qexplaining

50 videos = 5 topics × 10 subjects each. 10–15 min per video, narrated by Janet over animated cards/text + D3.js / Cytoscape.js graphics. Brand-synced with the two product sites and the two product app shells.

A weekly-refreshed window into the explainer arc that runs alongside the private working repository. The output of the work is scripts (plus the data behind every graphic); rendering happens in Claude Design; voice + lipsync in a separate pipeline. This repo publishes the master plan and the scripts as they land.

• Parent organisation: https://github.com/quantapix
• Engineering output: https://quantapix.com
• Motivational record: https://femfas.net

Profile-area tags

Every subject below carries one or two of these tags. Order within a topic is shootable order — earlier subjects motivate later ones.

• P1 — rigorous proofs for LLM "reasoning"
• P2 — legal applications (Family Court → 1st Cir. → Qnarre)
• P3 — financial applications (analyzing + accounting → Qresev)
• P4 — grounding competing technical-analysis approaches
• P5 — agentic software development (Claude Code + monorepo)

---

Status — 2026-05-09

Production cadence has crossed from planning into the first full cuts. The week's deltas:

• 1.0 channel preview — pilot moved from script-only to a full B-roll layer. A-roll (Janet narration) was rendered earlier; the redo adds a five-component B-roll (cold-open stinger, topic-rotator chip strip, 5×10 video matrix bloom, Janet self-credit lower-third, end-card lockup) plus the channel background bed at 8% opacity. Dual-aspect: long-form 16:9 + Shorts 9:16.
• 1.1 The Hallucination Tax — spec rewritten end-to-end against the locked production architecture (see "Working stack" below) without touching Janet's narration text or the rendered HeyGen takes. The script is frozen; the render path is what changed.
• 1.5 Negative Verification and 2.1 The Family Court that ate the file — priority-10 #2 and #3 (after 1.1 at #1) drafted to full per-video shape: Janet narration, Claude Design B-roll prompts, DaVinci Resolve assembly plan, HeyGen call payloads. 1.5 is the topic-1 demo episode; 2.1 is the topic-2 entry.
• Channel background bed locked at round 2: twelve stair-fragment plates rendered as a derivative-not-copy of the impossible-staircase visual idiom (named-Escher motifs excluded). The plate's narrative load is the multi-rational contradiction motif — every plate composes several independently-rational stair flights that share an explicit reference (engraved elevation grid, "datum ±0.00" line, claimed- landing band) and visibly disagree on it. The design IS the problem statement; the same shape Quantapix is designed to detect in legal filings and financial statements.
• Dual-aspect (16:9 long-form + 9:16 Shorts) production lock — same Resolve project, two parallel timelines, identical A-roll audio. The long-form carries Janet as a corner PIP and full-frame B-roll cues; the Shorts cut flips the layout (Janet hero, B-roll in a 1080×280 lower inset band).

What's coming up:

• HeyGen render runs for 1.5 (9 beats × 2 aspects = 18 takes) and 2.1 (8 beats × 2 aspects = 16 takes), tightened lipsync and pronunciation gates.
• Per-cue Remotion B-roll renders against the per-video Claude Design bundles.
• First end-to-end production cycle through the full pipeline (1.0 ships first as the gating smoke test for the cross-aspect Resolve scripting; 1.1 follows; 1.5 + 2.1 inherit the same shape).

Working stack (locked)

• A-roll (voice + face). HeyGen Photo Avatar V against a single locked still. Voice generated via HeyGen's Design-a-Voice (channel voice locked retroactively to the production preview at https://quantapix.com/videos/Janet-preview.mp4). Per-beat generation — one MP4 + sidecar SRT per script beat per aspect — rather than monolithic takes; cheaper re-cuts, tighter lipsync.
• B-roll (cards + graphics + animations). Remotion: React-based programmatic video. Per-video components live alongside the script as JSX in a small workspace package; brand tokens (palette, type, easing) are read through a generated tokens.ts so no hex literals leak into renders.
• Composition + finishing. DaVinci Resolve Studio, scripted via an in-house Python wrapper plus a recipe library that turns per-episode assemble.py drivers from ~260-line scripts into ~80-line ones. The wrapper handles import, two-aspect timeline setup, A-roll comping with breath gaps, B-roll cue placement at marker frames, and back-to-back YouTube-master + Shorts render. Fusion node-tree authoring (Janet PIP corner ring + soft shadow, Shorts inset band, per-cue compositing) is manual at v1, guided by per-skill node-tree force-graph diagrams.
• Captions + loudness. Resolve native subtitle generation from the A-roll track (V4 burn-in plus sidecar SRT for upload metadata); Fairlight loudness normalization to −14 LUFS / −1 dBTP on the master bus.
• Music + SFX. YouTube Audio Library (free, no-attribution- required filter, monetization-safe); per-topic register pivot (T1 ambient electronic / T2 string-pad / etc.). SFX library selected per cue, kept sparing — no whooshes, no risers under Janet.

