Explaining
Explaining
5 topics × 10 subjects video-explainer arc. Janet narrates animated cards + D3.js / Cytoscape.js graphics over a Remotion → HeyGen → DaVinci Resolve pipeline. Scripts only here; rendering downstream.
Stage 1–6 pipeline
Script (here) → Claude Design bundle → Remotion B-roll → HeyGen narration → DaVinci Resolve composition → S3 publish.
Script
Stage 1 — Script. Janet narration with per-beat timing, card cues, and graphic specs, authored here as scripts/<topic>/<n.m>-<slug>/long/script.md. Every claim is anchored to a repo file; the public-messaging floor applies (concede-and-preempt legal voice, synthetic Doe v. Acme demos only). 19/50 subjects scripted.
Claude Design
Stage 2 — Claude Design bundle. Per-video design.md prompts drive the G1–G7 card + B-roll component set under artifacts/<slug>/. Reads brand tokens from rendering/ — no hex literals in rendering code. 19/50 subjects have a design bundle.
HeyGen
Stage 4 — HeyGen narration. Janet (locked Avatar-V voice) reads each beat as a 9:16 take with mandatory silence handles. Fires only after the Stage-3 b-roll gate lifts (b-roll-first, 2026-07-11). 11/50 subjects have recorded takes on disk.
Remotion
Stage 3 — Remotion B-roll. React + brand-token compositions (plus a Blender plate lane where declared) render to ProRes 4444 cue clips anchored to markers.csv word boundaries. Under the b-roll-first gate the full comp set is authored and visually quality-gated before any narration take fires.
DaVinci Resolve
Stage 5 — DaVinci Resolve composition. Off-Resolve PIP decoration, segmented caption bake, and loudness normalization compose the take, B-roll, and captions into a finished master, driven by the per-episode phase{0..9}_*.py scripts over the resolving/ wrapper. 8/50 subjects have rendered cuts.
S3 / CloudFront
Stage 6 — Publish. Finished masters upload to s3://qagents-videos/<topic>/ and serve immutably from videos.quantapix.com via CloudFront. explaining/ is NOT_YET_LIVE as a status surface — its runtime is the downstream CDN.
5 × 10 video arc — progress
Pipeline-stage counts over the 50-subject arc, with per-topic filtering of the artifact-presence matrix below.
| ID | Subject | Profile | Script | Design | Record | Render |
|---|---|---|---|---|---|---|
| Topic 1 — Why theorem provers… instead of semantic searches | ||||||
1.1 | The Hallucination Tax — Where LLMs Lose at High Stakes | P1 | ✓ | ✓ | ✓ | ✓ |
1.2 | What Semantic Search Actually Does — and What It Can't | P1, P5 | ✓ | ✓ | ✓ | ✓ |
1.3 | The Lean4 Kernel — a 10kloc Oracle for Proof Correctness | P1 | ✓ | ✓ | ✓ | ✓ |
1.4 | Predicates — Where Judgment Lives Between LLM and Kernel | P1 | ✓ | ✓ | — | — |
1.5 | Negative Verification — When the Kernel Says "Not Proven" | P1, P2 | ✓ | ✓ | ✓ | ✓ |
1.6 | Axioms vs. Evidence — What We Don't Trust the LLM With | P1 | — | — | — | — |
1.7 | Civil RICO, End to End — Predicate, Sub-Axioms, § 1962 Elements | P1, P2 | ✓ | ✓ | ✓ | ✓ |
1.8 | Title VI — Discriminatory Intent Has Structure | P1, P2 | — | — | — | — |
1.9 | Same Kernel, Different Axioms — Scaling to Financial Contracts | P1, P3 | — | — | — | — |
1.10 | The Cost — Latency, Surface Area, and Auditability as a Feature | P1, P5 | — | — | — | — |
| Topic 2 — Why narrative analysis… or Qnarre | ||||||
2.1 | When a Docket and Its Record Disagree — Why I Built Qnarre | P2 | ✓ | ✓ | ✓ | ✓ |
2.2 | What a "Narrative" Is, Formally — Predicates Over Named Entities | P1, P2 | ✓ | ✓ | — | — |
2.3 | Three-Zone Qnarre — Input, Kernel, Witness | P2, P5 | ✓ | ✓ | ✓ | ✓ |
2.4 | Watching a Proof Elaborate Live — SSE Event Streaming | P2, P5 | ✓ | ✓ | — | — |
