Authority — the rules everything else obeys

• Original page pixels are the visual authority. Original page pixels.
• CAP owns the exact captured text. No other role may silently rewrite it.
• Uncertainty becomes an explicit HOLD. Omission never converts uncertainty into acceptance.

Six things that are not the same thing

A candidate is not human truth. Human truth is not audit acceptance. Audit acceptance is not release. Release is not production. Nothing in this system may skip a rung.

Readiness describes whether an implementation exists, not whether its output is accepted. No step in this system has produced released or production content.

Invariants this graph enforces

• CAP-RAW and MAP-GEO both start from the frozen original page, in parallel.
• SWEEP is the first reconstruction operation, but it cannot start until CAP polygons and the MAP/PIC/TABLE/STYLE protection evidence are frozen.
• SWEEP has no feedback edge into any evidence-producing step. It is a renderer substrate derivative only.
• Both instrument and instrument_candidate enter SPELL and REV.
• Corrections produced by REV are stored separately from CAP-RAW.

The source is right-to-left Arabic. This sample is here so a rendering fault in this portal is visible immediately rather than discovered later on real page text.

هذه عينة نصية بالعربية للتحقق من اتجاه الكتابة من اليمين إلى اليسار، ومن تشكيل الحروف المتصلة.

Expected: reads right to left, letters joined, the definite article ال attached to its word and not transposed.

Registry

registry/roles.json

Derivation

28 legacy roles - VER - HUMAN + REV = 27 active steps

Legacy source SHA-256

6f5d5d7ed8869e45307424c2193f95ffab84abc44a6a188f5c7f98c2a48ec64d

Generated

2026-09-04T22:13:12.433Z

The previous single-file report was preserved before replacement at portal/archive/E2T_SYSTEM_V1_2026-09-04_ORIGINAL.html, hash recorded.

How it works

Performer and AI boundary

Performer

Deterministic ingest/orchestration tool.

AI

None.

Starts

Immediately after source delivery.

Hands off to

CAP, MAP-GEO, All provenance envelopes

Training

Not a trained component. No model, no weights, no training corpus. Behaviour is fixed by code and configuration, and identical inputs must produce identical outputs.

Active tool

Invocation

python tools/seal_package.py <package_dir>

Input

package directory tree

Output

PACKAGE_SHA256.txt

Dependencies

Python 3.12 (NOT installed on this machine)

Why this version

seal_package.py is the mechanism actually used to freeze and hash-bind every package from V47 through V118, and PACKAGE_SHA256.txt appears in every version folder. It provides the freeze half of this role. It does NOT assign issue_id/page_id/page order/dimensions/colour profile from a source manifest, which is the other half of the V8 contract. Marked PARTIAL for that reason.

Tool receipt

1 file(s) copied, all hash-verified

Known limitations

• The freeze half of this role is implemented; the identifier-and-dimension half is not.
• Page order and colour profile are not independently asserted by any tool — they are inherited from the source manifest.
• Because the sealing tool is Python and Python is not installed on this machine, the step cannot be executed here.

Archived versions (1)

How it works

Performer and AI boundary

Performer

Pinned Google Document AI OCR.

AI

OCR recognition is allowed; generative rewriting is not.

Starts

After INGEST; parallel with MAP-GEO.

Hands off to

COMPLETENESS, ROLE/TITLE, TYPE, ORDER, TABLE-ASSIGN, CRITICAL

Training

Not trained by this project. Google Document AI processor `pretrained-ocr-v2.1-2024-08-07`, pinned by version. Its training data, accuracy characteristics and failure modes are the provider's and are not disclosed to us.

Active tool

Invocation

python tools/acquire_all_docai_cap_v33.py — calls the Cloudflare Worker marsoom-rev6-phase1-cf-trial-003, which calls Google Document AI

Input

frozen page images + issue manifest

Output

raw Document AI response JSON + request_manifest.json per page

Dependencies

Python 3.12 (NOT installed on this machine) · Cloudflare Worker marsoom-rev6-phase1-cf-trial-003 · Google Document AI processor pretrained-ocr-v2.1-2024-08-07 · R2 frozen-raw storage

Why this version

These four tools are the CAP acquisition path actually used to produce the sealed issue-4547 capture inside the SUCCESS_001 trial (1,416 verified files, 68 pages). The v33 suffix is the highest CAP tool generation present anywhere. The remote-raw runner is the one that freezes provider bytes before parsing, which is the V8 gate.

