When a prediction fails, the hypothesis is only one of the things that could be wrong. The test
configuration, the sample, the retention assumption, the tool's own coverage — any of them can
produce the same negative, and the Duhem–Quine thesis is the observation that a hypothesis is never
tested alone but always as a bundle with its auxiliaries.
peira has attack edges and a grounded extension that computes what survives them. What it does not
have is any requirement that an attack say what it defeats. attacks: [c1] asserts that c1 falls;
it does not say whether the claim itself is wrong, or whether one of the assumptions the claim rests
on has failed. Those are different findings with different remedies, and the graph currently cannot
tell them apart.
The forensic consequence is concrete. "The tool found no trace, therefore the file was never
present" competes with "the tool does not index that structure" and "the retention window closed
before the event". Only the first is a finding about the world. In the current model all three are
written the same way — an attack edge, or an absence — and the grounded extension treats them
identically.
Proposed: an attack edge may name the link it breaks, and a claim resting on stated auxiliaries
must declare them. Two candidate shapes:
PEIR-AUXILIARY-UNDECLARED — a claim at the intervention or counterfactual rung with no declared
assumes: list. Those rungs already require an executed protocol; a protocol has assumptions.PEIR-ATTACK-UNSCOPED — an attack edge that names no target link, where the attacked claim
declares auxiliaries it could have named instead.
Depends on nothing already open, but composes with the Instrument work (#8): "the tool does not
index that structure" is an instrument-validity fact, and "the retention window closed" is a
coverage-boundary fact. Both are auxiliaries in the Duhem–Quine sense.
The design risk to watch: an assumes: list is a field an author can fill mechanically. The gate
must not reward having a list — the question is whether an attack can be scoped to one, which is a
graph property rather than a text property.
When a prediction fails, the hypothesis is only one of the things that could be wrong. The test
configuration, the sample, the retention assumption, the tool's own coverage — any of them can
produce the same negative, and the Duhem–Quine thesis is the observation that a hypothesis is never
tested alone but always as a bundle with its auxiliaries.
peira has attack edges and a grounded extension that computes what survives them. What it does not
have is any requirement that an attack say what it defeats.
attacks: [c1]asserts that c1 falls;it does not say whether the claim itself is wrong, or whether one of the assumptions the claim rests
on has failed. Those are different findings with different remedies, and the graph currently cannot
tell them apart.
The forensic consequence is concrete. "The tool found no trace, therefore the file was never
present" competes with "the tool does not index that structure" and "the retention window closed
before the event". Only the first is a finding about the world. In the current model all three are
written the same way — an attack edge, or an absence — and the grounded extension treats them
identically.
Proposed: an attack edge may name the link it breaks, and a claim resting on stated auxiliaries
must declare them. Two candidate shapes:
PEIR-AUXILIARY-UNDECLARED— a claim at the intervention or counterfactual rung with no declaredassumes:list. Those rungs already require an executed protocol; a protocol has assumptions.PEIR-ATTACK-UNSCOPED— an attack edge that names no target link, where the attacked claimdeclares auxiliaries it could have named instead.
Depends on nothing already open, but composes with the
Instrumentwork (#8): "the tool does notindex that structure" is an instrument-validity fact, and "the retention window closed" is a
coverage-boundary fact. Both are auxiliaries in the Duhem–Quine sense.
The design risk to watch: an
assumes:list is a field an author can fill mechanically. The gatemust not reward having a list — the question is whether an attack can be scoped to one, which is a
graph property rather than a text property.