2.5 | From DOCX to Lean — How a Complaint Becomes a Claim Object | P2 | — | — | — | — |
2.6 | When the Appendix Is the Source of Truth — RAv:p Record Citations | P2 | — | — | — | — |
2.7 | Pro Se on Appeal — Encoding the Procedural Traps as Predicates | P2 | — | — | — | — |
2.8 | Counterfactual Narratives — What a Defense Story Has to Deny | P1, P2 | — | — | — | — |
2.9 | Stacking Predicates — Civil RICO and § 1983 Over One Fact File | P1, P2 | — | — | — | — |
2.10 | What Qnarre Never Does — Write Your Brief for You | P2, P5 | — | — | — | — |
| Topic 3 — Why result evaluations… or Qresev | ||||||
3.1 | Backtests Lie — the Survivorship and Look-Ahead Problem | P3 | ✓ | ✓ | ✓ | — |
3.2 | Six Defined-Risk Strategies — and a Hard Refusal of the Rest | P3 | ✓ | ✓ | — | — |
3.3 | Five Frameworks — TREND, MOMENTUM, OPTIONS-RISK, SECTOR, DRAWDOWN | P3, P1 | ✓ | ✓ | — | — |
3.4 | Why a Sector Cap Needs a Kernel, Not a Spreadsheet | P3 | — | — | — | — |
3.5 | One Bar Schema, Ten Consumers — the OHLCV Parquet Hub | P3, P4 | — | — | — | — |
3.6 | Portfolio In, Verdict and Witness Out — Qresev Live | P3 | ✓ | ✓ | ✓ | — |
3.7 | Drawdown as a Theorem — the Conservative Mandate as a Prop | P3 | — | — | — | — |
3.8 | Where the LLM Lives — Predicate Judgments Over Price Action | P1, P3 | — | — | — | — |
3.9 | Three PMs Disagree, the Kernel Doesn't — Per-PM Axioms | P3 | — | — | — | — |
3.10 | What Qresev Refuses — Naked Options, Leverage, Look-Ahead | P3 | — | — | — | — |
| Topic 4 — Why rigorous debates about… technical analysis | ||||||
4.1 | The TA Bestiary — 117 Indicators, Mostly Redundant | P4 | ✓ | ✓ | ✓ | — |
4.2 | TA-Lib as Ground Truth — the MACD EMA-Realignment Quirk | P4 | ✓ | ✓ | ✓ | ✓ |
4.3 | DuckDB and Parquet — Why Columnar Beats Row-Store for Bars | P4, P5 | ✓ | ✓ | — | — |
4.4 | Lightweight-Charts v5 — 10 Years of Bars at 60fps | P4 | — | — | — | — |
4.5 | Aggregators — Small-Multiples, Heatmaps, Three.js Surfaces | P4 | — | — | — | — |
4.6 | Three PMs on One Feed — Where the Disagreements Live | P3, P4 | — | — | — | — |
4.7 | Even the Chart Viewer Refuses Naked Legs — the Options Floor | P3, P4 | — | — | — | — |
4.8 | The Bar You Saw vs. the Bar You Traded On — Data Sourcing | P4 | — | — | — | — |
4.9 | "Tech" Is Not a Sector — GICS as the Only Referent | P3, P4 | — | — | — | — |
4.10 | What an Indicator Is, Formally — TA as Testable Predicates | P1, P4 | — | — | — | — |
| Topic 5 — About Quantapix… the two-teammate practice (one engineer + an AI assistant) | ||||||
5.1 | Two People, One Repo — What "Jack and Janet" Means | P5 | ✓ | ✓ | — | — |
5.2 | The qagents Monorepo — 20 Subprojects, One Venv, One Workspace | P5 | ✓ | ✓ | — | — |
5.3 | Claude Code as the Third Teammate — Agents You Can Fire | P5 | — | — | — | — |
5.4 | CLAUDE.md — Instructions for the Colleague Who Never Reads Twice | P5 | — | — | — | — |
5.5 | memsearch — Remembering What Was Said Last Tuesday | P5 | — | — | — | — |
5.6 | The 5×10 Plan — Made by Claude, for Claude to Make | P5 | — | — | — | — |
5.7 | Two Products From One Kernel — Qnarre and Qresev | P1, P5 | — | — | — | — |
5.8 | The Legal Arc — From Pro-Se Filings to a Verifier Product | P2, P5 | — | — | — | — |
5.9 | The Financial Arc — From Analyzing TA to Evaluating Portfolios | P3, P5 | — | — | — | — |
5.10 | What Quantapix Is Betting On — Kernels Under Everything | P1, P5 | — | — | — | — |
| 19/50 scripted · 19/50 designed · 11/50 recorded · 8/50 rendered | ||||||
Metrics
- topicCount
- 5
- subjectsPerTopic
- 10
- totalScripts
- 50
- scripted
- 19
- designed
- 19
- recorded
- 11
- rendered
- 8