Tool receipt

4 file(s) copied, all hash-verified

Known limitations

• Arabic OCR error modes are inherited wholesale and cannot be corrected at this step by design.
• Direct-pixel auditing found cases where CAP text does not match what the page actually prints — 25 of 160 audited candidates in the 2026-09-04 audit were CAP/pixel mismatches.
• No independent witness currently checks CAP, because MAP-GEO has no active tool.
• Recall against visible ink is unmeasured: the completeness ledger that would measure it is only partly implemented.

Archived versions (0)

No superseded versions recorded.

How it works

Performer and AI boundary

Performer

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

AI

Learned layout proposals are allowed.

Starts

After INGEST; parallel with CAP-RAW.

Hands off to

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

Training

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.

Active tool

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.

Archived versions (2)

How it works

Performer and AI boundary

Performer

Constrained structured vision AI with deterministic schema validation.

AI

Yes, proposal-only.

Starts

After MAP-GEO and CAP aliases are frozen.

Hands off to

MAP-REFINE, ROLE/TITLE, ORDER, TAG

Training

Not trained by this project. Gemini 3.8 Flash, pinned, prompted under response schema v65. It is constrained to proposal-only: it may reference existing ids and propose merges, splits, roles and relations, but may not return free OCR text or invent coordinates.

Active tool

Invocation

python tools/run_autonomous_parallel.py --issue <id>

Input

schemas/planner-response.schema.json + indexed CAP/MAP references

Output

schemas/detail-response.schema.json (compact schema v65)

Dependencies

Python 3.12 (NOT installed on this machine) · Gemini 3.8 Flash via API · Frozen CAP aliases and MAP-GEO region index

Why this version

V114 is the highest-numbered package in the Flash MAP pipeline chain (V47 to V50-V100 to V108-V114) that contains executable code, and each README explicitly supersedes the prior: V111 added resume, V112 fixed the token budget, V113 added tier amendment, V114 fixed candidate/feedback pairing. V114 is also the package captured as 02_SYSTEM_SNAPSHOT inside the sealed SUCCESS_001 issue-4547 trial — independent confirmation that this is the version which produced the live output.

Tool receipt

12 file(s) copied, all hash-verified

Known limitations

• A general-purpose vision model is doing structural work; its failure modes are not fully characterised on this corpus.
• Duplicate-ownership collisions occurred often enough that a dedicated repair loop (step 04) exists to resolve them.
• Proven on one issue (4547, 68 pages). Generalisation to other eras and layouts is not established.
• The trial that succeeded is explicitly marked HOLD and candidate-only — not accepted, not released.

Archived versions (2)

How it works

Performer and AI boundary

Performer

Deterministic geometry/component tool.

AI

None.

Starts

After MAP-AI requests or deterministic MAP conflicts.

Hands off to

COMPLETENESS, ROLE, PIC, TABLE-PHYSICAL, ORDER

Training

Not a trained component. No model, no weights, no training corpus. Behaviour is fixed by code and configuration, and identical inputs must produce identical outputs.

Active tool

Invocation

python tools/repair_selected_ranges.py --ranges <list>

Input

accepted refinement request + duplicate report

Output

revised measured regions with parent/child lineage

Dependencies

Python 3.12 (NOT installed on this machine) · Frozen MAP-AI output · Duplicate-alias report

Why this version

The V114 README states this generation mechanically binds each rejected parsed candidate to a freshly recomputed duplicate report — the fix that closes the repair loop. Latest of the V105, V108, V111, V112, V113, V114 repair chain.

Tool receipt

4 file(s) copied, all hash-verified

Known limitations

• Refinement quality is bounded by MAP-AI proposal quality, and by the absence of an independent MAP-GEO measurement to check against.
• The bound on repair cycles is a safety limit, not a correctness guarantee: a range can exhaust its cycles and still be ambiguous.

Archived versions (0)

No superseded versions recorded.

How it works

Performer and AI boundary

Performer

Deterministic component matcher.

AI

None for production evidence.

Starts

After CAP-RAW and current MAP geometry are frozen.

Hands off to

PROTECTION, ASSEMBLE, QA

Training

Not a trained component. No model, no weights, no training corpus. Behaviour is fixed by code and configuration, and identical inputs must produce identical outputs.

Active tool

Invocation

inspect READY.json gates: dangling_references, overlap_conflicts, missing_ownership

Input

review_data.json

Output

READY.json gate block

Dependencies

Built by BUILDER_STAGE_A.py / BUILDER_STAGE_B.py — see step 18

Why this version

Completeness currently exists only as gates inside the review-package builder (dangling_references, overlap_conflicts and missing_ownership all required to be zero), not as a standalone tool. Those gates account for CAP-anchor ownership but NOT for source-ink components that have no CAP token at all. The V8 dispositions CAP_TOKEN, STRUCTURE, MEDIA, ORNAMENT, BACKGROUND_NOISE and HOLD are not produced by any implementation.

