A research demonstration of an uncertainty certificate — a signed record of a measurement, its confidence bound, its assumptions and its validity window. It is not a compliance, conformity, or security certificate, and this page does not offer conformity-assessment or certification services.
Certificate: executed two-stage agentic pipeline, chain interval , point . Verdict rule: ALLOW if the certificate's lower confidence bound clears the policy threshold, DENY if its whole interval falls short, ESCALATE when the threshold lands inside the interval.
The threshold only matters for the valid certificate. The seven mutation scenarios return DENY at every threshold, by design — no policy setting rescues a tampered, expired, or mis-bound certificate.
Policy: minimum lower confidence bound
A composed score multiplies two link scores. The paired bootstrap resamples both links from the same draws, so their joint wobble is measured, not assumed. Below: the actual 2,000 paired replicates of the German-credit chain — link A against link B. The tilt of the cloud IS the cross-link correlation.
fixed policy threshold →
Theory (evaluated live, client-side — the page's only live computation): with average cross-correlation ρ̄ over peff equally exposed components, the honest variance is R² = 1 + (peff − 1)·ρ̄ times the independence-shortcut variance; the shortcut understates interval width by 1 − 1/R.
ρ̄ peff → understatement
Measured systems for comparison: German chain ρ = 0.712 → 23.6% (on the peff = 2 curve exactly); executed chain ρ = 0.319 → 11.0% measured (its links are unequally exposed, so its effective dimension sits below 2); volume composites and TOPSIS → -7.6% to +1.4% (logistic model alone: -3.1 to +1.4%) across models and variants, with sign flips — the shortcut can also run wide.
An operator clears a reading only if the whole interval lies above the threshold, rejects only if it lies below, and is undecided otherwise. The two readings of the same chain — measured covariance vs cross-term declared zero — disagree on a set of thresholds. The curve (the paper's own figure, recomputed here from the stored widths) shows the fraction of a threshold window on which they disagree.
Known-truth check: nine simulation scenarios, nominal 95%, 1,000 outer replications each (Monte-Carlo s.e. ≈ 0.7 pp). The paired percentile interval holds 94.1–97.1%. The independence-shortcut interval wanders 80.2–99.8% — sometimes anticonservative, sometimes wastefully wide. Wider is not the point; correct is.
Two fixed-design recorded runs of the deposited chain estimator — n chosen in advance in both, not sequentially. Same policy threshold, 0.750, both times.
Each row names the mutation scenario in section 1 that corrupting the field triggers. Every one of those scenarios is DENY at every threshold.
console-v2-data.json, generated by
build_console_v2.py); this page performs no cryptographic
verification — the only live computation is the closed-form interval
arithmetic above, evaluated client-side. The volume-integration and TOPSIS
composites are the constructions of Giudici and Kolesnikov (Machine
Learning with Applications, 2025; Kolesnikov, MSc thesis, Univ. of Pavia);
this work adds the confidence regions around them. Chain-certificate
values as published (DOI 10.5281/zenodo.21547845). The flip-rate,
coverage, evidence-collection and composite panels show results from a
draft in preparation with the University of Pavia. Research demonstrator,
not an operational trust configuration.