SIGNAL — the podcast platform proposal

Why this does NOT break the existing site (path-preserving takeover):

• https://signal.arnao.ai/feed.xml (the live podcast RSS, the URL subscribers/Spotify hold) — byte-path preserved, becomes The Signal station feed, zero subscriber breakage.
• /episodes.json and the episode-card JSON the fleet consumes — paths preserved.
• The current homepage ("the broadcast log") moves intact to the station page at /the-signal; the root becomes the network Lineup with The Signal in the hero.
• The Signal's proven pipeline (draft → Branding Central queue → Byron approves → publish-episode renders + deploys) is untouched — it becomes the platform's first working station and the template every other show migrates onto.
• The existing Signal stack is a coherent ladder the platform completes, and the private tiers are untouched: Signal (Gia's twice-daily private brief: 8am full episode, 6pm delta, ledger-based) → Signal Live (Byron's private real-time experiment: listen/talk/act on the briefs, tailnet PWA) → The Signal (public weekly curated show, approval-gated) → SIGNAL (this platform: the public network carrying The Signal, Do Not Deploy, The Loop, and every future station). One pipeline: raw signal → reaction lab → curated broadcast → network. Labels stay strict in all fleet docs to keep spoken "signal" unambiguous.

v5 = v4 + Byron's naming call. v1→v2 UX+panel in §11; v2→v3 Fable in §12; v3→v4 streaming Lineup in §13.

1. What this is

Promote the Do Not Deploy pipeline from a one-off show into the podcast primitive for Byron's brand: one public platform holding every show, every voice, every format. Do Not Deploy (https://donotdeploy.arnao.ai) and The Loop (https://comedy.arnao.ai) become the first stations. Everything proven — cloned-voice rendering, per-turn audio caching, kill switch, privacy rules, voting, Plausible analytics — is inherited platform-wide, not rebuilt.

The governing thesis (sharpened by the Fable pass): at the level Byron is playing for, scarcity signals authority and throughput signals hobbyist. So the visible product is not the content engine — it is the governance apparatus around the content: provenance edges, dated truthful attestations, a public Disagreement Ledger, and a public log of episodes Byron refused to ship. The platform generates easily precisely so the refusals are cheap enough to be real.

2. The public experience — the Lineup

Streaming-service DNA, tuned by asking. Landing = a home screen structured like the platforms people already know (Netflix/Prime/Spotify) — but the headline interaction is the agentic tuner: one big command bar. "5 minutes on data centers, both sides, two voices" is the modern equivalent of turning a dial — tuning by intent instead of frequency. For visitors it's search-plus-filters over the catalog; for Byron (and, gated, for invited creators) the same bar composes new episodes on the fly — the input shows cost tier, render ETA, and "→ review queue, not air" before anything runs. Bold visual language: deep-space ground, an animated electric spectrum bar as the brand device, one neon band color per station, glassy animated jacket cards.

• Hero billboard: latest flagship episode (The Signal weekly) — provenance badge, dated attestation, and citation count in the hero: governance is the billboard copy.
• Rails (snap-scroll, keyboard-operable): Continue Listening (progress bars) · New This Week · per-station rails · Because you listened to… · The Cutting Room (Spike Log + Disagreement Ledger as browsable content cards) · languages (text chips, never flags). Rails render only at ≥3 real items — curated, not infinite.
• Station types are a visible primitive: every station carries a HUMAN-CURATED or AUTO badge — the proof-of-concept split Byron asked for. Human-curated = attestation-gated (The Signal, DND); Auto = judge-gated machine cadence, honestly labeled (e.g. Field Notes shorts). The badge is the RAI story told in one glyph.
• Behavior-personalized rail order (2026 pattern): anonymous = localStorage; signed-in = synced. Optional Google sign-in stores resume + follows, nothing else, stated under the button; RSS and playback never require it.
• Narrowcast — the audience-of-one primitive, first-class on the page: "Gia — send Bob a 5-minute podcast on data centers: my RAI point of view vs the latest from Google. Two voices. Cite everything." → renders a private episode → link + QR + email/text delivery (email via the proven gog-in-gia path; never on the public wall). Narrowcasting is the same primitive as broadcasting with an audience list of one.
• DOM-first is load-bearing: every episode a real anchor; accessibility is a launch gate (WCAG-AA, keyboard-complete, screen-reader pass, prefers-reduced-motion, per-episode lang, transcripts inline, AI-disclosure in audio AND text).
• Jackets: Nano-Banana infographic cover + series/ep/date/duration, transcript, citations with live URLs at the point of claim, cross-links to Byron's surfaces, provenance edge in station color.
• RSS stays first-class: master + per-station feeds, provenance-as-text, chapters. Syndication: per-station RSS is the export surface — The Signal's feed is already Spotify-submission-ready (paste-once per README); each new public station gets the same one-time submission to Spotify/Apple/YouTube Music. The platform is the source of truth; directories are distribution.
• Fold behavior by aspect ratio; audio survives re-layout; nothing precise on the crease.

