The Eight Lenses › Lens 4: Scientific Evidence
Lens 4: Scientific Evidence
What this lens looks for
This lens has two blades, and the second is the one that distinguishes it from ordinary fact-checking. The first asks whether claims are supported: are there studies, data, benchmarks, citations, and do they actually say what the communication says they say? The second asks whether the limits of that evidence are preserved: sample sizes, populations, effect sizes, confidence, contested replications, funding sources, and the boundary beyond which the evidence says nothing.
A communication can pass the first blade and fail the second completely — every citation real, every number accurate, and every hedge from the original study stripped away in transit. The result is scientific in costume and overconfident in content. The core axiom — factually correct, epistemically impoverished — was practically written for this case.
Why it is an architectural constraint
Including a science lens in a framework grounded in critical humanities is deliberate. Preservative AI's commitment to plurality is not relativism: evidence matters, methods matter, and some claims are better supported than others. What the lens refuses is scientism — the use of science's authority without its discipline. The lens treats the scientific tradition the way the framework treats every tradition: as a powerful way of knowing with a defined domain and real limits, which must be preserved along with its findings. Benchmark culture in AI is a standing example: a leaderboard number is evidence about a benchmark, routinely circulated as evidence about the world.
Characteristic gaps this lens detects
- Stripped hedges — "may, in mice, under these conditions" becomes "does"; uncertainty removed between the paper and the press release.
- Citation laundering — real sources attached to claims they do not make, or chains of secondary citation that never reach a primary source.
- Benchmark inflation — performance on a narrow evaluation presented as general capability, without the benchmark's scope or known gaming.
- Absent denominators — risks and effects without base rates; "doubled" without "from what".
- Methodological silence — results presented with no indication of how they were produced, making independent assessment impossible.
- Boundary violations — evidence from one domain quietly extended to another (lab to deployment, one population to all populations).
When the lens binds
It binds strongly on any communication invoking studies, data, percentages, benchmarks, or the authority of research — which includes most corporate AI communication. It binds moderately on narrative and opinion pieces, where the question becomes whether empirical claims smuggled into the narrative are flagged as such. It binds on this site too: where the OtherAI's own claims are conceptual rather than empirical, this lens requires saying so plainly (see Limitations).
Scoring guidance
| Score | Indicates |
|---|---|
| 0–24 | Authority of science invoked with no inspectable evidence, or evidence misrepresented. |
| 25–49 | Evidence present and broadly accurate, but limits stripped: no hedges, no scope, no methods. |
| 50–79 | Claims anchored to inspectable sources with limits preserved; uncertainty stated where it exists. |
| 80–100 | Evidence and its boundaries fully legible: primary sources, methods, contested points, and what remains unknown. |
Relation to Suffixscape
This lens pairs naturally with the Epistemic Inflation diagnostic: inflation is the grammatical signature ("state-of-the-art", "unprecedented") of the content-level gap this lens scores (superiority asserted without methodology). When the diagnostic fires and the lens finds no supporting evidence, the two findings corroborate — a pattern the audit workflow exploits deliberately.