Tool receipt

1 file(s) copied, all hash-verified

Known limitations

• Only partly implemented. The existing gates check anchor ownership, dangling references and overlap conflicts within CAP output.
• They do NOT detect visible text-like ink that CAP never captured, which is precisely the failure this role exists to catch.
• 'Every CAP token exactly once' proves the provider's output was preserved. It does not prove every printed word was captured. That distinction is the whole point of this step and it is currently unenforced.

Archived versions (0)

No superseded versions recorded.

How it works

Performer and AI boundary

Performer

Deterministic pixel sampler.

AI

None.

Starts

After physical regions and exclusion masks exist.

Hands off to

TYPE, PROTECTION, SWEEP, PDF

Training

Not a trained component. No model, no weights, no training corpus. Behaviour is fixed by code and configuration, and identical inputs must produce identical outputs.

Active tool

Invocation

python tools/mechanical_pipeline.py

Input

original pixels + accepted regions + exclusion masks

Output

renderer.config.json palette block

Dependencies

Python 3.12 (NOT installed on this machine) · OpenCV · Original page pixels

Why this version

The mechanical pipeline performs the paper and region fill sampling described in the V4 README step 3. It is the only implementation that measures colour from pixels rather than assuming it. It does not emit the full V8 STYLE contract — sample denominators, per-palette confidence and sampling lineage are missing — so it is PARTIAL.

Tool receipt

2 file(s) copied, all hash-verified

Known limitations

• Sampling exists, but the contract's evidence fields do not: sample denominators, per-palette confidence and sampling lineage are not emitted.
• Without denominators, a colour measured from three pixels is indistinguishable from one measured from three thousand.

Archived versions (0)

No superseded versions recorded.

How it works

Performer and AI boundary

Performer

Pinned layout classifier plus deterministic evidence gates.

AI

Constrained proposal is allowed.

Starts

After geometry is frozen; CAP geometry may assist.

Hands off to

TITLE, TYPE, ORDER, TAG, PROTECTION

Training

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.

Active tool

Invocation

structural role is emitted by the MAP-AI detail call: field `f` = H|F|R|X (header, footer, rule, other furniture); field `x` = table boolean

Input

MAP geometry + CAP geometry

Output

detail-response fields `f` and `x`

Dependencies

Step 03 MAP-AI

Why this version

ROLE has no separate binary. It is realised as typed fields inside the MAP-AI detail response schema v65 and validated deterministically against that schema. Recorded ACTIVE because the contract is enforced and the output exists — but the performer is step 03, not an independent classifier, and the report says so.

Tool receipt

2 file(s) copied, all hash-verified

Known limitations

• Realised as typed fields inside the MAP-AI response rather than as an independent classifier, so it inherits MAP-AI's failure modes.
• The furniture vocabulary is coarse: header, footer, rule and a catch-all 'other'.

Archived versions (0)

No superseded versions recorded.

How it works

Performer and AI boundary

Performer

Structured AI proposal with deterministic exact-CAP and ownership validation.

AI

May share the MAP-AI/ORDER/TAG request.

Starts

After ROLE evidence and ordered candidate units exist.

Hands off to

ORDER, TAG, DGTL, ARCH

Training

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.

Active tool

Invocation

title binding is emitted by the MAP-AI detail call: field `t` with `k` = P (printed) or A (AI-proposed), and segment `m`

Input

ROLE candidates + ordered CAP line aliases

Output

detail-response field `t`

Dependencies

Step 03 MAP-AI · Exact CAP line IDs

Why this version

The schema distinguishes a PRINTED title from an AI-PROPOSED one via the `k` discriminator. That distinction is exactly what the owner rule requires: only exact printed PRIMARY_TITLE text may enter the TOC as a source title. A derived label for a titleless unit is carried separately and must never be presented as source text.

Tool receipt

1 file(s) copied, all hash-verified

Known limitations

• Only exact printed PRIMARY_TITLE text may enter the table of contents as a source title. A derived label is not source text and must be visibly marked as derived wherever it appears.
• Subtitle contamination is a known, named failure mode with no independent detector.

Archived versions (0)

No superseded versions recorded.

How it works

Performer and AI boundary

Performer

Deterministic typography and fitting solver.

AI

None.

Starts

After CAP geometry and ROLE/TITLE are available.

