You don't. Process mining starts from an event log: no case-linked timestamped events, no discovery. Mine the artefact you have instead — parse the Visio files, BPMN exports and written procedures into graphs, check each model for soundness, then compare models against each other. Where three departments document one process three ways, the contradictions are the finding.
Ross Jones — Founder, The Hopium Lab. Last modified 21 July 2026.
Twenty-four shipped products across fifteen sectors. I wrote the parser for this problem because I needed it on my own documentation.
What does process discovery actually require?
An event log. The field's founding document — the IEEE Task Force on Process Mining's Process Mining Manifesto (2011) — opens its definition of the field with "Starting point for process mining is an event log". Every technique assumes events recorded in sequence, each tied to an activity and a case.
The manifesto grades logs on a five-star scale. At the bottom sit events recorded by hand — some wrong, some missing; the examples are paper trails routed through an organisation and paper-based medical records. Discovery is not worth attempting there. The founders of the field drew that line, not me. What survives is the rest of Stage 1: the manifesto's L* life-cycle lists handmade models as a first-class output of its first stage. Documented models are an input to the canonical methodology.
Is "no event logs" ever literally true?
Usually not, and the correction is worth making before anyone spends money. Guiding principle 1 of the manifesto is explicit: events need not live in dedicated log files — database tables, message logs, mail archives and transaction logs all hold event data. Most organisations claiming no event logs have event data nobody extracted.
The genuine case is narrow and common: work executed across email, phone calls, paper forms, spreadsheets and human judgement, where no system anywhere holds a case-linked timestamped trace. Approvals sitting in inboxes with no case identifier, no activity label, no thread linking one to the next. Handovers in corridors. Discovery tooling has nothing to read.
Process mining cannot run without event data. Not a tooling gap — a definition. A vendor demo always begins with a log the vendor already has.
How do you get a graph out of Visio files, PDFs and written procedures?
Parse them, because the formats are open and boring. A .vsdx is a ZIP of XML built on the Open Packaging Conventions (ISO/IEC 29500-2), and Microsoft's file format specification defines its pages, masters, shapes, connectors and text. A canvas description, not process semantics: nothing in it resembles BPMN 2.0.2 (OMG formal/2013-12-09, ISO/IEC 19510:2013). Extraction is table stakes: free browser converters emit BPMN 2.0 XML, and iGrafx and Nintex shipped Visio import years ago. Diagram import is solved, not innovative.
Prose is older than it looks. Friedrich, Mendling and Puhlmann generated BPMN from natural-language text at CAiSE 2011, producing 77% of models correctly across 47 text–model pairs. Language models did not invent the technique; they removed its dependence on tidy prose and named roles. Hand-drawn diagrams remain hard: Arrow R-CNN raised fully recognised scanned flowcharts from 37.7% to 78.6% (Schäfer et al., IJDAR 24(1), 2021), nowhere near a clean import.
I built FLOW for the parsing; the audit on top is a fixed-scope first look.
What does diffing a documentation library find that reading the diagrams cannot?
Disagreements no single reader holds in their head at once. A corpus diff is a machine-readable comparison of every documented version of a process, run across an entire diagram library, that reports where the documentation contradicts itself.
Five classes of finding, none needing a single event. Two are decidable from the models alone; three need only a directory or an inventory to join against:
- One process, three versions. Three departments differing on step count, ordering, and who owns the exception path.
- Orphaned decision branches, which break soundness in a formally named way. A workflow net is sound only if it can always still reach completion (option to complete), terminates with nothing left behind (proper completion), and contains no transition that can never fire (no dead transitions). A branch leading nowhere kills option to complete; a branch nothing can select is a dead transition (van der Aalst, 1997; Formal Aspects of Computing, 2011).
- Steps referencing decommissioned systems, surfaced by joining node text to the application inventory.
- Swimlanes naming roles that no longer exist, surfaced by joining lane labels to the current directory.
- Terminology collisions. One label meaning two things; two labels meaning one step. "Approve", "Authorise" and "Sign-off" are usually one activity in three vocabularies.
