s instead of 3s), falling accounts get full density. Optional, phase 2.
4) Re-upload loop abuse: a rejected video re-uploaded unchanged should be detected (file hash) and auto-rejected without burning a new scan.
6. OPEN QUESTIONS FOR THE DIRECTOR
1) Should video rejection for opted-in users offer a one-tap "let Brenda re-voice the caption" partial fix, or is full rejection cleaner doctrine?
2) Does pre-existing content skip Mitchel's queue when Brenda clears it (like Redraw does), or does uploaded third-party content ALWAYS require a human/Mitchel pass on first use?
3) Content Library storage: do rejected uploads get purged after N days, or retained for appeal?
7. RECOMMENDED PHASING
- Phase A (now): Content Library + image pipeline (reuses Redraw). Ship fast, prove the flow.
- Phase B: Video scan (frames + transcript), reject-only with timestamped reasons.
- Phase C: Trust-tier sampling relaxation + re-upload hash dedupe + appeal flow.
BOTTOM LINE: images are a near-free win; video is fully feasible with the integrations we already have — the discipline is that Brenda JUDGES video but never EDITS it, and opted-out users get the clean, doctrinally-pure straight-to-reject path.