ElevenLabs is deferred — only revisited if HeyGen's voice quality fails a real production take. So far it hasn't.

---

Topic 1 — Why theorem provers… instead of semantic searches?

Profile mix: P1 dominant, P2 / P3 supporting.

1. The hallucination tax — where LLMs lose at high stakes. [P1] A complaint that hinges on a § 1983 element the model "remembered wrong"; probabilistic recall vs. provable derivation; the cost curve as stakes rise.
2. What semantic search actually does (and doesn't). [P1, P5] Cosine similarity demoed as a magic trick that's not magic; embeddings cluster, they do not derive.
3. The Lean4 kernel — a 10kloc oracle for proof correctness. [P1] "Every proof is checked by a program small enough to read." Type checker as the trust boundary; one .lean file's type elaboration is the verdict.
4. Predicates as the LLM-Lean bridge — where judgment lives. [P1] Not the kernel, not the LLM — the markdown predicate. Three-layer split: kernel + predicate sub-agents + thin driver; each predicate has provenance + cite + question.
5. Negative verification — when the kernel says "no, that's not RICO." [P1, P2] The demo where the build fails, and that's the right answer. sorry on the affected theorem; other theories still elaborate.
6. Axioms vs. evidence — why we don't trust the LLM with everything. [P1] Which knobs are policy, which are facts, which are model output. Axioms first-class with cite + scope; LLM output never enters the trust base.
7. Civil RICO walkthrough — predicate, sub-axioms, verdict. [P1, P2] A single complaint reduced to a Prop. Person → predicate acts → enterprise → pattern → § 1962 elements; each one a provable lemma.
8. Title VI walkthrough — discriminatory intent has structure. [P1, P2] "Intent" sounds soft. It isn't. Protected class, recipient of federal funds, intentional treatment, causation; predicates render each as a question with required cites.
9. Why this scales to financial contracts (and on to TREND / MOMENTUM). [P1, P3] Same machinery, different axioms. TREND, MOMENTUM, OPTIONS-RISK, SECTOR, DRAWDOWN as kernel-checkable predicate vocabularies over an OHLCV trace.
10. The cost — latency, surface area, and auditability as a feature. [P1, P5] 30s elaboration vs. 1s vector search; predicate library upkeep; but every result ships with a witness.

---

Topic 2 — Why narrative analysis… or Qnarre?

Profile mix: P2 dominant, P1 supporting, P5 cameo.

1. The Family Court that ate the file — a real narrative-analysis use case. [P2] A real docket, dispositions misread, a brief built on a wrong premise. Why Qnarre's first job is procedural posture.
2. What is a "narrative" formally? — predicates over named entities. [P1, P2] A story is a graph plus a calendar. Actors, acts, edges, time; the narrative reduces to a Lean structure.
3. Three-zone Qnarre — input, kernel, witness. [P2, P5] A tour of the live /app island: docx/text in (left), kernel elaboration (middle), witness + fail trace (right).
4. SSE event streaming — watching a proof elaborate live. [P2, P5] 12 seconds of a real Lean elaboration as a stream of {stage, msg, kind} events; the events drive the React island's colors.
5. From DOCX to Lean — how a complaint becomes a claim object. [P2] A paragraph of pleading mapped to four structured predicates. Extraction → predicate match → axiom selection → theorem assembly.
6. RAv:p record citations — when the appendix is the source of truth. [P2] Every assertion carries RAII:203. 10-vol Combined Record Appendix; ground-truth pagination by footer stamp; never cite raw filings on appeal.
7. Pro se on appeal — why ordinary people need this. [P2] The asymmetry between counsel and pro se on procedural traps. Rule 4(a), tolling motions, the SO 2-99 page ceiling — encoded as predicates so a brief doesn't trip them.
8. Counterfactual narratives — what a defense story has to deny. [P1, P2] Every theorem has a contrapositive. Opposing counsel's narrative as a competing predicate set; Qnarre shows which axioms they would need.
9. Federal civil RICO + § 1983 — stacking predicates. [P1, P2] One fact pattern, two statutory frames, one kernel. Shared facts file, framework-specific axioms, theorems compose.
10. What Qnarre never does — write your brief for you. [P2, P5] The verifier's modesty. Predicates check, witnesses cite, but legal-strategy and prose are the lawyer's; auditability over autonomy.