Per-model, none of these five checks is new: similarity metrics were benchmarked in the Process Model Matching Contest 2013. Running them across a whole departmental library, and joining role and system names to live inventories, is the part nobody ships.
Why is a corpus diff not conformance checking?
Because conformance checking replays an event log against a model, and the comparison here has no log at all. The distinction between discovery, conformance and matching surfaces in the first review meeting containing a BPM academic.
| Technique | Needs | Establishes | Cannot establish |
|---|---|---|---|
| Process discovery | An event log, three stars or better | Behaviour as executed | Work no system touches |
| Conformance checking | A log and a model | Where execution deviated | Anything without the log |
| Soundness verification | One model | Dead transitions, branches that cannot complete | Whether anybody follows it |
| Model-to-model matching | Two or more models | Differences between documented versions | Which version is correct |
Comparing documented models cannot tell you how work runs. It tells you what the organisation believes, in three incompatible versions, and which of those beliefs cannot all be true.
Have the tooling vendors already answered this?
Partly. Claiming the process mining vendors have no modelling layer dates you to before November 2023: SAP closed Signavio in March 2021, Celonis bought Symbio on 13 November 2023 and folded it into its Process Cockpit. Every vendor bought a modelling layer because the algorithm cannot run without event data.
What no vendor prices for is the archaeology: files nobody has opened in four years, in departments that never bought a licence. Meanwhile US patent 11,710,313 B2, granted 25 July 2023, generates event logs from video of worker screen activity. Filming staff to synthesise a log is a patented invention; reading the documentation those staff already wrote is not yet a product category.
How does documentation-first analysis fail?
In five recognisable ways, none involving the parser. Name the failures, because unnamed risks never get budget.
- Timestamp authority. Last-modified records who opened the file, not who validated it.
- String-equality matching. Literal comparison reports four processes where four vocabularies describe one. Match on normalised labels plus structure, then let a human adjudicate.
- Silent lane loss. Converters routinely drop pool and lane containment, and roles are what make a handover visible. Verify swimlanes survive parsing.
- The confident synthesis. Ask a language model to reconcile three contradictory diagrams and it invents a fourth. Route contradictions to two named owners; never auto-merge.
- Model-for-reality substitution. Guiding principle 5 of the manifesto treats models as purposeful abstractions. A library that agrees with itself proves coherence, not compliance.
What should a pre-discovery corpus audit check?
Inventory, parse, verify each model, compare across models, then report without merging anything. Below is what I run against my own infrastructure documentation: one organisation, illustrative, never a benchmark.
PRE-DISCOVERY CORPUS AUDIT — v1.0 (21 July 2026)
Run before buying discovery tooling. Not process mining.
INVENTORY
- Every .vsdx, .bpmn, .drawio, .pdf and wiki page: path, owner,
last-modified, last-validated (usually null)
PARSE
- .vsdx opened as an OPC package (ISO/IEC 29500-2), not screenshotted
- Connectors resolved to shape IDs; an unattached arrow is a drawing
- Pool and lane containment preserved, roles carried onto nodes
PER MODEL — statically decidable, zero telemetry
- One start; at least one end reachable from every node
- No dead transitions: every step reachable by some execution
- Option to complete: every branch terminates or rejoins
- Proper completion: nothing in flight at the end
CROSS MODEL — the corpus diff
- Models clustered by label and structural similarity
- Per cluster: step-set, ordering, exception-owner differences
- Role names joined to the directory, systems to the CMDB
- Terminology collisions: one label two meanings, two labels one step
REPORT
- Contradictions routed to two named owners, never merged; each
carries both source files and becomes a Stage 2 question
GATE
- Which processes hold three-star-or-better event data? Answer before
any discovery purchase. If none, the log is the project
Diffing the library costs parsing time. Discovery tooling gets bought before anyone confirms a log exists to point it at. Auditing the documentation first is arithmetic, not methodology.
Once the contradictions are settled, the interesting question arrives: for each surviving step, agent or process.
Ross Jones, Founder, The Hopium Lab. Last modified 21 July 2026.