Questions a buyer actually has

It hears a US president name a public company on live TV or radio, checks the audio is live and not a rerun, and timestamps it. It then records how many independent feeds heard the same thing and what the company's stock did against a benchmark over the following minutes.

22 names are tracked across 9 feeds. You can watch the raw transcript arrive on the console.

No. It is an observation feed. There are zero confirmed live outcomes so far. It reports what was said and what the price did, and it makes no claim that one predicts the other.

There is no win rate here, no backtest result, no edge claim and no profit claim. 3 events are on file and 2 of them are reconstructed from archived broadcast audio, from the same speech, the same stock and the same day. That is not a sample you can conclude anything from, and the publisher's own event study refuses to try.

Detection is seconds from the utterance. The detector itself is not the slow part: matching a tracked name and reading the surrounding words takes 0.17ms at the median, and there is no cloud model in the hot path.

The time goes into the feed. A broadcast source hands you several seconds of already-buffered audio the moment you connect, and that buffer depth dominates everything else. Continuous radio buffers 2 to 5 times shallower than segmented video. The measured numbers per source are on the feed rail, in the BUF column.

One honest limit: buffer depth is a prior for lag, not lag. Standing latency per feed needs a simulcast to measure and we do not have that measurement, so we do not publish a number for it.

Acoustic fingerprinting against everything already heard, plus cross-feed simultaneity. A soundbite re-aired during a live event is still caught, because the audio itself matches something already in the store.

A replay defeats every other check. The gazetteer matches, the valence is right, the negation test passes, because the president really did say it. So the default verdict is REPLAY and live has to be earned.

Two or more independent feeds reported the same tracked company inside a lag-corrected window. Detections are compared at utterance time, not arrival time, so a slow feed does not break the match.

One feed is not evidence. Two independent decoders do not mishear the same rare proper noun at the same instant, which is exactly why the threshold is two.

Recency, and nothing else. Every tier returns the identical schema, the identical fields and the identical verification block. A cheaper tier is older, never thinner. The cutoff is applied server side against the event timestamp and is a property of the route, never of a query parameter.

Responses also tell you what you cannot see. The field withheld_newer_than_tier counts events that exist and are newer than your rung. It is a count, never a timestamp, because knowing that something happened at a precise second is the product.

x402. USDC on Base mainnet. No account, no API key, no signup.

Call a priced route with no payment header. You get HTTP 402 with the amount, the asset, the recipient and the network. Sign the authorization with your own wallet tooling and retry the same request with the payload in the PAYMENT-SIGNATURE header.

If you would rather buy once and poll freely, take a prepaid pass: $50 for 24 hours of the live tier, $750 for 30 days of it, $100 for 30 days of the archive. A pass returns a signed token you send as X-SYGNL47-TOKEN. It buys unmetered access to its own tier, never fresher data than that tier, and an expired pass falls back to a normal 402 rather than to a free serve.

No. A tier whose window contains no events returns HTTP 404, and a 404 is not settled. An agent polling a quiet afternoon pays nothing.

The same is true when our publisher sync is missing or unreadable. That returns 503, also unsettled. We would rather fail loudly than serve you yesterday's artifact as though it were current, so every successful response also carries sync.feed_age_seconds, the measured staleness of the artifact it came from.

22 names, with their aliases and their mapped instruments. The count is on the console tiles. The full list with current guard state is sold at GET /v1/potus/entities and is also in the free public JSON at /api/public/signals.

Being tracked and being actionable today are different facts. A name is blocked when it has an earnings print inside the holding window, because the print dominates the move and wrecks attribution, and blocked when it fails the $10M average dollar volume floor. A guard that could not run says UNKNOWN. It never silently passes.

Some entries are watch-only. A country or an index only counts when a policy instrument is nearby, which is the rule that correctly held a mention of a foreign head of state rather than firing on it.

Yes. There is an MCP endpoint at https://47.sygnliq.com/mcp/. Paste the config from /mcp.json into Claude Desktop and the sygnl47_* tools appear. Start with sygnl47_manifest, which is free and states the inventory honestly.

For any other agent there is /skill.md, a drop-in skill file that teaches the field meanings, the payment flow and the things the agent must not conclude, and /agent.py, a stdlib-only starter that polls a tier and handles the 402.

The starter agent supports Alpaca. It defaults to paper and stays there unless you set two separate environment variables on purpose, one to ask for live and one to confirm you understand it trades real money. One stray variable is not enough. It sends limit orders only, with a per-fire notional cap and a hard daily cap, and it refuses to act on reconstructed events or on any event whose liveness check did not come back live.

Read this part plainly. Nothing about this signal is evidenced. There are zero confirmed live outcomes, the sample on file is two reconstructions from a single speech, and nobody has measured what you would actually get filled at. If you point this at real money you are trading an unevidenced hypothesis with your own capital, and the outcome is entirely your risk. The honest use is to observe, log, and build your own evidence first.

3 events. 2 reconstructed from archived broadcast audio and labelled source: replay. 1 confirmed live.

That is the inventory, stated plainly, and it is also published in the free manifest at GET /v1/potus so you can check it without paying. A thin feed is a thin feed. Selling it as anything else would be the fastest way to lose the only buyers worth having.