Discovery workbench
Lontra
Nothing gathered yet. A signal is a person describing a problem in their own words, with a link and a date.
Ask a source anything in the chat, bottom right. It reads what you want out of a sentence, shows you what it understood, and asks nothing until you press the button.
Nothing has checked whether the capability register still matches the documents it was written from. Everything a campaign may claim rests on it.
Look at the register before scanning, or carry on — nothing here is blocked.
Questions put to a source
Each one ran the same search and the same scout a scan would, and wrote nothing into the store. Only what you tick becomes evidence — the store is what every report reads, so a question you were merely curious about must not end up in one.
Nothing has been asked yet
Open Ask a source in the bottom right and say what you want to know, in your own words. Nothing it finds is kept until you say so.
Sources
A source is a place to go and ask a question, not a databank. Each has its own scout, because the candidates are not the same thing — Reddit hands over a pseudonymous comment whose identity is the subreddit, X a post plus the author's bio, a job board an organisation describing a problem it has funded.
Posts and comments — pseudonymous, with the subreddit as the only identity, plus search hits nobody opened.
Google for breadth, Reddit's own RSS for depth: one thread read per ~60s window, so a read budget is a time budget (ADR-0041).
Posts of up to 280 characters. Read through Grok on OpenRouter (#149): a synthesis of the market with citations, of which only a cited permalink carrying a quotation, dated from its own post id and inside the window, becomes a signal. Everything else is context and may never be cited as evidence.
Grok through OpenRouter native X search. Only a dated permalink with a quote matched to returned source text becomes evidence; other findings remain context.
Video descriptions written by the channel as marketing copy, and comments written by viewers. Almost opposites.
The YouTube Data API as the project's Google identity — no separate key, the same ADC path as Vertex.
Talk abstracts with a speaker and their organisation — written to attract an audience, so they overstate.
Programme pages fetched directly. Finding them with --find costs a search per topic.
Comments where people describe what happened at their own company — pseudonymous, but a population that builds and runs software.
Algolia's public search API. Free, unauthenticated, and the only source here that costs nothing to run.
People describing what is wrong with the tool they already bought, usually with their role and company size.
Google, via DataForSEO — two query shapes a topic at about $0.0175 each, then the pages they name are fetched directly.
A problem someone has put budget behind — the role, the organisation hiring, and what they say it has to fix.
Google, via DataForSEO — two query shapes a topic at about $0.0175 each, then the pages they name are fetched directly.
Practitioners quoted by name inside an industry article. The journalist's own narration is not kept.
Google, via DataForSEO — two query shapes a topic at about $0.0175 each, then the pages they name are fetched directly.
A named analyst quoted on the market, or a practitioner quoted inside an analyst piece.
Google, via DataForSEO — two query shapes a topic at about $0.0175 each, then the pages they name are fetched directly.
Public forum threads — Discourse and Stack Exchange communities, trade boards — where people post to solve a problem they have now.
Google, via DataForSEO — two query shapes a topic at about $0.0175 each, then the pages they name are fetched directly.
Public posts and Pulse articles, where the speaker's real name and job title are normally attached.
Google, via DataForSEO — two query shapes a topic at about $0.0175 each, then the pages they name are fetched directly.
Any page where someone describes the problem in their own words — blogs, forums, write-ups.
Google, via DataForSEO — two query shapes a topic at about $0.0175 each, then the pages they name are fetched directly.
Channels — where the audience gathers
Derived from where the signals were said, and kept by hand where the signals have not reached. A community that forbids promotion stays listed: it is still where the audience listens. A size or an activity figure is a measurement with a date, or absent.
No channel has been mapped yet
Mapping reads the signals already in the store and groups them by where they were said, so it needs signals first. With none, open a source above and add the one you already know about by hand.
A source you want that is not here
Describe it in a sentence — what is said there, and by whom. What comes back is the registry row it would take, whether one of the twelve sources or seven lenses already covers it, and what somebody would still have to write.
Worth asking
- “Somewhere HR software buyers compare products and say what went wrong after they bought.” — Names a population and what they say. Most likely already the `review` lens — which is the answer worth getting before writing anything.
- “Public Discourse and Stack Exchange communities where practitioners ask each other how to handle something.” — A shape of place rather than a named site, which is what a lens is.
- “Regulatory filings and consultation responses where organisations describe a problem they have.” — A source nothing here reads. Dated, quotable, institutional — the test of whether this really is new ground.
- “Podcast transcripts where practitioners describe their own work.” — Tests the date rule hard: a transcript is quotable, and whether an episode carries a readable date decides whether it can produce a signal at all.
- “Our competitors' changelogs and release notes.” — Not a market source at all — it is competitor intelligence, which is a different module. A good proposal says so.