@buildtovalue/replay

Headless conformance checking for BPMN replay — "does reality behave the way the model says?". Streaming XES 2.0 / CSV parsing, one-pass frequency/time pre-aggregation, token-replay fitness (never alignments), deviation detection and variant extraction.

Imports nothing from the ecosystem — not even @buildtovalue/core. It replays over an abstract graph injected by the caller, so the same machinery works on a BPMN model, a future DMN DRD, or a fake graph. Zero dependencies.

import { parseXes, aggregate } from '@buildtovalue/replay';

const traces = parseXes(xesText); // or parseCsv(csvText, { case, activity, timestamp })
const graph = {
  nodes: [{ id: 'brief', name: 'Coletar briefing' }, /* … */],
  edges: [{ id: 'e1', source: 'brief', target: 'gate' }, /* … */],
};

const log = aggregate(graph, traces);
log.fitness;          // { fitness, conformingCases, totalCases, … }
log.edges;            // per-edge frequency + avg time (heatmap thickness)
log.nodes;            // per-node frequency + avg sojourn (⌀ chip)
log.bottleneckNodeId; // slowest node (GARGALO)
log.deviations;       // transitions with no model edge (▲ DESVIO · N casos)
log.variants;         // top sequences by share (▶ Reproduzir)
log.unmapped;         // log activities with no matching node

Design constraints (cercas §0)

• Token-replay fitness only — a transition move with no corresponding model edge is a deviation; fitness = fit moves / total moves. Optimal alignments (A* over model×log) are process-mining literature and stay out (recorded in pendencias.md).
• One pass, no DOM — aggregate computes every frequency/time in a single O(n) sweep; a 100k-event log aggregates in well under 2s (tests/benchmark.test.ts). Rendering (Handoff 7B-2) animates only sampled variant traces, never one token per event.
• Streaming parse — XES scans trace-by-trace and CSV reads line-by-line into compact tuples; the full log is never materialized as a rich object graph.

Matching

Events map to nodes by normalized activity name (normalizeName: trim, lowercase, collapse whitespace) against each node's name (falling back to id). Unmatched activities are reported in unmapped and every transition touching them counts as a deviation.

Decoupling

replay imports only relative paths — pinned by tests/independence.test.ts — and the acid test (tests/acidez.test.ts) drives it with a coffee-brewing graph to prove it knows nothing about BPMN. Integrations with registry/audit (bindRun filtering, attaching analysis to a promotion) happen in the host by injection, never here.