Evidence and confidence
A husbandry fact is not a number. It is a claim, made by somebody, on some basis, at some time, possibly contradicted by somebody else — and the parts that usually get thrown away are the parts you need when your tank does not match the care sheet.
Evidence type is not confidence
“Two hundred aquarists agree” and “one paper measured it” are different KINDS of support, and neither is automatically stronger. AquariumHQ stores them as separate axes and displays them as separate axes, so a confidently held editorial position and a tentatively held measurement do not look the same.
Worked examples
Two cases where the evidence did not say what AquariumHQ had previously reported. Each renders its figures live from the evidence layer rather than restating them, so the page cannot drift from the data behind it.
- Is the zebra danio a tropical fish?It is sold as one, kept as one, and the sources do not support it. This is also the case study where AquariumHQ's previous account of the evidence turned out to be wrong about which source said what.
- How big a tank does an angelfish need?The recorded figures differ by a factor of nearly two, and the disagreement is not about measurement. It is about what the number is for — and about a dimension that volume cannot express.
- The two rams are not the same fishSold from adjacent tanks under one word. The wild records for the German blue ram and the Bolivian ram do not overlap in temperature or in pH — and the ranges AquariumHQ advises for them do. Both of those facts are true, and the gap between them is the interesting part.
- Why mollies and tetras are usually a water problemBoth are sold as community fish. That phrase describes temperament, and the difficulty here is not temperament — it is that the water one is recorded from is water the other is not. It also is not one answer: some tetras are far closer to the molly than others.
Evidence types
| Type | What it means |
|---|---|
| Observed / measured | A measurement or direct observation reported by an identified source. |
| Sourced recommendation | A recommendation attributable to a specific named source. |
| Aquarium-literature consensus | Broad agreement across published aquarium husbandry practice, without a single citable locus. Not a measurement. |
| Derived from other data | Follows from other stored claims by a stated rule. |
| Calculated | Produced by a published formula from inputs shown on the page. |
| Legal requirement | Imposed by law in the named jurisdiction. Verify locally before acting. |
| AquariumHQ editorial position | An AquariumHQ position taken where evidence is incomplete. Labelled so you can discount it. |
| Insufficient evidence | We looked and there is not enough reliable information to state a value. |
Confidence levels
Deliberately a five-point word scale. Nothing in this domain supports two significant figures of epistemic precision, and “confidence: 0.82” would be false precision dressed as rigour.
| Level | How to read it |
|---|---|
| High | Well established across multiple independent lines of support. |
| Moderate | Widely agreed, with some variation between sources or contexts. |
| Low | Thinly supported. Treat as a starting point rather than a specification. |
| Very low | Barely supported. Included because omission would be worse. |
| Unknown | We have not assessed it. |
Stored levels: high, moderate, low, very-low, unknown.
Disagreement is data
Where sources give different ranges, AquariumHQ keeps all of them and derives an interpretation that says so. Collapsing three sources into one confident number is the single most common way aquarium information becomes wrong, because the collapse is invisible afterwards.
The conflict detector distinguishes a SPREAD — three sources describing the same animal through different lenses, whose ranges overlap — from a genuine DISJOINT conflict, where the ranges do not overlap at all. Only the second is presented as a dispute; reporting the first as one would be noise.
Saying we do not know
“Insufficient evidence” is a valid stored answer and must remain cheap to give. A schema with no way to say “we do not know” will be filled with guesses, and a claim marked insufficient-evidence is structurally forbidden from carrying a numeric value.
The same principle runs through the compatibility engine, which returns “insufficient evidence” rather than a fabricated verdict when trait data is missing, and through the coverage report, which counts worthwhile questions we cannot answer rather than quietly omitting them.