# Fable adversarial pass

## (a) Cost integrity — the "cheap" claims don't survive arithmetic

**The verification pass is mispriced by design, not by amount.** "One Haiku-class pass per edition" is not what you described. You described: per-claim refutation + per-link fetch-and-resolve. That's N model calls scaled by claim count, plus HTTP fetches whose HTML bodies — if you feed them to a model to confirm the link *supports* the claim rather than merely *resolves* — are the single biggest token sink in the whole system. A resolved 200 on a page that says something else is worse than a dead link. Decide now which one you're building, because they differ in cost by ~50x. Also: a cheap-model refute pass with no retrieval grounding is vibes. Haiku saying "seems plausible" verifies nothing; it launders unverified content into a "verified" label. That's a credibility landmine for someone whose pitch *is* verification.

**Preview + one re-roll can silently double your daily spend.** If you use it every evening — and you will, because you built a studio to use it — that's 2x edition renders/day forever. The bound is "one re-roll," but the *preview itself* is a full render. Cache aggressively or accept the 2x.

**The sleeper cost is arXiv, not finance.** cs.LG + cs.CL + cs.AI is routinely 300–600 papers/day. "Top-N filtered by your rules manifest" means scoring hundreds of abstracts daily with a model. That's plausibly more tokens than the entire rest of the daily edition. Nobody in the document priced it.

**Mentor pilot: "configuration, not construction" is false on cost.** Each mentor manifest with their own topics means their own source pulls, their own scoring, their own render, plus a report sheet (a new artifact type) — 2–3 parallel pipelines, not config rows. Small in absolute dollars, but the framing hides that it's new marginal research spend, exactly what the public-weekly design correctly avoided.

**Genuinely cheap and correctly argued:** public weekly as pure synthesis over seven paid editions. That's the best cost decision in the doc. The 4-story podcast is also honestly priced.

## (b) Manifest — real control only if the renderer can't disobey it

The manifest is real control **exactly where it's enforced structurally and theater everywhere it's interpolated into a prompt.**

- **Real:** private-only section flags enforced in code (the public renderer literally cannot see those sections). Source list as data. Length ceiling as a hard truncation/regeneration check. These are structural — build them.
- **Theater:** "tone," "what kind of messages I want in and out," "topics never-in" — if these become paragraphs in the system prompt, you've built a YAML frontend to the same prompt nobody edits. Config theater with a versioning veneer.
- **The diffing claim is the biggest overreach.** "'The signal got worse' is diffable to which rule change did it" requires an *output quality measure*. A versioned config with no eval harness gives you correlation storytelling, not attribution. Non-determinism in the model swamps most rule deltas. If you want diffability, you need a golden set: re-render the last 7 days' inputs under old-manifest and new-manifest and diff the outputs at save time. That's the feature. Without it, delete the diffability claim from the pitch.
- **Cost-impact-before-save** is theater unless it's derived from measured per-source token/fetch history. An estimate a model makes up is worse than no number — it trains you to trust fake precision.

## (c) Mentor pilot — scope creep wearing a reuse costume

It reuses the pipeline; it does not reuse the *validation*. You have N=1 quality evidence (you), a verification pass that doesn't exist yet, and you're proposing to charge money for a derivative before the core is proven. "Premium pilot" means pricing, delivery, expectations, and a mentor telling someone your feed shipped a bad claim — with Byron's name on the mechanism. It's also the only P3 item that competes for attention with the thing that actually compounds (the weekly papers digest → LinkedIn material). Correct move: keep the *idea*, gate it behind a measured bar — e.g., 4 consecutive weeks of the personal daily with verification pass rate >95% and zero silent failures — and run it first as a **shadow pilot**: generate 2 mentor editions, review them yourself, ship nothing. If shadow output is good, *then* it's configuration.

## Financial data — definitive verdict

**Split it, and the split is not negotiable.**

- **Build now:** funding/M&A/deals. It's already flowing through your news feeds, it's editorial not numerical, it's zero new sources, and your existing verification approach (link resolves, claim is supported) actually works on it. Ship it in P2.
- **Do not build now:** market moves via free-tier APIs. Free tiers (Alpha Vantage's 25 req/day, yfinance's scraping fragility, Finnhub's caps) give you delayed, rate-limited, occasionally wrong numbers — and your verification pass **cannot check them** (refute against what? another free API?). A newsletter whose brand is "citations-always, check-on-check" printing a stale or wrong price is a self-inflicted credibility wound in the one section readers can trivially fact-check on their phone. A wrong number is strictly worse than no number. Wait for a real data source — even a cheap paid one ($30–100/mo tier) — and add price lines then. The "AI basket" also needs a maintained constituent list; that's a curation job the doc hand-waves.

---

## Three highest-leverage changes

1. **Build the golden-set regression harness into the studio, or drop the diffability claim.** On manifest save: re-render the last 7 days under old vs. new manifest, show the diff. This converts the manifest from config theater into a measured instrument, and it's the only way "which rule change made it worse" is ever answerable.
2. **Split verification into two tiers and price each honestly.** Tier 1: deterministic link resolution + quote-in-source check (near-free, no model, run on everything). Tier 2: grounded refutation (retrieval + model) only on claims tagged high-stakes. Emit a per-edition verification scorecard: claims checked / dropped / flagged. Without the tiering, verification is either too expensive to run or too shallow to mean anything.
3. **Reorder P3: gate the mentor pilot behind the quality metric from change #2, and pull the weekly arXiv aggregate digest forward.** The through-line digest is the compounding asset (LinkedIn, credibility, eventually the mentor product's actual content). The pilot is a distribution experiment that's worthless until the content is proven.

## The bold idea that's missing

**Publish the verification ledger.** Every public weekly ships with claim-level provenance: each claim, its source link, fetch timestamp, and refute-pass result — visible, not buried. Every AI newsletter on earth is unverified prose; none can show their work because none do the work. You're building check-on-check anyway — making the ledger *public* turns an internal cost center into the differentiator, and it's the one move that converts Byron's personal credibility stake from a risk into the product's moat. "The only AI signal that shows you its receipts" is a positioning nobody in this space can copy without rebuilding their pipeline.