Hands off to

PDF, QA

Training

Not a trained component. No model, no weights, no training corpus. Behaviour is fixed by code and configuration, and identical inputs must produce identical outputs.

Active tool

Invocation

node tools/build_v7_runtime.js, then tools/audit_v7.py for the fit and collision gate

Input

CAP line/token geometry + ROLE/TITLE + font inventory

Output

renderer runtime data (data/renderer_data.js)

Dependencies

Node.js (present on this machine) · Python 3.12 for the audit step (NOT installed) · Bundled font inventory

Why this version

V7 is an explicit append-only successor to V6 Typography Alignment, which succeeds V5 and V4. It is the highest typography generation containing a working runtime builder. Note that build_v7_runtime.js is JavaScript and therefore runnable on this machine; the audit gate is Python and is not.

Tool receipt

4 file(s) copied, all hash-verified

Known limitations

• The runtime builder is JavaScript and runs here; the fit-and-collision audit gate is Python and cannot run on this machine.
• Font matching is approximate — the original newspaper faces are not available, so a closest supported bundled font is used and the approximation is recorded.

Archived versions (1)

How it works

Performer and AI boundary

Performer

Deterministic component/energy detector; layout ML may nominate only.

AI

No generative image operation.

Starts

After MAP media candidates and CAP masks.

Hands off to

PROTECTION, ORDER, PDF, DGTL

Training

Not a trained component. No model, no weights, no training corpus. Behaviour is fixed by code and configuration, and identical inputs must produce identical outputs.

Active tool

Known limitations

• No active tool. There is no bounded media evidence for any page.
• Caption-to-picture and picture-to-article links therefore cannot exist, which also caps what step 13 can deliver.
• The reconstruction path has no verified media layer.

Archived versions (0)

No superseded versions recorded.

How it works

Performer and AI boundary

Performer

Deterministic grid/rule detector; table-layout ML may nominate only.

AI

Proposal-only if used.

Starts

After MAP table envelopes.

Hands off to

TABLE-ASSIGN, PROTECTION, PDF, DGTL

Training

Not a trained component. No model, no weights, no training corpus. Behaviour is fixed by code and configuration, and identical inputs must produce identical outputs.

Active tool

Known limitations

• No active tool. The only table signal in the system is a boolean marking a group as table or not-table.
• A boolean is not geometry: there are no cells, so no table can be faithfully rebuilt and the cell-level QA gate cannot be evaluated.

Archived versions (0)

No superseded versions recorded.

How it works

Performer and AI boundary

Performer

Deterministic geometry assignment.

AI

None.

Starts

After TABLE-PHYSICAL and CAP-RAW.

Hands off to

ORDER, CRITICAL, PDF, DGTL

Training

Not a trained component. No model, no weights, no training corpus. Behaviour is fixed by code and configuration, and identical inputs must produce identical outputs.

Active tool

Known limitations

• No active tool, and blocked upstream: there are no cells to assign tokens to while step 11 is unimplemented.

Archived versions (0)

No superseded versions recorded.

How it works

Performer and AI boundary

Performer

Constrained structured AI may propose; deterministic graph validator accepts.

AI

May share the MAP-AI/TITLE/TAG request.

Starts

After CAP and page-model roles are available.

Hands off to

ISSUE-LINK, TAG, CRITICAL, ASSEMBLE

Training

Not a trained component. No model, no weights, no training corpus. Behaviour is fixed by code and configuration, and identical inputs must produce identical outputs.

Active tool

Invocation

sequence emitted by MAP-AI as field `s` (array of p/a/v); unique ownership verified by analyze_duplicate_aliases.py

Input

frozen MAP/ROLE/CAP

Output

detail-response field `s` + duplicate-ownership report

Dependencies

Step 03 MAP-AI · Python 3.12 for the ownership analyser (NOT installed)

Why this version

Unique-anchor ownership is genuinely proven — the SUCCESS_001 trial recorded 13,815 source anchors each owned exactly once, with zero missing, zero duplicate and zero overlap conflicts. That is the every-token-exactly-once half of the contract. The typed link set the contract demands (title to article, body paragraphs to article, picture to article, caption to picture) is only partly present, and picture/caption links cannot exist at all while step 10 has no tool. Hence PARTIAL.

Tool receipt

2 file(s) copied, all hash-verified

Known limitations

• Unique ownership is genuinely proven — 13,815 anchors each owned exactly once, zero duplicates, zero orphans, zero overlap conflicts on issue 4547.
• The typed link set is only partly present. Picture and caption links cannot exist while step 10 has no tool.
• Multi-column reading order was previously found unsafe enough that mechanically ordered OCR was withdrawn from presentation. Treat ordering as candidate evidence.

