ONLY FOUND HERE · METHOD
How an edition is made
What the public signals show, what the provider says, and what would change our read.
- 1 · Evidence packetAn immutable snapshot of ORBIT's ledger and approved public feeds. Every record has an ID, URL, publisher, source family, type, observed/published/fetched UTC, window, scope, freshness and reuse permission. Unavailable sources are recorded too.
- 2 · Rules, not a modelAngles are detected by deterministic rules; numbers are computed in code; sentences come from templates. Feed text is data: instructions inside it are withheld and never followed.
- 3 · ValidatorJSON Schema plus checks that every evidence ID exists, scope and clocks match the records, cited times and regions appear in the cited evidence, root-cause wording is backed by the provider, "conflict" is used only for comparable claims, and no private data or novelty claims slip in.
- 4 · Rubric & decisionpublish_candidate, hold or skip. A routine recap with no original contribution is skipped.
- 5 · Cloudy_Day reviewsShadow mode: every real edition waits for his approval.
Analyst rules
- All times normalized to UTC; unknowns are labelled, never estimated.
- Gap when the public record leaves something unanswered; conflict only when comparable claims (same component, operation, region and window) disagree.
- Interpretations are ranked with evidence for and against.
- One original contribution per edition: a normalized timeline, scope correction, measured recovery lag, baseline comparison or documented dependency chain. No “nobody else reported this” claims.
- Every verifiable claim carries evidence IDs, a type (observed, provider-reported, inferred, unknown), a confidence label and the reason.
- Never fabricate quotes, times or sources.
- Public status notices, disclosed postmortems and permitted aggregates only. Private reports, inboxes and case files never go in.
Rubric (0-2 each)
- Provenance: 2 = every claim linked and at least one source family independent of the provider; 1 = linked but single lineage; 0 = an unlinked claim.
- Scope & timing: 2 = scope and clocks validated with a known impact start; 1 = an unknown start; 0 = a scope or clock mismatch.
- Originality: 2 = a measured contribution (lag, baseline, timeline with numbers); 1 = qualitative (scope, dependency); 0 = none.
- Counterevidence: 2 = a ranked alternative with evidence for or against and stated missing evidence; 1 = an alternative without evidence; 0 = none.
Exact auto-publish rule (built, switched OFF)
An edition could publish unattended only if all hold: rubric total ≥ 7/8; no zero on any line; overall confidence not low; no hard-stop flag; validator clean; decision = publish_candidate; Cloudy_Day's switch is ON in his owner desk; the Pages env var OFH_AUTOPUBLISH=1; neither the global nor the per-incident pause is set. Today the switch is OFF: every real edition is held for his approval.
Hard stops (always reviewed by a person): suspected compromise, named wrongdoing, sensitive disclosure, material conflict, root cause claim.
Triggers: a material revision of the ledger (new or changed evidence, not just a re-fetch) or a 10-minute cadence. Duplicates are grouped by canonical incident. A broad news spike alone goes to "watch", not to a draft.
Live metrics
Edition JSON SchemaAll editions
Framework cloudy-day-1.0 · schema 1.0.0.