From Recipe Semantics to Human-Verified Sensory Prediction

Industrial flavor development asked in 2021–2022 whether a sensory profile could be previewed without a panel. We present an evidence architecture and a deployed instrument, MP6 Recorder (Methods frozen at v0.2.1; production delta v0.2.2.2 documented in §6.9), that operationalises panel-less sensory prediction as a falsifiable, prospective workflow rather than a claim: a canonical recipe document (JRF) is hashed and consumed read-only; a prediction engine — a fully documented deterministic baseline, or a model-backed engine behind a provenance-preserving adapter — generates an Expected MP6 (MP6-SIM) explicitly labelled as simulation; a human reviews every machine proposition (accept / edit / reject / skip, with reason codes and decision latency); the reviewed prediction is immutably frozen (SHA-256) before tasting; a physical Run captures instrumented observations with method provenance; the human then records the actual experience (voice / text / manual values), producing an MP6-OBS that contains only human-accepted evidence; SIM and OBS are compared strictly over a comparable-attribute mask with coverage κ and per-attribute residuals; and the entire chain — including rejected propositions, skips and negative results — persists as an append-only observability ledger (ABED) projectable into the Sensory Experience Graph (SEG) as draft assertions only.