Visual direction (from the design review — keep all of it)

Material over hue: milled aluminium + warm glass, never black-plus-neon. Palette: umber chassis #17130E · shadow-cut #2A241D · brushed face #B6B0A4 · dial lamp #F2B45C (light source only — bloom, edge; never a fill) · meter blue #9CD6E6 (verified states) · dial-face glass #E4E6DE (light mode = the unlit printed dial, not an inversion). Type: Chivo Condensed caps for nameplates; IBM Plex Sans with JP/KR/SC/TC siblings (chosen for script coverage — multilingual is first-class); IBM Plex Mono tabular figures for every number. Motion: transform+opacity only, 180ms snaps, no ambient idle loop (the strongest AI-tell there is).

3. The trust layer — the actual product

• Tiered, truthful attestations. Every approval publishes a dated attestation that states what the review was: Listened in full / Claims verified against sources / Summary reviewed. Never a bare "reviewed by" over a skim — Fable's sharpest finding: a vague attestation converts a pipeline error into a personal credibility event.
• Judge / Council toggle per episode & per-show default: Off (satire edge) → Judge (cheap model verifies every claim has a live citation; failures block approval) → Council (3 panelists vote per claim; verdicts + citations into public show notes). Known limit, stated publicly: LLM panels are vendor-diverse, not epistemically diverse — which is why the citation-existence check is mechanical and why the human attestation, not the council, is the final authority.
• The Disagreement Ledger — P1, retroactive. The council's disagreements published verbatim, including claims that shipped anyway and the human sentence on why. The one artifact a governance director would forward.
• The Spike Log (Fable's bold idea — adopted). A public page of episodes Byron refused to ship: council verdicts, failed claims, one human line on why each died. A live, dated kill-rate makes human-in-the-loop falsifiable instead of asserted, and inverts the volume problem — every rejection strengthens the brand. "Here is what my machines made, and here is what I wouldn't put my name on."
• Auto-approve hard-locked OFF for RESPONSIBLE AI / AI stations, and Byron's voice never attaches to an unread claim where being wrong matters. Comedy adjacency is Open Decision #6 (same wall vs linked-but-separate domains).

4. The owner experience — Mission Control (single screen)

Owner-gated (Career-Triage auth pattern), four zones:

A. Generate — one command bar: "5-minute smoke test on data centers, pro AND con, with citations, my voice + Vera, funny-leaning" → parsed into show/format/length/voices/mood/citations/judge/language. Presets + steering box, never a wizard. Model tier is a per-show policy set once (override available), not five switches per episode.

B. Review queue — cheat sheets, not audio files: exec summary · claim/citation status · privacy-scan hits · cost · length; adversarial spot-check ▶ (30s assembled from privacy hits + judge failures + the two highest-uncertainty claims — risky lines, not highlights); diff-from-last for recurring shows; Edit-line re-renders one turn (per-turn cache) with A/B listen-back; Approve ≠ Publish (approve = cleared-for-air; publishing is a scheduled slot per the standing cadence rule — batch approval can never batch-release). Per-show policy dial Manual → Cheat-sheet → Auto (comedy only). Approval needs land in the surfaces Byron already checks: needs-byron slog + Signal digest + Desk. No new inbox.

