02 — MAP-GEO / MAP-M

Represent the page’s physical geometry without deciding article meaning or writing text.

NO TOOL Nothing implements this step

MAP-GEO cannot act as an independent witness to CAP. Any error CAP makes about where things sit on the page is inherited rather than caught. The V8 gate — every significant physical object accepted with evidence or marked MAP_AMBIGUOUS/HOLD — cannot currently be evaluated.

Why it is held: Two candidate detectors were evaluated and both rejected on evidence: Transkribus model 49272 (rejected for MAP-quality failure) and Eynollah (3 of 12 pages correct, 10 false bands on continuation pages). No replacement has been selected.

ML proposal-only Parallel page model 2 archived step 02 of 26

Where this output sits

Source evidence›Candidate evidence›Human decision›Audit acceptance›Release›Production

Source evidence, candidate evidence, human decision, audit acceptance, release and production are six separate things. This step produces the highlighted one and nothing further.

Scope

Intended to measure the page's physical geometry directly from the original pixels — regions, columns, gutters, frames, rules, boxes, table envelopes and media candidates — independently of what the text says. No implementation currently does this.

This role is allowed to

Propose physical class and geometry
Retain multiple candidates

How it works

fed by 00 INGEST

Inputs

Immutable original pixelsNative PDF geometry when present

ML proposal-only Pinned non-generative layout model plus deterministic line/component detector.

No internal sequence is documented for this step — only the contract above and below it. Nothing was invented to fill the gap.

Must pass

Every significant physical object represented or held
All geometry in page bounds with source lineage

passes →

Outputs

Regions, probable text blocks and containmentColumns, gutters, frames, rules and boxesTable envelopes and media candidatesConfidence and geometry HOLDs

hands off to MAP-AI · COMPLETENESS · STYLE · ROLE · PIC · TABLE-PHYSICAL · ORDER

fails →

HOLD

MAP_AMBIGUOUSMAP_OBJECT_MISSINGMAP_OUT_OF_BOUNDSMAP_UNSUPPORTED_GEOMETRY

Nothing continues on a failed gate. Uncertainty becomes an explicit HOLD, and no later step may read an unanswered item as an accepted one.

Non-scope — what this role does NOT own

Text transcription
Semantic category
Reading order acceptance
Article meaning

Explicitly forbidden

Transcribe text
Accept semantic category
Accept reading order
Generate unsupported coordinates

Dependencies and position

Starts: After INGEST; parallel with CAP-RAW.
Previous step: 01 CAP-RAW
Next step: 03 MAP-AI
Hands off to: MAP-AI, COMPLETENESS, STYLE, ROLE, PIC, TABLE-PHYSICAL, ORDER

Exact inputs

Immutable original pixels
Native PDF geometry when present

Exact outputs

Regions, probable text blocks and containment
Columns, gutters, frames, rules and boxes
Table envelopes and media candidates
Confidence and geometry HOLDs

Performer and AI/ML boundary

Performer: Pinned non-generative layout model plus deterministic line/component detector.
Class: ML proposal-only
AI boundary: Learned layout proposals are allowed.

Training information

Not trained by this project. A pinned general model is prompted under a strict response schema. It proposes; deterministic validation accepts or HOLDs. It cannot create text or coordinates.

Deterministic validation and acceptance gates

Every significant physical object represented or held
All geometry in page bounds with source lineage

HOLD and failure behaviour

Failure codes this step may emit:

MAP_AMBIGUOUS
MAP_OBJECT_MISSING
MAP_OUT_OF_BOUNDS
MAP_UNSUPPORTED_GEOMETRY

Uncertainty becomes an explicit HOLD. Omission never converts uncertainty into acceptance, and no downstream step may treat an unanswered item as an accepted one.

Downstream handoff

MAP-AI, COMPLETENESS, STYLE, ROLE, PIC, TABLE-PHYSICAL, ORDER

Active tool

NO ACTIVE TOOL — HOLD. No implementation was invented to fill this gap.

Actual performer today: In the live V114 chain, physical geometry is taken from Document AI CAP output rather than from an independent layout detector reading the original pixels.
Missing implementation: A pinned, non-generative layout detector that measures regions, columns, gutters, frames, rules, boxes, table envelopes and media candidates from the ORIGINAL PIXELS, independently of CAP.
Consequence: MAP-GEO cannot act as an independent witness to CAP. Any error CAP makes about where things sit on the page is inherited rather than caught. The V8 gate — every significant physical object accepted with evidence or marked MAP_AMBIGUOUS/HOLD — cannot currently be evaluated.
Why it is on HOLD: Two candidate detectors were evaluated and both rejected on evidence: Transkribus model 49272 (rejected for MAP-quality failure) and Eynollah (3 of 12 pages correct, 10 false bands on continuation pages). No replacement has been selected.

Recorded in tool/TOOL_STATUS.json.

Archived versions

Archived version	Why superseded	Retained value	Bytes
MARSOOM_REV6_PHASE1_TRANSKRIBUS_SIX_PAGE_TRIAL_2026-08-31	Rejected for production: MAP-quality failure	Negative evidence — records why this detector was not adopted	referenced
MARSOOM_REV6_PHASE1_UNIFIED_MAP_COMPARISON_2026-08-30	Comparison run, not a production tool	Side-by-side detector comparison evidence	referenced

2 archived version(s). Historical packages are never deleted or mutated; large ones are referenced with verified paths rather than copied, because the source archive is 11 GB.

Known limitations

No active tool. Geometry is currently inherited from CAP output rather than measured independently.
Two candidate detectors were tested and rejected on evidence: Transkribus model 49272 for MAP-quality failure, and Eynollah at 3 of 12 pages correct with 10 false bands on continuation pages.
Consequence: the system has no independent witness to CAP's view of the page.

Security and privacy

Page images and indexed references are sent to an external provider. Credentials live outside this repository. Model output is candidate evidence, never truth.

Cost behaviour

Metered per call by the provider. Token budget and tier are set per run and recorded in the run receipts.

Copying a tool into this repository does not authorise running it, retraining it, calling a model, or processing a new issue. No paid call may be made without the owner's explicit authorisation and a hard cost cap.

Provenance

Registry: registry/roles.json (schema marsoom.e2t.roles.v1)
Derivation: 28 legacy roles - VER - HUMAN + REV = 27 active steps
Legacy role id: MAP-GEO
Legacy source SHA-256: 6f5d5d7ed8869e45307424c2193f95ffab84abc44a6a188f5c7f98c2a48ec64d
Generated: 2026-09-04T22:13:12.433Z

Feedback and decisions

Feedback is recorded per source and never merged into an invented consensus. Where sources disagree, both positions stand and the owner decides.

Source	Events
Naser / owner Final authority. Overrides every other source.	none recorded
Codex / orchestrator Architecture and sequencing.	none recorded
Builder / designer Implementation reality and constraints.	none recorded
Independent reviewer Adversarial review of claims.	none recorded
Auditor Evidence verification against artifacts.	none recorded
Human REV reviewer Page-level truth from the review site.	none recorded

Recorded events

No feedback events recorded yet. The ledger exists and is append-only: feedback/FEEDBACK_LEDGER.jsonl.

Editing feedback is not possible: a change is a new event whose supersedes names the one it replaces, and the original stays exactly as written.

Generated from registry/roles.json by tools/build-reports.mjs. Do not hand-edit — edit the registry and rebuild.
Original page pixels are the visual authority. CAP owns the exact captured text.