012 Multi-platform media extraction API

Synthesis

The council converges hard: as specified, this is a $500 learning bet, not a path to $8–10k/mo — and every seat that ranked it placed branch 013 above it. The 011 engine (2,747-test stack) is real, reusable capital; the business around it is truncated by one fact the seats reach independently.

Consensus

• The buyers who pay live on the two platforms you can't serve. Every paid comp in the brief — Apify's $45/mo scraper, the RapidAPI premium wrappers — sells Instagram or TikTok, an estimated 60–80% of extraction demand. Claude, codex, groq, mistral, openrouter and nvidia all note the willingness-to-pay evidence does not transfer to the 23 remaining platforms.
• The moat is invisible at the point of sale. "Reliability" can't be verified until after integration — which is why buyers tolerated bad wrappers. yt-dlp is free and covers all 23; a copycat clones DNS-pinning + GCRA + error taxonomy in a weekend (gemini, glm, groq, cloudflare).
• RapidAPI is the channel, not the enemy. With outreach refused and AdSense barred, its 25% cut is your entire CAC and your free demand-measurement surface — subscriber counts are public per listing (claude, codex, groq, glm, cohere).
• The real unit cost is maintenance labor, not servers — ~100–350 solo debugging-hours/year keeping 23 extractors alive (claude, codex, gemini).
• "Pay to remove ads" is muddled — API buyers never see the front-end ads (codex, nvidia).

Live disagreements

• Is $8–10k reachable? glm and cloudflare say yes — ~185–200 subscriptions is feasible if you dominate the "Reddit .json fix" niche. codex, mistral, groq, openrouter, claude underwrite only 20–60 customers → a $2–5k/mo business at best.
• The free ad-funded front end. Most call it a mismatched funnel — consumers pasting one URL never become $29/mo devs (claude) — and would skip it entirely (gemini, codex, groq operators). The founder wants it; the board doesn't.
• Outreach. groq insists closing one $500/mo contract from public job posts is unavoidable — a direct challenge to the no-outreach constraint.

Kill criteria

• A narrow RapidAPI listing yields 0 paying customers, or <100 authenticated calls from <10 accounts in 30 days (codex); or <5 organic key requests in 72h (groq).
• Public subscriber counts on YouTube/Reddit-only listings come back near-zero — selling to ghosts (claude, openrouter).
• Live maintenance exceeds ~5 hrs/week on a 30-day soak (codex).

Next actions

1. Read the market for free (2h): pull RapidAPI/Apify subscriber + review counts for YouTube/Reddit/Pinterest-only listings. This is the cheap niche-sizing the founder demanded — no outreach.
2. Ship one narrow listing: a Reddit (+YouTube) media resolver, honest unsupported-platform errors, $9/$29 tiers, public status page. Let 30-day subscriber velocity be the signal.
3. Automate a per-provider canary — hourly probe → public uptime badge — so "reliability" is observable and maintenance never depends on daily motivation.

The way through

Biggest objection: paying demand lives on IG/TikTok. Realistic route — stop selling breadth. Own the one segment where physics favors you inside the founder's constraints: Reddit killed .json and closed OAuth, stranding thousands, and its media resolves with an ordinary client — no forging, no proxies. That's the single wedge no incumbent owns and where "the reliable one" is genuinely needed. It won't hit $8k alone; if the RapidAPI counts come back thin, the one pivot worth testing before walking away is folding this verified engine into branch 013 as infrastructure — selling extraction as a feature inside a creator/agency outcome, where the buyer never compares you to yt-dlp.

Dissent

By: codex

• The revenue math is broken. At $29/month, 200 subscribers produce $5,800 gross and $4,350 after RapidAPI’s 25% cut—not $8–10k. Even the “20–60 customers → $2–5k/mo” estimate requires an unstated $45–$333 ARPU. The council has not modeled plan mix, churn, or usage costs.

• “Independent convergence” may be correlated groupthink. Ten models reasoning from the same brief and comparables are not ten independent market observations. Ranking branch 013 higher says little without normalized evidence, capital requirements, and probability-weighted returns.

• The 60–80% demand claim is load-bearing but apparently estimated, not measured. Paid Instagram/TikTok competitors prove demand there; they do not establish what fraction of addressable demand remains elsewhere.

• “Reddit killed .json and closed OAuth” is dangerously broad and possibly false as stated. The entire wedge rests on exactly what broke, for whom, and whether Reddit tolerates commercial resolution traffic. “Ordinary client—no proxies” is an assumption, not durable platform physics.

• The experiment cannot distinguish weak demand from zero distribution. A new RapidAPI listing getting no customers in 30 days may test ranking, naming, trust, or marketplace visibility—not the niche. Five key requests in 72 hours is especially arbitrary.

• Public subscriber counts may be stale, bucketed, gamed, or incomparable across listing ages. They are directional evidence, not a free market-size read.

• “Cloneable in a weekend” confuses code with service operations. DNS pinning and rate limiting are copyable; accumulated failure fixtures, provider-change detection, incident history, and reputation are not. Conversely, 2,747 tests are not capital unless they predict live failures.

• Hourly canaries do not make reliability observable. They test chosen URLs and paths, can miss regional/authenticated failures, and may themselves trigger blocking.

• The proposed Reddit+YouTube bundle muddies the supposedly narrow wedge. Test Reddit alone first; otherwise results cannot identify which demand exists.

• Survives: Maintenance labor is probably the dominant cost, and the ad-funded consumer funnel is poorly matched to developer subscriptions.

Decision

KILL as a developer API. Competes with free yt-dlp and excludes the two platforms with the wallet (Instagram/TikTok). Engine frozen, not deleted.

Seat health

Who answered and who did not. Kept at the bottom on purpose: it is diagnostics, not findings.

Seats present: claude, codex, mistral, gemini, groq, nvidia, glm, openrouter, cohere, cloudflare No seats missing or failed.