Signal Studio — shape the feed, verify it, and close the two holes

One line: a small studio page where you shape what the daily signal is allowed to contain (rules, sources, tone), a check-on-check pass so nothing ships unverified, two new source classes you named as holes (financial data, lab papers), and a cheap public weekly edition derived from the private daily.

Source: tonight's drive-time dictations (2026-07-17). Direct quotes drive each section.

---

The studio — "craft and shape better what was in the signal"

Today the signal's editorial policy lives in a prompt nobody edits. The studio makes it a manifest you own (same pattern as the fleet app manifest):

• Rules panel: topics always-in / never-in, must-have sections, tone, length ceiling, "what kind of messages I want in and out." Every edition renders against the manifest, and the manifest is versioned — with the critique's addition that makes "diffable" true instead of decorative: on every manifest save, the studio re-renders the last 7 days under old-vs-new rules and shows you the diff. "The signal got worse" becomes answerable to which rule change did it, measured against editions you already know.
• Sources panel: the feed list stops being code. Add/remove/weight sources; every added source shows its cost impact before you save (the $-per-edition estimate, same warn-and-steer instinct as the swarm console's $2 rule).
• Preview + one re-roll: tomorrow-morning's edition rendered on demand against the current manifest, with one manual re-roll — bounded, so the studio can't become a token furnace.

Check-on-check — the verification pass

Your phrase: "maybe a check on check kind of thing." Two tiers, priced honestly (the critique's restructure — one flat pass is either too expensive to run or too shallow to mean anything):

• Tier 1 — deterministic, near-free, runs on everything: every link fetched and must resolve; every quoted claim checked against its source text. No model involved.
• Tier 2 — grounded refutation, only on claims tagged high-stakes: retrieval + a cheap-model refute pass.
• Every edition emits a verification scorecard — claims checked / dropped / flagged — and failures are dropped or rendered unverified in de-emphasized style, never silently shipped.

And the critique's bold idea, adopted as the public positioning: publish the verification ledger. Every public weekly ships with claim-level provenance — claim, source link, fetch timestamp, refute result — visible, not buried. Every AI newsletter on earth is unverified prose; "the only AI signal that shows its receipts" is a moat nobody can copy without rebuilding their pipeline, and it converts the personal-credibility stake from risk into product.

The two holes, closed

1. Financial data — "definitely a huge hole," split in two by the critique, and the split is right. Build now: funding / M&A / deals — it's editorial, it already flows through the news feeds, and Tier-1 verification actually works on it (a link either supports the claim or doesn't). Ships in P2 as a ~5-line daily section. Explicitly NOT now: market prices from free-tier APIs — delayed, rate-limited, occasionally wrong, and unverifiable by our own pass (refute against what — another free API?). A feed whose brand is check-on-check printing a stale price is a self-inflicted wound in the one section readers can fact-check on their phone. A wrong number is strictly worse than no number. Price lines wait for a real (even cheap paid) data source.

2. Lab white papers — "a way of understanding the aggregate of all the reports." The source you were reaching for is arXiv (arxiv.org — cs.AI / cs.CL / cs.LG daily listings, free API) plus the labs' own research feeds (Anthropic, DeepMind, OpenAI, Meta AI). Daily: top-N new papers filtered by your rules manifest. Weekly: an aggregate digest — "what the labs actually published this week, and the through-line" — which is the piece that compounds into LinkedIn material.

Editions — private daily, public weekly, cost-guarded

Your instinct, kept exactly: "the public one gets produced once a week… the private one daily as it does now."

• Private daily: unchanged cadence, upgraded contents (finance section, papers section, verification pass).
• Public weekly: derived exclusively from the seven private editions already paid for — one synthesis call, zero new research spend — rendered as a public page + the LinkedIn feeder. Public never contains watchlist/personal items (the manifest marks sections private-only, structurally).

Podcast — more signal per episode, and the mentor play

• Open the daily podcast to 4 stories. Your observation stands on its own: at current length the episode "clearly did not have enough signal." Costs pennies more; the pipeline doesn't change.
• Mentor podcast (premium, gated — not scheduled): your idea to point the same signal→podcast mechanism at the mentor program — each mentor gets a small-scale personal edition (their topics manifest, weekly cadence) wrapped with a report sheet. The critique's correction, adopted: reuse of the pipeline is not reuse of the validation — charging money for a derivative before the core is proven puts your name on an unproven mechanism. So the pilot is gated behind a measured bar (4 consecutive weeks of the personal daily at >95% verification pass rate, zero silent failures), and runs first as a shadow pilot: generate 2 mentor editions, review them yourself, ship nothing. If shadow output holds, the real pilot is configuration.

Roadmap

• P1: manifest + studio page (rules, sources, preview, golden-set diff); two-tier verification + scorecard; 4-story podcast. (The daily is already running — everything upgrades in place.)
• P2: funding/M&A section; arXiv + lab-papers daily section; the weekly labs through-line digest (pulled forward — it's the compounding asset that feeds LinkedIn and, eventually, the mentor product's content).
• P3: public weekly edition with the published verification ledger; mentor shadow pilot once the quality bar is met.

Self-critique — Fable adversarial pass

Full text: fable-critique-20260717.md. What it caught and what changed:

• "Diffable manifest" was config theater without a way to see a rule change's effect. Fixed: golden-set regression — every manifest save re-renders the last 7 days old-vs-new and shows the diff.
• One-pass verification was either too costly or too shallow. Fixed: two tiers (deterministic link/quote checks on everything; grounded refutation only on high-stakes claims) + a per-edition scorecard.
• Free-API market prices would betray the feed's own brand — unverifiable numbers in the most fact-checkable section. Fixed: funding/M&A now, price lines only with a real data source.
• The mentor pilot competed with the compounding asset. Fixed: weekly labs digest pulled forward; pilot gated behind 4 weeks >95% verification + shadow run.
• Its bold idea — publish the verification ledger — is now the public weekly's positioning: the only AI signal that shows its receipts.