From principle to verifiable reading

The method defines the rules, the pipeline organizes their execution

From the image of the source to a traceably justified reading.

The method page explains why observation, analysis, reading and interpretation remain separate. This page shows how that becomes a controlled workflow.

The pipeline does not replace looking at the source. It organizes that looking: through limited module roles, stored intermediate states, traceable handovers, checks, locks and human review steps.

Simplified workflow

A reading emerges only at the end of a verifiable path.

The following graphic is an orientation, not a rigid scheme. Depending on the source, steps may be repeated, skipped, branched or returned to review.

Light, low-text workflow graphic of the HistoriaMP pipeline with eleven numbered modules, evidence flow, audit points, uncertainty, review and export
Low-text, brand-neutral visualization of a HistoriaMP workflow. Module names and the legend are controlled in the HTML.
SourceImage qualityLayoutSegmentsGlyphsMinim clustersAbbreviationsReadingConsensusReviewExport
  • Gold: evidence flow and audit points
  • Blue: uncertainty
  • Red: review required
  • Grey: not released or blocked
Source→Image quality→Layout→Segment→Line→Glyph→Minim cluster→Abbreviation→Reading→Variants→Consensus→QC→Review→Export
01

Modules are roles

A module is not an omniscient instance. It works with a limited task and may only output what its level methodologically permits.

02

Artifacts instead of text flow

Documented findings move between modules: image areas, coordinates, segments, uncertainties, variants and checking status.

03

Not everything may continue

An artifact may remain stored without being reused as a reliable reading. Uncertainty can trigger review or block handover.

Controlled transitions

Modules take on limited tasks.

The pipeline orders analysis tasks into stages. An early step describes the visible finding, a layout module records zones and structures, segment and line modules organize crops and assignments. Glyph and minim modules mark critical sign areas; an abbreviation module prepares possible abbreviation findings. Transcription creates reading candidates only with a visual basis. QC checks them against previous artifacts.

The strength of the pipeline does not lie in allowing every module to do a lot, but in each module fulfilling its limited role in a controlled, checkable and traceable way.

Artifact chain

What is passed between the modules.

An artifact is not a final judgment. It is a verifiable intermediate state with reference to image, segment, coordinate, module origin, uncertainty status, checking status and handover status.

This makes it possible to see later which image area a finding came from, which module produced it, which uncertainty remained and whether it was confirmed, deferred, commented on or blocked.

Source to segment to uncertain finding to preliminary reading candidate.
Artifacts keep visible findings, segment references, status information and evidence points separate. The graphic is a schematic illustration, not a real source finding.
Reading proposal

From visible finding to checkable proposal.

In the workflow, a reading is not treated as an isolated text answer. It emerges from a documented image crop, detailed findings, variants and a visible checking status.

The presentation separates source, crop, detail enlargement and proposal. This keeps visible which level is image finding and which level is already a preliminary reading.

From visible finding to reading proposal: crop, detail finding and variants remain separately visible. The graphic is a schematic illustration, not a real source finding.
Orchestrator

The orchestrator is not a chief model. It is a guard.

The orchestrator does not decide what is written in the source. It checks transitions: Has a module respected its role? Is the image basis missing? Has uncertainty disappeared? Is there an artifact reference? May a result be handed on?

If rules are violated, it blocks automatic handover or routes the finding into review. It does not ask: Which answer sounds best? It asks: May this answer continue on the basis of the previous artifacts?

Orchestrator as control instance between modules, review, blocking and handover.
Errors become rules

Errors reveal missing control layers.

A documented error can trigger a pipeline rule, a lock, a review step or an additional checking module. If a model reads too early, a stricter observation layer is needed. If uncertainty disappears, it must be carried forward as an artifact. If a minim cluster becomes a smooth word, it needs its own checking layer.

If a coordinate cannot be traced back, the finding remains technically restricted. If a special sign is ignored, a glyph or abbreviation check may become necessary.

Documented errors become rules, locks or review steps.
Uncertain problem areas remain visible in the workflow.
Consensus

Model majority is not source evidence.

Comparing several models can provide indications, but it does not replace a return to the source. Three models can overestimate the same weak finding, prefer the same plausible addition or smooth away the same uncertainty.

Consensus becomes reliable only when a reading can be traced back to shared visible basis, identifiable artifacts, documented uncertainty and a traceable checking decision.

Review is not an afterword

Human checking can take place at several points: after layout, segmentation, glyph finding, reading candidate or quality control. Review can confirm, comment on, defer or block artifacts.

No rigid pipeline

The standard pipeline is a starting point. A source may require different analysis paths; the workflow can branch, repeat modules or jump back to earlier stages.

Example analysis paths

Not every source needs the same pipeline.

HistoriaMP is designed to organize workflows depending on the source. Example analysis paths may be:

Medieval manuscript

Image quality → Layout → Segment → Glyph → Minim cluster → Abbreviation → Reading → QC → Review

Archive document

Image quality → Layout → Writing areas → Stamps / signatures → Transcription → Review → Export

Comparison project

Image → Layout → Reading variants → Source comparison → Commentary → Export

Result package

At the end there is not only text.

At the end stands an analysis package: original image, image versions, segments, coordinates, module findings, variants, uncertainties, review comments, quality report and export.

The pipeline does not only make a reading visible. It makes visible how that reading came into being.

View methodological foundationsView limits and open problems