How a score is worked out › ledger-strict@0.2.0
ledger-strict@0.2.0
The cautious counterpart. It refuses to score the claim types where reasonable models differ most — causal and interpretive — rather than produce a number it would have to hedge. Where the two models disagree about a claim, that disagreement is itself reported, on the weaknesses page.
Version 0.2.0. Evidence from a source the claim was taken out of stops counting as support for it — that is provenance, not corroboration. Ungrouped evidence is keyed by origin and passage, so one page recorded twice as two sources counts once.
Released at entry seq 4002 ·
config sha256:19333a3ad9ad67be52a3ee10299037849ff47532fc592a9bba63a2278eb35342 ·
code sha256:884a61eba176cfd67305a75bf1fa5557dff0f30aad7807326839db1bc0513acc
A released model never changes. Those two hashes are recorded in every score it produced, so any
number it has ever given can be reproduced exactly, by this node or by anyone else.
What this version changed
| Setting | Was | Now |
|---|
| how ungrouped evidence is grouped | not declared | origin |
| evidence from the claim's own source | not declared | self 0.0, independent 1.0 |
Released alongside it, from the same code, with different settings.
| Setting | ledger-default@0.2.0 | ledger-strict@0.2.0 |
|---|
| types it marks as model-dependent | causal, interpretive | none |
| observation weight: expert analysis | 0.3 | 0.0 |
| prior: causal | 0.35 | not declared |
| prior: interpretive | 0.50 | not declared |
| claim types it scores | observational, quantitative, historical, textual, comparative, definitional, attributed belief, causal, interpretive | observational, quantitative, historical, textual, comparative, definitional, attributed belief |
What it scores, and what it declines to
Scored: observational, quantitative, historical, textual, comparative, definitional, and attributed belief.
Every other claim type gets not applicable from this model and no probability at all —
a refusal to score, not a score of nought.
Every number it declares
What each of these means, and why it is there, is on
how a score is worked out.
Prior by claim type
| Setting | Value |
|---|
| textual | 0.50 |
| historical | 0.50 |
| comparative | 0.50 |
| definitional | 0.50 |
| quantitative | 0.50 |
| observational | 0.50 |
| attributed belief | 0.50 |
Relevance
| Setting | Value |
|---|
| weak | 0.2 |
| direct | 2.0 |
| strong | 1.2 |
| moderate | 0.6 |
| context only | 0.0 |
Observation
| Setting | Value |
|---|
| other | 0.1 |
| hearsay | 0.1 |
| eyewitness | 0.4 |
| calculation | 0.7 |
| direct text | 0.9 |
| measurement | 1.0 |
| model output | 0.0 |
| archaeological | 0.9 |
| dataset result | 0.9 |
| expert analysis | 0.0 |
Authenticity
| Setting | Value |
|---|
| doubtful | 0.3 |
| verified | 1.0 |
| unverified | 1.0 |
Extraction
| Setting | Value |
|---|
| verified | 1.0 |
| unverified | 1.0 |
Provenance
| Setting | Value |
|---|
| self | 0.0 |
| independent | 1.0 |
State thresholds
| Setting | Value |
|---|
| supported | 0.80 |
| leans supported | 0.60 |
| unresolved above | 0.40 |
| leans contradicted above | 0.20 |
The rest
| Setting | Value |
|---|
| interpretive step penalty | 0.15 |
| independence | strongest only per group and direction, ungrouped evidence keyed by origin |
| rounding | mode half_even, weights 6, coverage 2, probability 4, boundary guard 1e-6 |
| review checks | primary source reviewed, opposing search done, independence reviewed, qualifiers reviewed |
| primary source types | primary text, dataset, measurement, legal document |
Other released models
Galedra