status-quo dataset

Raw and LLM-interpreted incident data collected from public status pages of 10 SaaS providers (Atlassian Statuspage instances: GitHub, Cloudflare, Discord, Reddit, Vercel, Linear, Notion, Netlify, DigitalOcean, npm). Feeds the status-quo project — public code, public dashboard at ronniechong.com/status-quo. This dataset itself stays private; the dashboard only ever ships small precomputed JSON summaries derived from it, never this raw data directly.

See the code repo's LLM_WORKFLOW.md for exactly how interpretation works — what the model is/isn't asked to do, and how the "no invented facts" constraint is enforced in code, not just by prompting.

Structure

Partitioned by provider and month, appended to by a scheduled pipeline (~6-hourly fetch, batched export). Never rewritten in place — each file covers one provider's one calendar month.

data/{provider_id}/{YYYY-MM}.parquet — raw fetch snapshots

column	type	notes
id	string	snapshot id
provider_id	string	one of the 10 provider ids above
fetched_at_utc	string (ISO)	when this snapshot was collected
http_status	int	response status from the provider's API
body	string	raw JSON response body, verbatim
normalized_timestamps	string (JSON)	per-incident {raw, utc} timestamp pairs derived from body

interpretations/{provider_id}/{YYYY-MM}.parquet — LLM-tagged incidents

One row per resolved incident, keyed by (incident_id, provider_id, prompt_version). Fields split into two groups:

LLM-produced (title/summary/taxonomy only — see LLM_WORKFLOW.md for the exact prompt and constraints): title, summary, affected_surface, fault_origin, workaround_offered, workaround

Computed in code, never asked of the model: time_to_first_update_min, updates_per_hour, component_count, is_retroactive, severity (provider's own reported word, unmodified), source_url, created_at, resolved_at, duration_hours (null if the incident is still open, or if the provider's own created_at/resolved_at gap is under 60 seconds or negative — treated as unreliable rather than published as a misleading number), incident_status

Provenance (always present, never hidden): model_used, prompt_version, schema_version, interpreted_at_utc

Currently-open incidents get a lightweight no-LLM row instead (model_used: "none", prompt_version: "raw-v1") — upserted every cycle until the incident resolves and a real interpretation replaces it.

coverage/latest.json — per-provider collection metadata

Single file, fully overwritten each export. Collection start time, last successful fetch, and any observed collection gap windows per provider — feeds the dashboard's "history depth" and coverage indicators.

Non-negotiables

• No fact in title/summary beyond what the incident's own update text states — enforced by prompt constraints, not aspirational.
• Every metric that can be computed deterministically (durations, counts, rates) is, architecturally, never left to the model.
• severity is always the provider's own word, verbatim — never a model judgement, never compared or ranked across providers.