Archived versions (0)

No superseded versions recorded.

How it works

Performer and AI boundary

Performer

Deterministic graph adapter; structured AI may propose edges.

AI

Proposal-only.

Starts

After all PAGE-ORDER graphs for the issue.

Hands off to

TAG, DGTL, ARCH, ASSEMBLE

Training

Not a trained component. No model, no weights, no training corpus. Behaviour is fixed by code and configuration, and identical inputs must produce identical outputs.

Active tool

Invocation

continuation emitted by MAP-AI as field `b`: C = continuation, O = open, U = unbound, with `lg`/`rg` neighbour group ids

Input

accepted page graphs + issue manifest

Output

detail-response field `b`

Dependencies

Step 03 MAP-AI · All accepted PAGE-ORDER graphs

Why this version

Cross-page continuation is a first-class typed field with an explicit unbound state, and it was exercised as a dedicated run (V96 FLEX_MAP_CONTINUATION) inside the SUCCESS_001 lineage. The issue-4547 review package preserved 9 cross-page groups through V001 to V003 — direct evidence the mechanism works end to end.

Tool receipt

1 file(s) copied, all hash-verified

Known limitations

• Proven on one issue: 9 cross-page groups preserved through three package generations.
• Continuation is proposed by the same model that proposes structure, so a systematic misreading of continuation cues would not be independently caught.

Archived versions (1)

How it works

Performer and AI boundary

Performer

Rules/lexicon plus small supervised classifier or constrained structured AI.

AI

Allowed within a pinned, validated label set.

Starts

After complete PAGE-ORDER/ISSUE-LINK units and TITLE roles.

Hands off to

CRITICAL, SPELL, DGTL, ARCH

Training

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.

Active tool

Invocation

categories applied by the MAP-AI detail call against the closed ontology set

Input

complete linked units + ROLE + exact CAP text

Output

category per group with confidence

Dependencies

Step 03 MAP-AI

Why this version

ontology.json holds the closed category set actually enforced: NEWS, GOVERNMENT_INSTRUMENT, GOVERNMENT_ANNOUNCEMENT, COMPANY_ANNOUNCEMENT, TENDER, PAID_ADVERTISEMENT, TABLE, REFERENCE, HEADER, FOOTER, FURNITURE, NEEDS_REVIEW, OTHER, UNKNOWN. NEEDS_REVIEW and UNKNOWN provide the explicit uncertainty the contract requires.

Tool receipt

2 file(s) copied, all hash-verified

Known limitations

• The classifier is the same constrained vision model, not a separately trained text classifier as the original design intended.
• Instrument recall is not measured against an adjudicated gold set, so the recall-first requirement is asserted rather than proven.
• A previous generation produced a classifier false positive and two mislabelled issues that only direct image review caught.

Archived versions (0)

No superseded versions recorded.

How it works

Performer and AI boundary

Performer

Deterministic high-recall rules; optional pinned nomination classifier.

AI

Optional nomination only.

Starts

After TAG and complete unit ownership.

Hands off to

SPELL, REV, EXTRACT

Training

Not a trained component. No model, no weights, no training corpus. Behaviour is fixed by code and configuration, and identical inputs must produce identical outputs.

Active tool

Invocation

nomination is performed inside BUILDER_STAGE_A.py (step 18); gates are recorded in READY.json

Input

linked units + TAG results

Output

critical candidate ledger with state

Dependencies

Step 15 TAG · Step 18 builder

Why this version

Date nomination exists and is deliberately scoped: READY.json records date_scope_instrument_only true with instrument_date_count 11. That is real, bounded critical nomination. But the V8 contract also covers numbers, identifiers and instrument numbers with bbox/polygon and crop-hash binding, and the 100-percent-recall-or-HOLD gate is not evaluated anywhere. An earlier generation produced 1,915 unfiltered date detections which the trial itself marked INVALID_ALL_1915_DETECTIONS_WERE_NOT_REVIEW_ELIGIBLE.

Tool receipt

2 file(s) copied, all hash-verified

Known limitations

• Currently scoped to instrument dates only: 11 nominated on issue 4547.
• Numbers, identifiers and instrument numbers are not nominated by any implementation.
• The 100-percent-recall-or-HOLD gate is not evaluated anywhere.
• An earlier generation emitted 1,915 unfiltered date detections, which the trial itself recorded as not review-eligible. Quantity is not nomination.

Archived versions (0)

No superseded versions recorded.