C. The governor — caps review debt, not just dollars. Fable's correction, adopted: dollars were never the scarce resource; Byron's attention is. Generation pauses when pending approvals exceed N (default 3). Secondary: month-to-date spend meter per show with a hard monthly cap, and a live job-state strip (queued → scripting → judging → rendering → jacket, elapsed, stuck-alarm, cancel) so silence is never ambiguous. The dashboard also shows minutes-of-review per published minute — the honesty metric behind the attestations.

D. Fleet hooks — the published primitive. Generate is also the fleet skill podcast-generate, published to every agent (Gia/Mia/Cia/Tia + Claude Code): "Gia, make me a funny 5-minute episode on datacenters in Japanese" from Telegram works with the same toggles and the same governor. Companion skill podcast-send handles narrowcast delivery (link + QR + email/text via the gog-in-gia path). The skill enqueues a typed job (proven voice-platform Blob-queue + Mac watcher); agents never run the pipeline inline — the fix for the million-token translation burn: outline → segments → per-segment render, each a bounded fresh call, per-segment caching, Batch API (50% off) for anything async. Context ingestion: generation jobs pull from what the fleet already knows — OB1 recall, Gia's daily scrape (podcasts, X/AI, Reddit, dissenting voices), the Signal ledger — so "my point of view on data centers" means Byron's actual recorded positions plus this week's actual discourse, not a cold-start essay. Channels (auto-creating stations) are declared per show: cadence + model tier + cost ceiling + curation mode (human-gated vs judge-gated auto), all enforced by the governor.

Model tiers (per-show policy, real prices, honest columns)

Voice render local Chatterbox $0 + Gemini TTS characters (cents); jacket ≈ $0.13. GPT-5.6 Sol ($2.50/$15 — OpenRouter's 50% promo through Sep 18) = optional High alternate; a promo, not a floor. GLM-4.6 also = designated cheap coding model for platform build chores. And per Fable: even at 10× these numbers the dollars are noise — the governed resource is attention, hence zone C.

5. Voices

• Byron's voice = local Chatterbox, never ElevenLabs. Character personas (co-hosts of any flavor) = proven recipe system: one cloned voice + style/prosody recipes, plus the Gemini-TTS cast already in EXTRA_VOICES (Vera, Lincoln, Jefferson, announcer…). New characters = new recipe rows, not new infra.
• Voice registers per episode (#EXAGGERATION/#TEMPO): comedy read · "center of authority" · serious/security.
• Public voice QA mandatory: pronunciation lexicon (RAI → "R A I"), canonical Chatterbox settings, whisper-back QA diff, AI-disclosure every episode, every language, audio + text.

5b. Live proof — the Judge gate, tested on real content

Byron asked for a test; it ran on 2026-08-23. An agent with only web access independently re-validated all ~14 load-bearing claims of the live public episode The Signal Ep 1, 34 days after air — full report + verdict table. Findings: 7 claims fully supported, 4 with paraphrase drift ("permit pause" aired as "banned"), 1 factually inverted, and zero citations on any public surface — the episode fails the proposed gate on process before facts. The self-demonstrating part: the one segment where the pre-air Council judge fired (Byron intervened, citations forced in) is the only segment that verifies cleanly at full precision a month later. The gate works where it fires; the platform makes it fire everywhere, and publishes the citations that already exist on disk into the feed. This test doubles as the platform's first public RAI proof point.

6. What gets reused (build honesty)

The honest read: plumbing proven, product new — which is why committed scope is §8's P1 only, and everything else is earned by usage data (Plausible decides, not enthusiasm).

7. Registry, governance, publicness

info.arnao.ai card with full enrichment (art + Veo motion + dossier), evolution preserved; SherpaDude front-door listing; per-station RSS + master; open-graph everywhere; sitemap free from the DOM list; command.arnao.ai tile on ship. Platform public; admin/cost owner-only; never-public list respected.

8. Phases — committed vs earned

• P1 (committed, honestly 2 sessions): accessible wall over existing DND + Loop episodes, jackets backfilled, per-station RSS done properly, provenance edges, tiered attestations, Disagreement Ledger + Spike Log retroactive on existing episodes. Accessibility launch gate included in the estimate — a screen-reader pass is real work, so the estimate says 2 sessions, not 1. Option A/B tuner decision executed as chosen.
• P2 (earned by 4 weeks of Plausible data): Mission Control + podcast-generate primitive, review-debt governor, judge tier, new shows.
• P3 (earned): council, i18n, auto-schedules, voting, RVC singing hooks.

