Method
How claims are labelled, scored and kept from turning into fake translations.
Research compilation
Research compilation from a published ICIT-style digitization — not a complete or licensed photographic corpus
Evidence levels
- ObservedDirectly visible, catalogued or source-documented information.
- Structural findingA reproducible sequence or positional pattern validated against controls.
- Functional hypothesisA proposed role based on structure and object or context evidence.
- Semantic candidateA possible reading based on a named theory — not an accepted translation.
- Unverified proposalAn untested, weakly supported or external claim.
- Rejected or unsupportedA hypothesis that failed predefined tests or lacks sufficient evidence.
Evidence score
E(H) = 0.25S + 0.20R + 0.15D + 0.15X + 0.15C + 0.10V, where S is statistical strength, R replication across independent objects, D directionality, X cross-context, C contextual coherence and V visual-identity confidence. Components are shown separately. The total does not confirm a translation.
Null models
The default shuffle keeps each inscription’s length and inventory and randomises order. Observed transitions are compared with that null. Weak samples are labelled as such.
What we will not say
- The Indus script has been solved.
- This sign definitely means X.
- AI has translated Mohenjo-daro.
- A fish-shaped sign definitely means fish, star, god or king.