How it works

Detection contract

Eligibility routing

Both instrument AND instrument_candidate units enter SPELL. Routing only confidently-tagged instruments would let a single TAG false negative bypass detection entirely, so recall is prioritised at the TAG boundary and both classes are carried through.

What every flag must carry

The exact candidate value byte-for-byte; character offsets within its unit; the full surrounding context; and linkage to page, group, unit, token, bbox and source crop. A flag that cannot bind to its geometry is SPELL_SPAN_UNBOUND and is not emitted as a finding.

Candidate reason and provenance

Anomaly type, model score, the model identity and revision, and the frozen threshold that admitted it. SPELL_PROVENANCE_MISSING if any of these is absent.

Four things kept separate

Detected inventory — everything the model scored. Review nominations — the subset that survived the stage-2 exact-glyph verifier and is worth a human's time. Human truth — what a reviewer actually decided, which lives in step 18 and nowhere else. Corrections — stored in a separate adjudication layer, never written back over CAP-RAW. Collapsing any two of these would let a machine guess be read as a fact.

Suppression is not correction

Stage 2 may only remove a stage-1 candidate from the nomination list. It cannot create a candidate, edit a value, or assert that a spelling is right — only that this project has an exact-glyph rule saying the printed form is admissible.

Zero output is not success

Zero flags on a substantial eligible corpus is NOT_MEASURED or FAILED_COVERAGE until coverage is independently proven. A prior review-material generation produced zero spelling flags and the trial correctly recorded that as invalid rather than clean.

Model licence — resolved 2026-09-05

The sealed package contains no licence file, which was carried as a HOLD. Checked at source: the base model CAMeL-Lab/camelbert-msa-qalb14-ged-13 is published under the MIT licence — permissive, and it permits the use and redistribution this project needs. The authors ask that work using it cite *Alhafni et al. (2023), "Advancements in Arabic Grammatical Error Detection and Correction: An Empirical Investigation"*. Note their own usage caveat: the model was fine-tuned on morphologically preprocessed text, which is worth remembering when applying it to raw gazette text. HOLD lifted; the obligation is citation, not restriction.

Retraining authority

Only the owner may authorise a retrain. Nothing in this repository does so, and the last attempt (v2) was rejected on evidence — its weights are deliberately excluded so it cannot be run by accident. Five of seven priority error families still have zero training positives, so a retrain today would have nothing new to learn from.

Performer and AI boundary

Performer

Pinned project-trained spelling/anomaly model plus exact-form verifier.

AI

Model detection is allowed; correction authority is not.

Starts

After TAG; for instrument and instrument_candidate units.

Hands off to

REV

Training

Trained by this project, once. Stage 1 is a `BertForTokenClassification` head warm-started from `CAMeL-Lab/camelbert-msa-qalb14-ged-13` at revision `447179dc63d186e4bff09a993e90e73ad622d571`, fine-tuned for 3 epochs in 50.9 minutes on CPU, seed 20260903, at a external spend of USD 0.00 with zero provider calls. Training data was clean Saudi .gov.sa text with synthetic corruptions as positives: 361 pages, 2,341 paragraphs, 60,500 words, split by complete source so no publisher appears in two splits — TRAIN 1,137 paragraphs / 845 positives, DEV 460 / 412, TEST 744 / 522. The decision threshold 0.93 was selected on DEV against an F0.5 objective and frozen at 2026-09-03T22:24:28Z before the test split was opened. Stage 2 is not learned: it is a registry of 55 typed exact-glyph entries, version v3-registry-1.0.0.

Active tool

Invocation

python infer_spell_candidates.py --input <passages.jsonl> --threshold 0.93

Input

passages with exact CAP tokens and offsets

Output

bounded candidate spans with offsets, confidence, model identity, status candidate-not-truth

Dependencies

Python 3.12 (NOT installed on this machine) · transformers 5.16.1 · model.safetensors 433,992,496 bytes — carried via Git LFS

Why this version

This is the only genuinely trained project tool in the system, and it is sealed. Stage 1 is a BertForTokenClassification head warm-started from CAMeL-Lab/camelbert-msa-qalb14-ged-13 revision 447179dc63d186e4bff09a993e90e73ad622d571 and fine-tuned once in-project. Stage 2 is a deterministic exact-glyph verifier with 55 typed registry entries that can only SUPPRESS a stage-1 candidate, never create or edit text. Model v2 was evaluated and REJECTED; its weights are deliberately excluded so it cannot be run.