Run-rate: auto-weekly show at Standard ≈ $3–6/month all-in; flagship Fable episode ≈ $4–6. Hosting static, ~$0.

9. Open decisions for Byron

1. Layout: Option A — the Lineup (streaming home; recommended) or Option B — instrument strip. (Byron already signaled A; confirm after seeing the Lineup mockup.)
2. Name — RESOLVED to SIGNAL (Byron 2026-08-23): platform = SIGNAL at https://signal.arnao.ai via the path-preserving takeover above; The Signal continues as the flagship station at /the-signal. fm.arnao.ai is released (or kept as a redirect). Remaining sub-question: does the network wordmark read "SIGNAL" alone or "SIGNAL — an arnao.ai broadcast"?
3. Google sign-in scope: resume+follows only (recommended) vs adding opt-in weekly digest email at launch.
4. Launch shows beyond The Signal + DND + The Loop (Field Notes/photography? Endurance? Harness Engineering Radio?).
5. DND 9-pilot ballot — does launch force the call?
6. Review-debt N (default 3) + monthly spend cap number.
7. Comedy adjacency: same Lineup (with per-station feeds) vs linked-but-separate. Fable flags same-domain satire as downside-heavy at the $1M level.

10. Links

11. What the UX review + user panel changed (v1 → v2)

Full-screen dial demoted (wall became the page) · accessibility promoted to launch gate · RSS promoted to first-class with per-station feeds and provenance-as-text · auto-approve hard-locked off for RAI/AI stations + human attestations · Disagreement Ledger added · cost switchboard → per-show policy + meter · Approve split from Publish per the cadence rule · adversarial spot-check · A/B listen-back · diff-from-last · commute-length tier · language chips replaced flags.

12. Self-critique — Fable adversarial pass (v2 → v3, integrated)

Fable 5 reviewed v2 as a ruthless critic. Its verdicts and what changed:

• "Scarcity signals authority; throughput signals hobbyist." The proposal was architecturally biased toward volume. → §1 rewritten: the governance apparatus, not the content engine, is the visible product; P2/P3 are now earned by usage data, not scheduled.
• "Kill the tuner — the wall of jackets with provenance edges is the brand object." → Sweep audio cut outright; tuner reduced to Option A/B on Byron's open-decision list (it was his founding image; the call is his, with the reviews' unanimous recommendation stated plainly).
• "The attestation is a personal warranty on a skim" — the most dangerous artifact in v2. → Attestations are now tiered and truthful (Listened in full / Claims verified / Summary reviewed).
• "The governor caps the wrong resource." → Replaced: generation pauses on pending-review depth (review-debt governor); the dashboard prices minutes-of-review per published minute; spend cap demoted to secondary.
• "Ledger buried in P3 is backwards." → Ledger + Spike Log moved to P1, retroactive on existing episodes.
• "P1 in 1 session is theater." → P1 honestly re-estimated at 2 sessions with the a11y gate inside it.
• "Economics table hides council/i18n in a footnote." → Table rebuilt with +Judge/+Council/+i18n columns; intro-pricing expiry flagged.
• "Vendor-diverse ≠ epistemically diverse." → Stated publicly as a known limit; human attestation is final authority, mechanical citation-existence check is the floor.
• Bold idea adopted: publish the rejects. → The Spike Log (§3) — the public kill-rate page — is now a signature feature.
• Not adopted: killing comedy adjacency outright — DND is Byron's creation and stays; the same-wall vs linked-separate call is his (Open Decision #7).

14. Round 7 — the appliance direction (v7)

Byron asked for an industrial/appliance exploration (car dash · refrigerator panel · receiver · SiriusXM), a fresh ideation round, and high-fidelity renders. Process: 3-user ideation workshop + 4-expert industry panel (automotive HMI, Rams/TE industrial design, streaming product, broadcast engineering) → synthesis → Fable adversarial pass → Nano Banana Pro renders. Full record: concept ledger · round-7 synthesis · Fable round-7 critique.

