Generated from
cat-harness/skills/authoring/authoring-core/evidence-review.md— do not edit here.
Evidence review — the same rule, a second instance
Skill id:
evidence-review· Package:folio-core
This skill exists to demonstrate a claim rather than to assert it.
untainted-verification states a discipline and
says it is generic. One instance is an example; two unrelated ones are
evidence. The first is the translation round trip. This is the second, and it
is as far from translation as the platform reaches: the appraisal of evidence
behind a recommendation.
The rule is not a platform invention
The party that drafted the recommendation does not write the evidence appraisal for it.
That is untainted-verification’s sentence with the nouns changed, and it long
predates this repository. Guideline methodology separates the systematic
review team from the guideline development group for exactly the reason
the spine gives: a party assessing its own output is comparing a thing with its
own paraphrase of itself.
Which is the useful finding here. The platform did not need to invent a rule for evidence review — it needed to notice that the rule it already had was the same one.
untainted-verification | here |
|---|---|
| producer | whoever drafts the recommendation |
| checker — given the artefact, and nothing that would let it shortcut | the appraisal, given the question and the evidence, not the draft’s conclusion |
| adjudicator — given the intent and the checker’s output, never the artefact | the evidence-to-decision judgement |
Rule 1 of the spine has real teeth in this direction. An appraiser shown the recommendation the guideline group has already drafted is being told the answer before being asked the question, and a “the evidence supports it” produced that way measures the instruction, not the literature.
Recurring surveillance is staleness, generalised
A living guideline is continuously updated as evidence appears rather than revised on a multi-year cycle. So the appraisal is not a one-off chapter — it is a verdict with an expiry.
The spine already has that idea: a verdict is hashed to what it was about, so editing the subject stales it. A living guideline stales the same verdict for a different reason — the subject did not change, the evidence base did.
That gap is stated, not papered over. The spine’s field_hash covers
inputs that live in the repository. An external corpus does not, and nothing
here pretends otherwise. What the two share is the shape a consumer must
respect either way: “verified on DATE against THESE inputs” is a different
claim from “verified”, and a reader given the second cannot tell a current
verdict from a lapsed one. Any folio implementing surveillance owes an explicit
expiry, in the same way bean-blocking requires one on a
block: a verdict with no expiry cannot be told from an abandoned one.
The fencing analysis — which rules are the platform’s, and which are not
Run through domain-fencing’s three
questions, in order, first-to-answer wins:
| candidate rule | Q1 rationale survives translation? | Q2 domain in mechanism or vocabulary? | Q3 needs the domain installed? | verdict |
|---|---|---|---|---|
| appraiser ≠ drafter | yes, word for word | vocabulary | no | platform — it IS the spine |
| a verdict carries what it was verified against, and when | yes — it is the field_hash argument | vocabulary | no | platform |
| a finding is upheld by opening the cited source, never by trusting the rule’s wording | yes | vocabulary | no | platform — and already the practice in voice-who-guideline-development |
| certainty rated across the GRADE domains | no | mechanism | yes | fence |
| PICO question formulation | no | mechanism | yes | fence |
| evidence-to-decision framework tables | no | mechanism | yes | fence |
| a search strategy is re-runnable — databases, dates, query recorded | yes in rationale | mechanism is bibliographic | yes | fence the mechanism, keep the rationale (it is the row above) |
Three rows earn a place on the platform and four do not, and the four are
not a backlog. They are a folio’s methodology, correctly located. The honest
move is to fence them and say so — which is what q-usage-watcher does in its
first paragraph, and what this table does here.
Nothing is shipped for the fenced rows. Declaring empty criteria for them
would be bean dh4f exactly: a consumer scanning nothing and reporting a clean
run over it. A folio that wants them declares them, in its own repository,
where its methodology lives.
What this skill does NOT do
It asserts no clinical or empirical claim, and the platform must not. The mechanism carries process — who may rule on what, given what. The evidence is the folio’s, the judgement is the guideline group’s, and a platform that shipped a default answer to either would be asserting on their behalf.
This is the same line content-profiles draws for block kinds and
domain-fencing draws for QA axes, arrived at from a third direction.
Related
untainted-verification— the discipline. Read it first; this skill adds only what evidence review instantiates.translation-manager§”The agentic round trip” — the other instance, deliberately unrelated to this one.voice-who-guideline-developmentin the IRIS replica’s style guide — how a recommendation is worded. It states its own boundary: it governs the language of normative statements, not the process that produces them. This skill is on the other side of that line, and the two do not overlap.domain-fencing— the three questions, and why a fenced rule is correctly located rather than missing.