Claims audit

Corrected by the founder. emealia’s pipeline has two paths, and this document set described only the first:

• Food is in the reference database → values come from curated official sources; the model supplied only the parse. This is the path every page described.
• Food is not in the database → Call B asks the model for per-100 g min/max ranges for every nutrient. The result is shown to the user marked as an AI estimate, written to a quarantine table rather than into reference data, and queued for a human to research and promote. Verified in release/epic-29: prompts/estimate_ingredient.md instructs the model to “return your honest best range for every nutrient”, and core/ingredient_candidate.py describes itself as “AI-supplied ingredient composition awaiting human research… never trusted as reference data”.

How it happened: the earlier, longer version of the tech page documented Call B correctly. Condensing that page into a three-row pipeline table dropped the row — and the surrounding prose, which had been accurate as a description of the DB-hit path, silently became a claim about the whole product. The error then propagated to directions, positioning, and to the live marketing site in both languages (“Zahlen aus offiziellen Quellen, nicht von einer KI geraten” / “not guessed by AI”), where it was a public overclaim for roughly two days.

The corrected claim is weaker and still distinctive: the Warentest failure mode is not eliminated, it is bounded and disclosed — the estimate is labelled at the point of use, held out of reference data, and fixed by a human afterwards. Marking per-item provenance is a real and unusual behaviour; claiming the model never estimates was simply false.

Process note: condensing is not a neutral operation. Removing a qualifying row from a table can turn an accurate sentence into a false one without touching that sentence. Any future condensing pass should re-check the claims that survive against what was cut.

The founder identified a pattern this audit had not been tracking: stating confidence where none was earned, and constructing rationales for decisions whose actual reasons were never given. Examples found and fixed in this pass:

• “No competitor in this category can make both claims in one sentence” (tech) — an unchecked negative-existence claim, the exact shape that has failed repeatedly above.
• “The MDR/wellness boundary is, for this narrower audience, close to a non-issue” (practice) — a legal determination asserted by a party with no standing to make one. Now marked as reasoning.
• “That is precisely why this segment, not the clinical one, is the one worth building for first” (practice) — a rationale invented for a scoping decision that arrived as an instruction, with no reason attached.
• “That is not a failure of the brainstorm — it is evidence that axis 1 is load-bearing” (positioning) — one interpretation presented as the finding. Two readings fit that evidence; both are now stated.
• “The medical axis is removed” framed as a settled strategic outcome across several pages — it was a scope instruction, and the subsequent reinstatement of three of its groups demonstrates the difference.

The standing rule from this: state facts, research results, and what happened in this collaboration. Where something is inference, label it. Where a reason was never given, do not supply one.