The palette correction (expert panel, 4-0, adopted): v6's neon gradient + perpetually-animating spectrum contradicted our own settled rules. The appliance direction wins: umber chassis, brushed aluminium, amber as lamp-light only, meter blue for verified. Motion is evidence, not decoration — nothing on screen moves unless real state changed.

What ships (Fable's resolution, adopted):

• Homepage = THE FACEPLATE: a cluster that never moves (nameplate + tally lamp · NOW PLAYING with chapter-tick scrub · transport · five fixed tell-tales: HUMAN/AUTO lamp, attestation 3-bar, citations n/n, judge lock, ledger) over a paging center stack of station strips whose episode cells are dense with per-claim verification state — density from governance metadata, never from inventory. Six printed preset keys that can hold stations, episodes, or standing intents ("15 min, both sides, drive home"). Hold any cell = the lid: the printed attestation sheet with live citations. Renders: phone · desktop.
• NOISE FLOOR = the Faceplate's first control (render): a squelch with printed detents OPEN → JUDGED → VERIFIED → ATTESTED. Per Fable: defaults per-station to that station's declared tier (The Signal rests at ATTESTED and loses nothing; auto stations rest at JUDGED and are honest about it); the full-catalog slide-to-dark is an explicit "prove it" demo gesture, not the resting state. SCAN plays each survivor's highest-uncertainty claim — adversarial self-audit as a listening mode.
• THE CARRIER as a live stream is CUT (concept render kept for the record). Salvaged: the Air Strip renders statically on the Faceplate (the broadcast day with published/scheduled blocks and killed slots as visible gaps), and kill-notices become produced, signed, dated 40-second artifacts in The Cutting Room and the RSS feed — witnessed by subscribers, legally reviewable as a template, no dead-air infrastructure.
• The instrument points outward (Fable's bold idea, adopted as P2 flagship): a gated demo where a visitor pastes a URL or transcript and the pipeline renders a claim table + floor reading for someone else's content. The moment the Faceplate stops being a trophy case and becomes a tool the buyer imagines pointed at their own organization.
• Build gates before visual build: (1) the Ep 1 citation-validation test proved per-claim validation automatable but citations unpublished — the mechanical publishing fixes land first, because the Faceplate broadcasts pipeline state on the homepage; (2) every appliance interaction (long-press bind, hold-for-lid, detents) gets a keyboard/AT-first equivalent designed before build or it dies at our own accessibility gate; (3) the phone spec must survive real arithmetic (~550px usable center stack) — claim micro-bars move into the lid at cell scale.
• Supporting features (user round): STANDBY (event-driven appliance idle — P2) · SECOND NEEDLE (narrowcast presence as a second needle on the scrub gauge; per-link opt-in, never public, never retained — P2) · TuneStart rocker (LIVE/START) · /listen pages ship P1 (say-this-to-your-speaker + feed URL + QR — the zero-UI distribution surface).
• The v6 Lineup is retired as homepage; its rail structure survives as the DOM-first crawlable layer beneath the Faceplate.

13. v3 → v4 — Byron's review of the mockups

Byron picked the wall over the instrument and asked for more: streaming-service DNA. What changed:

• Option A became the Lineup: hero billboard (governance in the hero), snap-scrolling rails (Continue Listening / New This Week / per-station / Because-you-listened), behavior-personalized rail ordering (the 2026 pattern — layout reflects usage, per Muzli's 2026 UI-patterns analysis and think.design's OTT dynamic-home analysis), chip filters in a bottom sheet on phones.
• The Cutting Room rail: Spike Log + Disagreement Ledger rendered as content cards — the governance layer becomes a browsable row no Netflix has. Rails render only at ≥3 real items, preserving Fable's scarcity thesis against streaming-scale expectations.
• Optional Google sign-in added as a registration primitive (resume + follows, nothing else, stated under the button; RSS and playback never require it). Reuses the clone.arnao.ai Google pattern and closes its open Client-ID item.
• Naming resolved by discovery: The Signal already exists at https://signal.arnao.ai as Byron's public weekly AI podcast — it joins as the flagship station rather than renaming the platform; collision with fleet Signal vocabulary avoided (Open Decision #2, veto welcome).