Tool receipt

15 file(s) copied, all hash-verified

Known limitations

• No human ever adjudicated a real printed spelling error anywhere in this package. Every headline metric rests on SYNTHETIC positives.
• Recall on real gazette text is NOT MEASURED. Only precision-style counts exist there.
• On real pages the observed hit rate is 6 to 9 percent. Inspection of issue 4488 found roughly 91 percent of flags were extraction defects, not printed errors, and that issue was closed permanently.
• The 2026-09-04 direct-pixel audit of all 160 V004 candidates found 54 clear printed errors (33.8 percent), 75 valid or accepted forms, 25 CAP/pixel mismatches and 6 uncertain. Verdict: HOLD as a list of real spelling errors.
• Five of seven priority error families have zero training positives. Nothing is ready to retrain.
• Model v2 was REJECTED: it suppressed 73 of 75 known false positives but lost 29 of 54 real printed errors, 17 of them final-ha cases, through hard-negative over-generalisation. Its weights are deliberately excluded so it cannot be run.
• No licence file was found anywhere in the package. The base model's licence terms are therefore unconfirmed — HOLD.
• Zero output on a substantial eligible corpus is NOT_MEASURED or FAILED_COVERAGE, never a pass. A prior review-material generation produced zero spelling flags and the trial correctly recorded that as invalid, not as clean.

Archived versions (3)

How it works

The human-review website

https://review.marsoom.app  verified reachable

This is the bare root deliberately: the review client has no URL routing of any kind, so a deep path such as /assignment/<id> would 404 against its static asset map.

That site belongs to a different session. Nothing in this repository modifies, redeploys, or writes to it — the relationship is read-only in one direction.

Human review contract

Who may review

An authorised human only. The review site is private and owner-only, and reviewer sign-up is closed server-side — it is gated on an access-policy row that does not exist, and a missing row reads as closed. Anyone else following the link reaches a sign-in they cannot pass.

Reviewer instructions and calibration

Defined by the review contract, not by this repository. The site deliberately withholds the model's confidence from the reviewer until after they answer, so the machine cannot lead the witness.

Question types

Seven closed dimensions: grouping, label, title, subtitle, spelling, date, furniture. Each answer is addressed to an item id of the form ITEM:<dimension>:<id>. Free-form commentary is not a review answer.

What the reviewer is shown

The original page or crop — unaltered source pixels — beside the exact captured value and its surrounding context. The printed page reaches the reviewer unmodified; nothing rendered or reconstructed is substituted for it.

Allowed outcomes

CONFIRMED, CORRECTED, UNREADABLE, DEFERRED, CONFLICT. The export additionally carries cannot_determine and an excluded_by code (E1–E6: cannot-determine, illegible, owner void, integrity failure, subtitle true-negative, overlap duplicate) so an exclusion is always attributable to a stated reason rather than a silent drop.

Reviewer identity policy

assignment.reviewer_ref is an opaque id and, by contract, never a username. Reviewer identity must not appear in this repository, in the portal, or in any public artifact.

Timestamp policy

Every answer records answered_at, and an approved correction records approved_at, in UTC ISO-8601. Timestamps are written once at the moment of the decision. There is no update path that rewrites them, so the order of events stays reconstructable.

Append-only answers

One immutable official submission per assignment; it is never overwritten. A practice run is a separate mode and is marked as such, so practice can never be mistaken for an official result.

Conflict and adjudication

Where reviewers disagree, or a reviewer marks CONFLICT, the item goes to adjudication rather than to a majority vote. Both positions are retained. A correction carries status of proposed, approved or rejected, and only approved may reach designer feedback or training.

Escalation

UNREADABLE and DEFERRED escalate to the owner rather than resolving themselves. An unanswered or unreadable item is never converted into an accept by omission, and no downstream step may treat it as one. If the original evidence itself is insufficient, the escalation is against the source, not the reviewer.

Corrections stay separate from CAP-RAW

A corrected value is written to a separate adjudication layer keyed to the candidate. CAP-RAW is never edited. The captured text and the human's correction remain two distinct, independently readable facts.

Import contract — the identifier trap

marsoom.human-review-import.v47.profile-review-site-1 is the schema document's $id. The payload's schema field must read exactly marsoom.human-review-import.v47 — bare, with no profile suffix. The import gate compares that string and rejects the entire file at G1.1 if it differs; nothing is loaded, not even partially. The payload must also carry presentation_status: "HOLD" (gate G1.2). Two identifiers, two different jobs. (Confirmed by the VER V1 session, 2026-09-04.)

Export contract

marsoom.review-answers-export.v1 — here the $id and the emitted schema field are the same string. Owner-only, at GET /api/owner/export. Not public, so this portal cannot read it.

