Laboratory with a provoked failure

How it works

The exercise starts from a seed: the seeded pseudo-random generator (mulberry32), from the engine shared with the other demos, picks the fictional firm, the activity, the zone, the areas, the sum insured and the dates, and a separate sub-stream injects the failure of the chosen scenario. The scenarios form a domain pack for property insurance: a sum in words that differs from the figure, a fire suppression inspection certificate that expired before the policy starts or expires during it, and an area measured at the survey that differs from the declared one, either materially or by at most 0.5%, the size of a measurement rounding. The pipeline has four visible stages (read, extract fields, check rules, decision), and extraction is a deterministic reading of the document's structure, which is known because the page generated the document. A rule with no defined action stops the pipeline; the stopped stage is marked with text and colour, and the rule and the conflicting values appear only after the first diagnosis, so the learner looks for them in the document first. Diagnosis and fix options appear in an order drawn from the seed, and the fix (an escalation rule or a numeric threshold) is evaluated on the next run against the scenario key: a material discrepancy (at least 10% on the sum, 6% on the area, 15 days since the certificate expired) must be escalated to an underwriter, and a tolerable one (area rounding, a certificate expiring during the policy) must pass automatically, with a note. A threshold that is too high lets a material discrepancy through, and an escalation rule sends a tolerable one to an underwriter for nothing; the page marks both outcomes as incorrect. The log counts steps instead of recording the time, so the same seed and the same actions produce the same JSON file, byte for byte.