REF.03 own project · skillsintelligence.tools · 2026
A concept graph in which source agreement turned out to be a prompt echo
A map of the skills-intelligence domain — what differs from what, and what has to exist before what — with four evidence tiers and a separate register of claims the research prompt itself planted in the sources.
- T1–T2Declared · InferredInitial corpus sweep
- T3–T4Validated · Attested7 of 8 claims corrected
- ECHOPrompt-echo register48 concepts · 13 relations isolated
- Concepts / relations
- 191 / 322
- Prompt-seeded concepts
- 48
- Claims corrected in verification
- 7 of 8
- Sources
- 51
-
Problem
Eight deep-research reports from two systems agreed on the same maturity ladder. It looked like strong consensus — until it turned out the ladder had been pasted into the prompt that commissioned the reports. The sources' agreement carried no evidential weight at all.
-
Approach
Every concept and relation received an evidence tier: declared, inferred, validated, attested. Claims seeded by the prompt went into a separate register instead of entering the graph as facts. The project's own source sweep went through adversarial verification, and the data is built by a deterministic pipeline with ten invariants, followed by manual curation.
-
Outcome
191 concepts and 322 relations published on a public site, with 48 concepts and 13 relations explicitly marked as prompt-seeded. Adversarial verification forced corrections to 7 of 8 claims in the project's own sweep — before they reached the graph, not after publication.
This project shows best why agreement across many sources is not the same as confirmation. If every source got the same question with the same suggested answer, their agreement measures obedience to the question, not the state of the field.
The register of prompt-seeded concepts does not remove them from the graph — it lets the reader see which parts of the map stand on independent evidence and which stand on an echo.
What this engagement does not cover
Eight items have completed the highest evidence tier (adversarial verification against primary sources) so far. The remaining extracted concepts and relations are deliberately held at lower evidence tiers or in the waitlist register until independently corroborated.