Lodestake / How it works
How a Lodestake report is made.
A report is a chain of custody: questions, evidence, a draft, a critic, a revision and a verifier. This page shows each link, including the rule that removes unsupported numbers.
01Plan
Plan the questions
The idea is turned into a short list of research questions: who has the problem, what they use today, what it costs them, what the alternatives charge. Each question says what kind of source would answer it.
02Evidence
Fetch and label the evidence
Questions are answered from pages and data sources the pipeline fetches itself: web pages, public datasets and APIs. Every item gets an ID (E1, E2 …), a link and a method label. In our benchmark every one of the 2,380 evidence items was fetched, not written by a model.
03Draft
Draft with citations
A first report is drafted. Each claim carries the IDs of the evidence behind it, so the citation travels with the sentence.
04Critic
Argue against it
A second model reads the draft against the evidence as a sceptical investor would: claims the evidence does not support, wrong citations, filler, missing risks, weak plans. It returns up to eight concrete fixes and a verdict, and the draft is revised. In the benchmark the critic sent back 27 of 30 drafts; the other 3 skipped the pass.
05Verify
Remove what the sources do not carry
The verifier reads every sentence. If a sentence states a number, it stays only when that number appears in measured evidence the sentence cites. Otherwise the sentence is removed and counted. Kept citations become source links. Across the benchmark it removed 79 sentences: 0 to 6 per report.
06Unknowns
Write down what is not known
Open questions get their own section: market size that sources disagree on, missing pricing evidence, competitors nobody checked. Public data that could not be matched is shown as unknown, never as zero. The benchmark reports listed a median of 7 unknowns each.
Try the verifier
The rule, on your own sentence.
This box runs the same rule the pipeline uses, on illustrative evidence. Edit the sentence, tick or untick evidence, and watch the verdict. Notice what happens when you cite E3: a model-written summary can never support a number, however confident it sounds.
Two things the rule does not do. It does not judge whether prose is wise; that is the critic’s job. And it does not make a sentence true by matching a number; it only proves the number came from somewhere you can open.
Rule, as the pipeline runs it: a sentence that states a number stays only if every number appears in measured evidence it cites. Years are not treated as claims. Try changing 41 to 60.
The edge of the report
Where the report stops.
A report stops where testing stops. It does not tell you whether to build, whether a market will grow, or whether you can execute. Those are yours.
Inside the report
- What sources say about the problem and the people who have it
- What alternatives exist and what they charge, where published
- Reasons to build and reasons not to, with evidence
- What could not be established
Outside it
- Your judgement
- Market timing
- Execution
Your idea, your data
What happens to your idea
Ideas you save live in your library. Plans and reports attach to the idea they were made from. You can export a venture packet and hand it on. How long things are kept and how to have them deleted is set out in data handling, which is a draft pending review.
Why we show the failures
A report that hides its removed sentences looks cleaner and is less useful. The count of what the verifier removed, and the list of unknowns, tell you where the ground is soft. That is usually where the next week of work should go.
Next
Stake a claim on an idea you already have.
Ideas, plans, names and domains work in the app today. Evidence reports are in pilot on our side. Questions: [email protected]