01 · The problem
A notice is small. Its downstream impact is not.
Venues and vendors publish feed migrations, schema changes, entitlement updates, and fee notices. The operational task is to decide which feeds, datasets, systems, desks, owners, contracts, and entitlements a notice can disturb—and who needs to review it.
The toolkit models that path from unstructured notice to routed ticket. It is a technical demonstration, not a claim that an invented graph represents any firm’s production infrastructure.
02 · Truth boundary
Real inputs, synthetic estate.
Public notices
Twenty-one Nasdaq and five CME records retain source URLs, structured facts, and authored paraphrases.
Infrastructure graph
No firm publishes its topology. All relationships, contracts, entitlements, and lifecycle dates are invented.
Naming that boundary changes how the result should be interpreted. Classification can be evaluated against real notices. Impact traversal demonstrates mechanics and schema design, but not production coverage.
03 · The pipeline
Four stages, with confidence carried forward.
- Inbound notice: capture the venue message and its provenance.
- Router: infer notice type, venue, asset class, product, and effective date.
- Impact: traverse typed edges to resolve affected feeds, datasets, and systems.
- Decision: assign priority and owner, recommend an action, and require review when confidence is low.
The runtime Python package is dependency-free. The interactive D3 page contains a JavaScript port of the pipeline, so parity tests compare both implementations across the corpus and adversarial inputs. Rules are exported rather than manually retyped.
A low-confidence graph match should become a review task. It must not silently look like an unaffected system.
04 · Evaluation
Frozen before the rules existed.
A seeded stratified split assigned 19 notices to development and seven to holdout before classification rules were written. The rules were tuned on development only; held-out performance was scored once and published as-is.
Type accuracy
Feed recall
Priority accuracy
False clears
The dangerous failure
One Nasdaq notice contained no venue marker and no recognisable product name. It was confidently classified as a format change but attached to no feed, then incorrectly cleared. The escalation rule handled unknown products but not an empty resolution.
The obvious one-line fix was deliberately not applied to the published score because it was discovered on held-out data. It is a candidate for the next version and must be measured against fresh notices.
Seven held-out records are a smoke signal, not a benchmark. Labels, paraphrases, and rules share an author, so the split reduces tuning leakage but cannot eliminate labelling bias.
Known limitations
- The estate is synthetic even though the notices are real.
- Priority classification remained weak on the frozen holdout.
- One held-out notice was incorrectly cleared instead of escalated.
05 · Next iteration
Fix the failure, then earn a new measurement.
- Escalate whenever feed resolution is empty.
- Grow the real notice corpus and make fresh records the next holdout.
- Deepen the shallow CME records where public access allows it.
- Add the two unsupported governance checks only when the graph schema can express them honestly.
The point of the toolkit is not that the first router is complete. It is that the system makes its assumptions, boundaries, failures, and next decision inspectable.