The lab records position, adjacency, damage, and arithmetic before considering any social, lexical, or phonetic interpretation.
Linear A / evidence controls
Structure before translation
A source-linked lab for Linear A administrative documents: identified objects, preserved uncertainty, neutral structural fields, and auditable arithmetic.
Identified corpus records No phonetic decipherment claim Exact / conditional separation
HD, FLD, NUM, SUM, and UNK describe observable structure. They are not phonetic values, translations, persons, commodities, or offices.
Exact relations and conditional relations are displayed separately. Damage and ambiguous segmentation remain visible rather than being silently repaired.
HT 122b: exact cross-section total
a.8 KU-RO 31; b.5 KU-RO 65; b.6 PO-TO-KU-RO 97
Template match
SUM-NUM / SUM-NUM
A source-linked cross-section relation, not a universal translation rule.
Arithmetic validation
exact local relation
97 = 97
Residual 0
Published relation: 65 + 31 + 1 = 97. The additional unit is reported in the corpus commentary.
Role system
Primary labels describe evidence before they suggest meaning
Initial field
A sign-group or heading preserved at the beginning of a documented section.
Evidence rule
Positional description only; no person, office, or lexical meaning is asserted.
Record field
An attested sign-group or logogram associated with a numerical entry.
Evidence rule
The label records layout and adjacency, not a translation such as commodity or recipient.
Numerical field
A published Linear A numeral or fraction copied from the cited transcription.
Evidence rule
Numerical membership is the strongest directly testable class in the model.
Terminal summary candidate
A sign-group occurring in a terminal position beside a quantity that participates in a published total relation.
Evidence rule
KU-RO and PO-TO-KU-RO are treated structurally here; the label is not a phonetic or etymological claim.
Unresolved
A damaged, ambiguous, weakly supported, or structurally unstable field.
Evidence rule
Uncertainty is retained rather than converted into a convenient interpretation.
Audit cockpit
The model preserves an explicit boundary between observation and interpretation
Selected audit node 01
Document identity
Claim
Every public control is tied to an identified object rather than an invented sample line.
Evidence
HT 122b is HM 1366, GORILA I: 208-209; HT 102 is HM 1310, GORILA I: 168-169.
Risk controlled
A synthetic example can be mistaken for an inscription and contaminate later interpretation.
Verdict Source-linked
Template concentration
Only source-visible structural patterns are shown as current controls
Corpus-wide frequencies remain unpublished until the normalized line dataset and deterministic analysis are released. These patterns describe only the cited controls.
A source-linked record field followed by an attested numerical field.
A terminal summary candidate followed by an attested numerical field.
Separate section totals compared only where the corpus commentary publishes an arithmetic relation.
Validation register
Published, conditional, withheld, and pending claims remain distinct
| Test | Status | Evidence |
|---|---|---|
| Document identity | Published | HT 122b / HM 1366 and HT 102 / HM 1310 |
| Corpus provenance | Published | GORILA I references with GORILA-derived record links |
| Arithmetic relation | Separated | Exact HT 122b; conditional HT 102 |
| Semantic interpretation | Withheld | Neutral structural labels replace Agent / Commodity / Operation |
| Corpus-wide statistics | Not asserted | Requires a versioned line-level dataset and executable analysis |
| Permutation / stability | Pending corpus release | No p-value or stability score is presented before computation |
Source-linked arithmetic controls
One exact local relation and one explicitly conditional relation
PO-TO-KU-RO 97 is reported as KU-RO 65 + KU-RO 31 + 1.
The total closes only under the published conditional treatment of WI and the damaged entry.
Failure modes
A serious decipherment-adjacent system must publish its refusal rules
Damaged numeral or sign
A required operand, terminal quantity, or sign-group is restored, missing, or uncertain.
Retain the corpus damage notation and downgrade the relation to conditional or unresolved.Ambiguous segmentation
The source permits more than one division into sign-groups or numerical fields.
Show the alternatives; do not select one solely because it closes an equation.Arithmetic inconsistency
Attested operands do not equal the proposed total under the documented segmentation.
Report the residual without inventing an operator or tolerance that licenses it.Semantic overreach
A positional class is presented as a person, commodity, office, verb, or translated word.
Return to the neutral structural label unless independent evidence supports the interpretation.Method pipeline
Constraint-first inference keeps the model from becoming speculative decipherment
- Preserve the document ID, side, line, museum number, and corpus reference before normalization.
- Copy attested sign-groups, logograms, numerals, damage marks, and segmentation without silent repair.
- Assign neutral positional classes before considering semantic interpretations.
- Test arithmetic only when every operand and the proposed total have explicit provenance.
- Separate exact, conditional, damaged, and unresolved relations.
- Publish corpus-wide claims only from a versioned dataset and deterministic executable analysis.
Framework manuscript
ML-model-01: a method draft, separated from validated corpus evidence
The manuscript preserves the original formal proposal. Its aggregate tables are not treated here as independently validated results until line-level inputs and executable analysis are released.
Open method manuscriptHT 122b and HT 102 with document-level provenance.
GORILA-derived records and SigLA palaeographic reference.
Exact, conditional, unresolved, and pending results are never merged.
Structural evidence only; no lexical or phonetic claim is forced.