How RecoveryAssure decides whether recovery is feasible
RecoveryAssure separates deterministic engine truth from AI interpretation. Every finding is code-defined, every derived value traces back to source data or a named rule, and no LLM is allowed to change severities, RTOs, MTPDs, MBCOs, or triggered rules.
The pipeline, end to end
- 1. Source data
Business Impact Analyses, application catalogues, supplier registers, and DR test records enter the system as XLSX or CSV files. Nothing is executed — we open workbooks in read-only, data-only mode. Macros, embedded objects, and OLE content are rejected before parsing.
- 2. Deterministic normalisation
Numbers, durations, percentages, and enums are normalised by explicit rules. Ambiguous phrases like "1 business day" or "EOD" are surfaced as errors rather than silently interpreted. Every canonical value carries a Provenance record tying it back to the source file, sheet, row, and column.
- 3. Rule engine
Twenty deterministic rules evaluate the canonical dataset — each rule pins its version, its config, and the facts it used. Findings are code-defined; LLMs do not decide feasibility, severities, or triggered rules.
- 4. Recovery feasibility
For each process we evaluate nine dimensions — TIME, TECHNOLOGY, DATA, CAPACITY, PEOPLE, SUPPLIERS, LOCATION, DEPENDENCIES, VALIDATION — and derive an overall status: FEASIBLE, AT_RISK, NOT_FEASIBLE, or INSUFFICIENT_DATA. Confidence in the conclusion is a separate axis and is downgraded when data is stale, tests are missing, or duplicate patterns weaken evidence.
- 5. AI assistance · Available today
AI assistance is on-demand and interpretive only. The deterministic rule engine remains authoritative for every finding, feasibility status, severity, calculation, and any RTO / RPO / MTPD / MBCO value. The LLM cannot determine or change any of those; it is never on the code path that derives them.
Every AI output is labelled as AI-generated (INFERRED / RECOMMENDED) and requires professional review before being acted on. Quantitative claims returned by the model are guardrail-validated against the deterministic context — if a fact contradicts the underlying SOURCE FACT or CALCULATED value, the AI narrative is withheld and the user sees the deterministic finding on its own rather than a fabricated one. When the AI provider is unavailable (LLM_NOT_CONFIGURED / LLM_PROVIDER_UNAVAILABLE), the deterministic analysis remains fully usable — engine output, findings, feasibility, and PDF reports are unaffected.
- 6. Human review
Every finding, every dimension result, every derivation trace is retained and exportable. Historical runs are immutable — a run's rule library snapshot, mapping profile snapshot, and dataset are pinned by content hash so the same inputs always produce the same output.