Audit Hub
Seven scoring layers plus a re-check against platform rules turn "rejected, no idea why" into "which line, which rule, what to rewrite".
/ Internal Hub — Capability Overview
Creative, media and reviews —
all in one hub
From creative intake and framework breakdown to compliance scoring, daily media data and report delivery — all inside one interface. This page is a capability overview: accounts, brands and figures are redacted.
5pipelines
14dedicated roles
8hard guards
AuditLibraryResearchMediaHosts
/ 01 — Modules
Six modules, one data chain
Creatives in, verdicts out, data back. Each module works on its own; together they close the loop.
Creative Research
Two tracks: in-house performance on one side, external structural frameworks on the other. Counted separately, compared structurally.
Library & Visual Judgement
Paged browsing with on-demand thumbnails; audio-to-visual relations and per-shot voiceover are queryable, so even un-indexed creatives can be judged.
Daily Media Data
Pulled automatically every day: fields aligned, anomalies flagged. Silent failures surface the same day, not when someone notices.
Reports & Delivery
Chart types locked to a palette, full-page report assembly, deck export — fonts embedded, so it opens offline.
Host & Content Performance
Host-level curves and content-structure attribution side by side — mapped straight to "what to change in the next session".
/ 02 — Workflow
Four steps, one pipeline
Roles are split by action, not by person. Every step has a defined output, so a problem can be traced to a step.
STEP 01
Collect
Creatives ingested, fields cleaned, identities mapped — get the data right first.
STEP 02
Calculate
Align definitions, then compute spend, ROI, delivery share and concentration.
STEP 03
Produce
Charts, full-page reports, decks and finished cuts — all to spec.
STEP 04
Verify
Numbers cross-checked, layouts screenshot-verified, cuts frame-scanned. Not done until it passes.
/ 03 — Roles
Five pipelines, fourteen dedicated roles
One role, one job. When scope grows, add a role — never stretch one role to cover everything.
Media dataData
Metric guardianTime windows, delivery thresholds, refund formulas — confirm definitions haven't drifted before any report
Spend & efficiency auditComputes spend, ROI, share of delivering creatives and head concentration
Data cleaningMisaligned fields, missing values, duplicate rows in exports
Creative contentCreative
Hook miningPain point, efficacy, visual and structure dimensions — find high-potential items in the existing library
Creative gradingDelivering / losing / unspent — decide what to scale and what to stop
Script diagnosisStructural problems from a writer's perspective, with no data output
Report deliveryDelivery
Chart selectionChart type by data semantics, locked palette, aligned with the canonical definitions
Report assemblyCharts and conclusions into a full page, with complete footnote definitions
Font embeddingThe last step before sending — guarantees it opens offline
Deck assemblyTurns conclusions into pages that can be presented
VideoVideo
Render executionRuns the render and exports the final cut once the comp is locked
Cut verificationBlack frames, overlapping labels, colour handoffs — checked item by item
QualityQA
Number cross-checkIndependently recomputes KPIs, hunting definition/number mismatches
Screenshot proofingHeadless-browser captures to catch overlaps, gaps and missing renders
/ 04 — Guards & Self-Improvement
Eight guards, written into the code
Not discipline — mechanism. Five of the eight are where the model gets sharper with use: verdicts flow back, the lexicon grows, thresholds follow the category.
G/01
Verdict feedback
When the platform's real verdict lands, every line is reconciled into "same spot / same class / missed"; anything missed goes straight into the lexicon, so the next batch won't repeat it. This is where the model corrects itself.
G/02
Self-growing lexicon
Rejection reasons are mined into sentences automatically, each with its source, hit count and effective date. New entries sit in an observation pool until they have enough samples.
G/03
Versioned rules
Every re-judgement keeps a version: which line triggered it, which version ruled, how much it changed. Any conclusion traces back to the source text.
G/04
Three-engine cross-check
Transcription, context ruling and rule matching run independently; divergence beyond a threshold is flagged. Humans only read the disagreements; the rest passes in bulk.
G/05
Thresholds by category
Pass rates differ by category. Scores are calibrated against that category's history plus the account's own baseline — no single ruler for every creative.
G/06
No self-reporting
The model's "I fixed it" and the independent re-check are recorded separately; when they disagree, the re-check wins. Self-report is not evidence.
G/07
Sample floor
Fewer than N shifts or N creatives is labelled "low sample" — no conclusion, no place in the summary. Small samples must not bend the parameters.
G/08
Daily health check + market attribution
Task states are scanned daily; failures surface the same day. When scoring people, market drift and group medians are subtracted first — never book the market's bill to an individual.
/
All eight live in the code
Not written in a handbook for people to follow — grown into scripts and tables: gaps raise, drift flags, thin samples stay silent.
/ 05 — Engine
All six modules run on
DeepSeek Flash V4.1 (1017)
Rules aren't hard-coded — judgement is handed to the model, and nothing is computed until a creative arrives. Every clip is read for voiceover, visuals and batch context before its rules are chosen.
E/01
Dynamic frame sampling
Not fixed-interval sampling. The model picks frames by voiceover pauses and visual changes — a 30-second clip usually needs only 5–8 frames to be judged.
E/02
Dynamic criteria
Criteria are generated on the spot, not looked up. The same script under a different category or account yields different risk points — the same sentence judged differently in a different context.
E/03
Dynamic orchestration
The creative type decides how many chains run: voiceover-only takes two passes, visual creatives three. What it will cost is stated before it runs.
E/04
Instant write-back
Results land in the library and dashboards immediately and feed the next analysis. While a person reads the output, the next batch is already running.
- Model
- DeepSeek Flash V4.1 · build 1017
- Coverage
- 6 modules connected end to end, no manual handoffs
- Upgrade
- Swapping the version takes effect immediately — no rule rewrite
About the data on this page
This page is a capability overview of an internal system. Account names, brand names, creatives and media figures are all redacted: values appear as masks or unscaled placeholders and do not represent any real campaign result. Screenshots match the production layout, but the numbers are not real. For specifics, please get in touch.