Worked example
What the Diagnostic actually returns.
Below is a complete result for one described workflow. It was not written by hand to look convincing — it was produced by running the answer set in the appendix through the same scoring function the live tool calls, at build time. If the rubric changes, this page changes with it.
The scoring is deterministic: a fixed set of lookup tables, no model anywhere in the decision path. The methodology exhibit publishes every table used below.
Sample result
“Client-requested performance report generation”
Legibility grade
Medium
Exception mass
Low
Verification class
V1
Risk class
II
Max autonomy tier
A2 · Exception-gated
Capped from A3
Oversight mode
Sampled + pre-execution checks
Each score traces back to a specific answer — e.g. “Risk II because of how reversibility and reach were rated.” Inconsistencies are flagged, not hidden.
Recommended architecture pattern
Deterministic orchestration with model steps
Workflow complexity W2 · oversight: Sampled + pre-execution checks · floor: Exception-gated with documentary provenance
Ranked risks
Automation-decay risk: ad hoc workflows drift fastest, because there is no cadence forcing a recheck.
Counterfactual-baseline note: no automation decision here is complete without the current process's measured error rate to compare against.
Narrative memo
You described a workflow that runs on an ad-hoc basis, run by one person, with documentation you called “current and accurate” and practice that “rarely” deviates from it — a legibility grade of Medium — legible in parts, but not end to end. Your exception rate is low and the exception types are fully enumerable, which is why exception mass scored Low.
Because checking one output takes about half the time of producing it, and a bad instance would be caught by a simple reconciliation check, this lands at V1 — cheap to verify. Combined with Risk Class II (reversible with effort, affects many clients), that supports A2 autonomy: the system can run, but exceptions are gated for human review before they go out. The base lookup gave A3; a cap pulled it down to A2.
This tool scores what you told it. The paid assessment scores what we’d observe — and legibility debt is usually the gap between the two. On a Medium-legibility process like this one, that gap is usually worth finding before you automate.
Rubric v1 · same engine, same rubric as a live run
Appendix · the answers that produced this result
Nothing else was supplied. Each score above is a lookup against these inputs — including the one that mattered most here: the base autonomy lookup returnedA3, and the regulatory-obligation answer pulled it down to A2. The tool reports the cap rather than hiding it.
A · The workflow
- Runs ad hoc, performed by one person.
- Client-requested performance report generation.
B · Legibility
- Documentation: current and accurate.
- A competent replacement could roughly follow it.
- Practice rarely deviates from the documentation.
C · Exception mass
- Under 5% of instances need special handling.
- Exception types are all known.
D · Verification
- A bad instance would be caught by a simple reconciliation check.
- Checking one output takes about half the time of producing it.
E · Risk
- A mistake is reversible with effort.
- It affects many clients.
F · Constraint
- Judgment calls are rarely required.
- The work sits under a regulatory obligation.
G · Where information lives
- Structured systems.
What this sample does not show
- It scores what someone says about a workflow. The paid Assessment scores what we observe — and the gap between the two is usually where the legibility debt sits.
- It reads one workflow in isolation. It cannot see the undocumented judgment call that sits between two documented steps, because nobody thinks to report it.
- It produces a recommended ceiling on autonomy, not an implementation plan.