Conflicting signals

A witness conflict is two different official addresses stating different things about one document. This page states that they differ, shows what each said and when each was read, and stops there. It does not prefer the newer reading, the more official publisher, or the value this record already holds.

Two surfaces disagreeing and one surface having moved are different facts. A publisher editing its own page between two reads has revised it, and a revision is a fact about one page over time. Filing one as the other would have this record accusing a publisher of contradicting itself when all it did was publish an update, so the grouping key is the address and the two classes are counted apart.

What is held

• 61 reading(s) of an official surface, across 5 store(s), over 132 tracked instruments.
• 46 of 132 tracked instruments carry at least one such reading, held across 52 distinct (source, address) surface(s).
• 6 of 132 tracked instruments name two or more distinct addresses that state a status. That is the comparable population, and it is the only population a conflict can be found in. A source with one address is not a negative result; it is a source with no second surface to compare against.
• 0 of 6 comparable source(s) carry conflicting signals; 6 of 6 state one value across every surface. The zero is a comparison that ran and found agreement — not a comparison nobody attempted, and not one that failed.
• 0 of 52 surface(s) carry readings that differ from each other: one address, two answers. That is one page that moved or two readings of one page, it is counted here and never inside the conflict figure, and it is measured over every surface rather than only the comparable ones.

The comparable population, source by source

Ordered by source id. Ordering by verdict would put conflicts first, and geometry counts as a claim.

Surfaces whose own readings differ

None, over the 52 surface(s) held. Every address this record holds more than one reading for has been read the same way each time. That is a measurement over the whole population, not a sample of it.

The three kinds of value the order names

The stores this page reads

Intervals across the comparable population

• 0 of 6 comparable source(s) — same-day: both readings carry a day and it is the same day.
• 3 of 6 comparable source(s) — dated: both carry a day; the span between them is stated.
• 3 of 6 comparable source(s) — undated: at least one store records no read day at all.

What this page counts, and what it does not

• Nothing here is resolved. Where two surfaces differ, both values render with their citations and their read days. No branch prefers one, and there is no field on this page for a settled answer.
• The comparison is on the value each store recorded, in this record’s own status vocabulary. Publishers do not share one, so two surfaces can state different publisher tokens that this record wrote down as the same value — and this comparison would then read them as agreeing. That is a limit of the comparison and not a property of the documents, and it is stated because a limit left unstated reads as a census.
• Every reading below the comparable rows is a field reading — the publisher’s own field name and its own value, composed into a line by the collector that read them. It is not a passage of the publisher’s running prose, and it is not set as one. Retained passages live in state/status-evidence/; none belongs to a source with a second surface, so no comparison here rests on one.
• A line over the 14-word ceiling is withheld whole, with its word count, and never trimmed. The ceiling has one home, scripts/quote_cap.py, and this page reads it rather than spelling a second one.
• Two of the 5 stores record no read day, and the absence renders “—”. Admitting only dated readings would have cut the comparable population from 6 to 3, and a requirement widened to shrink the evidence is the same defect as a gate widened to admit it.
• The address is compared as the store recorded it, minus its query and any trailing slash. Nothing is redirect-followed: the answered URL is the record’s URL, and resolving here would ask the network what the record already holds.
• One rule has two readers, and that is a gap. The join between a source and FDA’s guidance datatable lives in scripts/signals.py; this page cannot call it, so src/lib/conflicts.ts restates it and scripts/witness_conflict.py calls the original, so the two derivations can be compared. The structural repair is a collector that writes those readings into a store, which both readers would then read.

Last updated: 18 September 2026