Governed Inverse-PINN Console
Discover an unknown physical parameter of an ODE from data + physics — then watch the governance work. Every discovered value is MODELED (a fit to data, never MEASURED), carries a 95% confidence interval, a GREEN/YELLOW/RED convergence badge, a Fisher-information self-doubt gate that refuses non-identifiable parameters, a Λ advisory, an F19/Bekenstein plausibility check, and a signed Khipu receipt you can re-verify.
Duffing oscillator m·x″ + c·x′ + δ·x + α·x³ = F·cos(ωt) — α is unknownRun Identify a parameter
Two one-click cases hit the LIVE engine at POST /api/a11oy/v1/pinn/identify.
The honesty story tells itself: one converges and asserts; one is non-identifiable and the engine refuses.
1 Convergence
The 3-state governance verdict from the causal-weight / gradient / FIM-conditioning criteria.
2 Discovered parameter(s)
Each value is MODELED with a 95% CI. A parameter the self-doubt gate cannot support is REFUSED verbatim — no value invented.
"We don't report what we can't support."
3 Identifiability — the self-doubt gate
Fisher Information Matrix conditioning decides whether the data actually constrain the parameter. κ(FIM) → ∞ or Fisher below floor ⇒ UNIDENTIFIABLE ⇒ the engine refuses.
4 Λ advisory
A single advisory trust score. Conjecture 1 — advisory, NOT a theorem — capped ≤ 0.99, never a proof, never 1.0.
5 F19 / Bekenstein check
The Bekenstein bound (F19) is a locked-proven inequality — here it is applied, not re-claimed. It checks the discovery doesn't encode more information than the system can physically hold.
6 Signed receipt + ledger
The discovery is sealed as a DSSE-enveloped Khipu receipt in the unified ledger organ
a11oy-pinn. Signature state is shown exactly as returned — never faked.
Doctrine Honest labels
What is proven, what is wired, what is advisory. The half-state — claiming more than is real — is the only unacceptable outcome.