Key idea & worked example
A processing candidate is a proposed interpretation of source material. Review and publication are additional decisions, with their own evidence and responsibilities.
Automation can propose a useful structure; an accountable workflow keeps the interpretation and its grounds inspectable.See the worked example →On this page
Follow the input, output and review boundary
The September 2026 portable module guide describes twelve modules in Processor 0.2.0, an unpublished development candidate. The following map explains that source snapshot’s responsibilities. It does not establish independent releases, broad benchmarks or public services for every module. Choose a module by its input and output, and retain the applicable decision and projection policy with the result.
| Module | Input → output | Interpretation and review |
|---|---|---|
| Ingestor | Registered transfer records → source-preserving staging records. | Schema-valid staging establishes an intake contract; it does not make the content an accepted canonical fact. |
| Normalizer | Staged source records → deterministic normalized candidates with original evidence. | Review and projection remain explicit. Exact replay preserves existing decisions instead of silently replacing them. |
| Bootstrap | Versioned policies, observations and decisions → governed projection receipts. | Projection depends on the authorized policy and its prerequisites; it is not unrestricted automatic acceptance. |
| Document evidence | Identified document bytes → pages, assets, exact selectors and candidate assertions. | A quotation must resolve to its source. Silence or an absent value must not become an invented numerical fact. |
| Localizer | Place mentions and provider evidence → ranked geographic candidates. | Retain ambiguity, temporal context and unknown positional accuracy. A high-ranked candidate is not automatically the historical location. |
| Language detection | Source text → language-identification evidence for quality review. | Short or ambiguous text can justify abstention. Uncalibrated scores are not probabilities of correctness. |
| Litparser | Citation strings → literature candidates with literal field evidence. | Conflicting years and indefinite page continuations remain uncertain; parsing does not authorize a canonical bibliographic write. |
| Enrichment | An authorized page or workflow → retained institution or virtual-instrument facts. | Fetching is opt-in and source-bound. Replaying retained evidence is distinct from obtaining a new page snapshot. |
| PLENUM | Claim-scoped evidence → versioned suitability and source-dependence assessments. | The result is a policy-based assessment for a claim, not permanent trust in a publisher or a truth probability. |
| Ambitus | MusicXML, compressed MXL or MIDI → nominal pitch-range evidence. | Instrument identity remains a separate question. MIDI pitch-control information is flagged rather than silently treated as measured acoustic pitch. |
| Imageseg | Image and SAM 2 configuration → aligned masks and annotations. | Masks are class-agnostic. A selected region needs interpretation and review before it becomes an organological component label. |
| Modelgenerator | Photographic bundle and pinned DA3 configuration → reconstruction artifacts; the September snapshot reports point clouds and camera paths. | The October hub adapter requests DA3 GLB output and can retain camera data. Inspect the actual output: registration, surface coverage, scale and transport verification remain separate concerns. |
Example: a historical account becomes a candidate
Consider a supplied document naming a workshop, place and construction year. Intake preserves the source identity. Document evidenceEvidenceMaterial offered as grounds for a claim. Its relevance depends on the claim, context, transmission and interpretation.Work through the Cuntz example → identifies the relevant page and passage. Language detection and Litparser can assist interpretation or bibliography; Localizer can propose geographic matches. Normalizer organizes candidate values without discarding the literal account. PLENUM evaluates how suitable that account is for the particular proposition and whether other accounts depend on it.
The editorial question is then precise: does the passage support construction by this workshop for this instrument stateInstrument stateA historically scoped condition or arrangement of an instrument. A statement about one state should not automatically describe the instrument at every date.Work through the Cuntz example → and date? A page mentioning a workshop nearby, or a contract that did not lead to completion, requires a different relation. The accepted projection should retain the reviewed claim and its grounds. If the evidence is insufficient, preserving an unresolved candidate is more informative than forcing a clean but unsupported value.
Where does an extracted number become an accepted claim?
A processing candidate is a proposed interpretation of source material. Review and publication are additional decisions, with their own evidence and responsibilities.
- ProcessingCandidate resultAn algorithm produces an attributable output.review the candidate
- ReviewCheck the evidenceA reviewer can accept, reject or defer.publish the scoped result
- PublicationA qualified claimKeep the decision and lineage.
Work through the example 3 STEPS
Use the Cuntz workbook observation to imagine a new import workflow. This is an explanatory mapping to the Processor architecture, not a claim that the published Cuntz release was built by all current Processor modules.
Preserve the source
Intake identifies the workbook bytes and source context. Extraction retains the literal cell, property code and unit rather than keeping only a convenient normalized number.
Inspect the candidate
A normalizer can propose a structured observation and a pipe identity. A reviewer checks the cell interpretation, the identity match and qualifications such as the timestamp note.
Publish the accountable result
An accepted projection keeps the evidence link and decision. A public view can be concise while a fixed release preserves the record needed to re-examine the claim.
- Cuntz Positiv · VAO 0.5.0-rc.2 ↗
Published manifest: identities, measurement observations, realization digests, profiles and rights. Exact examples checked against this release.
- MODAVIS Ontology Network 0.1.0 ↗
Identity, state, configuration, evidence, assertions, context and provenance. Teaching examples explain its distinctions without inventing new normative properties.
The example illustrates component responsibilities without attributing unverified processing history to the actual release.
The useful distinction. Automation can propose a useful structure; an accountable workflow keeps the interpretation and its grounds inspectable.
Example: images and scores support different claims
Imageseg can isolate a visible region for inspection; Modelgenerator can estimate spatial structure from a photographic bundle. Neither establishes the identity or date of the object pictured. Ambitus can describe the nominal range of a score or MIDI file; it does not establish that a particular surviving instrument can produce every encoded pitch under its current configurationConfigurationAn identified arrangement of components and their relationships. Rebuilding or reassigning parts can change a configuration while the instrument’s broader identity persists.Work through the Cuntz example →. Connect each derived asset to its source and the domain question it actually addresses.
Use diagnostics to answer a bounded question
Technical detail · Portable Processor 0.2.0 snapshot · after installing its documented environment
modavis-processor modules --check-imports
modavis-processor exampleIn the documented development package, the first command inspects the module inventory and imports; the second runs an authored, deterministic PLENUM example without requiring a database, model or service. It reports the modavis.plenum.v1 policy with automatic promotion disabled. These checks help establish that a local installation and example behave as described. They do not evaluate extraction accuracy over a historical corpus or validate an external provider.
For a meaningful module evaluation, retain the diagnostic input, expected task, actual output and failure cases. Separate transport success from semantic correctness, model availability from useful output, and a small authored fixture from a representative benchmark. The module guide’s boundaries are especially useful here: they show where an automated result stops and a research judgment begins.