---

Topic 3 — Why result evaluations… or Qresev?

Profile mix: P3 dominant, P1 + P4 supporting.

1. Backtests lie — the survivorship + look-ahead industry. [P3] A hand-picked equity curve that explodes when delisted names are removed. Where standard backtests cheat; what a kernel-checked evaluator must refuse.
2. Six defined-risk strategies, hard refusal of everything else. [P3] "We won't even compute it" for naked options. Long calls/puts, debit spreads, covered calls, protective puts — the allow-list.
3. TREND, MOMENTUM, OPTIONS-RISK, SECTOR, DRAWDOWN — five frameworks. [P3, P1] A portfolio is judged five ways at once. Each framework a Lean theorem set with its own predicate library.
4. Why sector-cap claims need a kernel, not a spreadsheet. [P3] "Max 25% Tech" — under whose definition? GICS sectors as the only acceptable referent; canonical 11 names; cap is a provable predicate.
5. OHLCV parquet hub — the same bars feed every framework. [P3, P4] One bar schema, ten consumers. {ts, o, h, l, c, v, adj_c}; vendor names die at the boundary.
6. Live evaluator walkthrough — portfolio in, verdict + witness out. [P3] 90 seconds of Qresev rendering verdicts on a 10-name portfolio. SSE events tagged by framework; the witness card carries the cite chain.
7. Drawdown as a theorem — the conservative PM's mandate. [P3] "No drawdown > 8%" is a Prop, not a wish. How the predicate phrases historical drawdown; how a conservative PM uses it as a precondition.
8. Where the LLM lives — predicate judgments over price action. [P1, P3] A chart segment, a question, an answer with a cite. Structured-output prompts return JSON predicate values; never free text into the kernel.
9. Aggressive vs. conservative — same kernel, different axioms. [P3] Three PMs disagree. The kernel doesn't. Per-PM risk_policy.md becomes per-PM axiom set; same theorems, different verdicts.
10. What Qresev refuses — naked options, leverage, look-ahead. [P3] Refusal as a feature. UI hard-refuse on disallowed legs; kernel rejection on look-ahead reads.

---

Topic 4 — Why rigorous debates about… technical analysis?

Profile mix: P4 dominant, P3 + P5 supporting.

1. The TA bestiary — 117 indicators, mostly redundant. [P4] Scrolling the catalog with a counter that ends at 117. The indicator-inflation problem; equivalence classes hidden in plain sight.
2. TA-Lib as ground truth — the MACD EMA-realignment quirk. [P4] A "wrong" MACD that's actually right. TA-Lib realigns both EMAs to slow-1; naive subtraction misses by one bar; parity is non-negotiable.
3. DuckDB + Parquet — why columnar beats row-store for bars. [P4, P5] "10 years of bars in 30 ms." Column-oriented IO, predicate pushdown, single-file portability.
4. Lightweight-charts v5 — rendering 10y of bars at 60fps. [P4] The chart that doesn't blink at zoom-out. Pinned via npm at 5.2.0; canvas-only rendering; v5 shape.
5. Aggregators — small-multiples, heatmaps, Three.js surfaces. [P4] From one symbol to a sector. Tab-switching; aggregates = symbol sets; operators DuckDB-first.
6. Three competing PMs on the same feed — where disagreements live. [P3, P4] Same bars, three verdicts. Separation of capital + differentiated strategies; per-PM watchlists; disagreement as a signal.
7. The defined-risk options floor — why the chart viewer enforces it too. [P3, P4] Even the chart viewer refuses naked legs. Cross-project rule; UI never offers a leg the runtime would refuse.
8. Alpaca IEX live + yfinance / Stooq — data sourcing tradeoffs. [P4] The bar you saw vs. the bar you traded on. Live IEX vs. consolidated tape; backfill sources; reconciliation at the canonical schema.
9. GICS as the only acceptable "sector" referent. [P3, P4] "Tech" is not a sector — Information Technology is. 11 canonical names; shared parquet, thin readers per project.
10. What an indicator is, formally — turning TA into testable predicates. [P1, P4] "RSI > 70" as a Prop with a witness. Predicate libraries promote indicators from cosmetic to provable; the bridge from analyzing to accounting.

---

Topic 5 — About Quantapix… the two-teammate practice

Profile mix: P5 dominant, all others as illustrations.

1. Two teammates, one repo — what the practice looks like. [P5] Sole developer + expert AI assistant; narrow team, wide repo, explicit CLAUDE.md contracts at every boundary.
2. The monorepo tour — twelve subprojects, one venv, one workspace. [P5] Sixty seconds, every subproject. Cytoscape repo map; edges are shared-data hubs.
3. Claude Code as the third teammate — agents you can fire. [P5] A sub-agent that did one job and vanished. Per-task agents; when to spawn vs. inline; CLAUDE.md as the contract.
4. CLAUDE.md — writing instructions for the colleague who never reads twice. [P5] A short, harsh CLAUDE.md beats a long polite one. Rules over preferences; pointers over prose; the trim discipline.
5. memsearch — how the assistant remembers what was said last Tuesday. [P5] Cross-session recall. ONNX bge-m3, per-subproject collection scope, fork-isolated recall.
6. The 5×10 video plan — made by Claude, for Claude to make. [P5] Outlining as a structured-output task; the design hand-off; the brand-sync constraint; the very plan you're watching being executed.
7. Verifying + Evaluating — shipping two products from one kernel. [P1, P5] Two SSE servers, two Astro islands, one type-checked Lean. How product variation lives above a shared kernel without copy-paste.
8. The legal arc — from pro-se filings to a verifier product. [P2, P5] A 2025 Family-Court motion that became an axiom set in 2026. Dogfooding across Family Court → 1st Cir. → Qnarre.
9. The financial arc — from analyzing TA to evaluating portfolios. [P3, P5] An indicator becoming a predicate becoming a portfolio verdict. Data hub → analyzer → evaluator; each layer narrower.
10. What Quantapix is betting on — kernels under everything. [P1, P5] The thesis in one card. An LLM is fast and frequently right; a kernel is slow and rarely wrong; both, audited together, change which industries can adopt this.

---

Brand sync (must hold across all 50 videos)

• Single source of truth for color / type / spacing is the tokens.css shipped to https://quantapix.com; D3 / Cytoscape themes pull palette from it, never hex literals.
• Wordmark, accent shapes, and closing-card lockup match the live product sites and the two /app shells (Qnarre, Qresev).
• Voice and pacing match the calm, declarative cadence of the site copy. No marketing exclamations. No rhetorical questions. No activist framing.
• Every closing card carries the same Quantapix wordmark + a one-sentence pull-quote.

Cadence

Refreshed weekly from the private working tree. Outline edits, new profile-area tags, and finalised scripts are committed as ordinary diffs — the commit log is the change record.

Per-video deliverables follow a four-file shape under scripts/<topic-slug>/<n.m>-<slug>/:

• script.md — Janet narration with timing, on-screen card cues, graphic-spec slugs, pacing notes for the voice generator.
• design.md — Claude Design prompts for the per-video B-roll bundle; tokens contract; cue-state spec for each Remotion component.
• resolve.md — DaVinci Resolve project preset, V1–V4 / A1–A4 track layout, A↔B sync strategy via marker CSV, render preset.
• SHOOTING.md — production-readiness gate (voice, takes, per-cue motion exports, marker CSV, SFX, music) plus the cut → review → render → publish sequence.

Plus a heygen-aroll.md with the per-beat HeyGen call payloads when the episode goes into A-roll production. Three episodes currently carry the full quad: 1.1, 1.5, 2.1. Channel pilot 1.0 carries an integrated identity README and shares the same production architecture.

Contact

quantapix@gmail.com