What human review does NOT authorise

Nothing automatic. A human decision is not audit acceptance, not release, and not production, and it does not by itself authorise training on the corrected data. Those are four further gates, each requiring its own explicit approval. spelling_metric.recall is a hard constant not_measured in the export contract — no code path may compute or display one.

Performer and AI boundary

Performer

Authorized human owner/reviewer.

AI

None as human authority.

Starts

When a role emits a reviewable HOLD or when acceptance sampling is required.

Hands off to

Adjudicated truth layer, Separate acceptance/release policy

Training

Not a model. Reviewer instructions, calibration and the closed question set are defined by the review contract. The review site deliberately withholds model confidence from the reviewer before they answer, so the machine cannot lead the witness.

Active tool

Invocation

python BUILDER_STAGE_A.py then BUILDER_STAGE_B.py to build the immutable review package; the package is then imported into the review website

Input

marsoom.human-review-import.v47 — see contract_identifier_warning

Output

marsoom.review-answers-export.v1

Dependencies

Python 3.12 (NOT installed on this machine) · review.marsoom.app — a separate system, read-only to this repository

Why this version

V005 is the highest version under review_packages/issue_4547 and was built 2026-09-04T06:34Z, AFTER the V004 pixel audit found V004's confidence heuristic yielded only 33.8 percent real errors. V005 replaces that heuristic with the trained SPELL detector plus SUPPRESSED_CANDIDATES.json and VERIFIER_REGISTRY_V3.json. The two contract schemas are the canonical copies placed by the review-site owner in the Human review Tool folder.

Tool receipt

9 file(s) copied, all hash-verified

Known limitations

• Access is private and owner-only. Reviewer sign-up is closed server-side, so anyone else following the link reaches a sign-in they cannot pass.
• The site runs under presentation status HOLD and refuses to start outside it. Nothing from it may be presented as an accepted result or a measured accuracy figure.
• No run has been submitted and no score exists.
• Spelling recall is a hard constant `not_measured` in the export contract — no code path may compute or display one.
• Human review does NOT by itself authorise training, audit acceptance, release or production. Those are four further, separate gates.

Archived versions (3)

How it works

Performer and AI boundary

Performer

Deterministic schema extractor; specialized model only under a separate approved contract.

AI

Not required in V1.

Starts

After instrument boundaries are complete and critical fields are resolved.

Hands off to

Phase 2 candidate environment, Audit and human review

Training

Not a trained component. No model, no weights, no training corpus. Behaviour is fixed by code and configuration, and identical inputs must produce identical outputs.

Active tool

Known limitations

• No active tool, and that is currently correct rather than a defect: no critical value has passed human review, so nothing is eligible for release.
• Building an extractor before REV produces adjudicated truth would create a path for unverified values to reach the database.
• Phase 2 extraction is on HOLD by owner decision.

Archived versions (0)

No superseded versions recorded.

How it works

Performer and AI boundary

Performer

Deterministic mask compiler.

AI

None.

Starts

After COMPLETENESS, STYLE, ROLE, PIC and TABLE evidence.

Hands off to

SWEEP, QA

Training

Not a trained component. No model, no weights, no training corpus. Behaviour is fixed by code and configuration, and identical inputs must produce identical outputs.

Active tool

Invocation

python tools/seal_package.py <package_dir>

Input

MAP/STYLE/PIC/TABLE evidence + CAP masks

Output

PACKAGE_SHA256.txt — freeze only; the mask manifest is missing

Dependencies

Python 3.12 (NOT installed on this machine)

Why this version

seal_package.py provides hash-binding and freezing, reused unchanged from V47 through V114. But hash-sealing a package is not the artifact the V10 closure defines as the protection manifest: a set of accepted TEXT and TABLE_TEXT masks with explicit removal authorisation, against protected classes MEDIA, LOGO, ORNAMENT, FRAME, RULE, BOX, TABLE_STRUCTURE, HEADER, FOOTER and PAGE_FURNITURE_TEXT, each with frozen width, height, encoded size and SHA-256. No implementation emits that manifest. Marked PARTIAL with the gap stated rather than papered over.

Tool receipt

2 file(s) copied, all hash-verified

Known limitations

• Package sealing and hash-binding are implemented. The protection mask manifest itself is not.
• Sweep therefore runs without a frozen, auditable statement of what it was forbidden to touch — which is the specific guarantee this step exists to provide.
• Protected classes cannot be verified as unchanged if they were never enumerated.

Archived versions (0)

No superseded versions recorded.

How it works

Performer and AI boundary

Performer

Deterministic CV/OpenCV.

AI

None.

Starts

Only after CAP polygons and PROTECTION are frozen.

Hands off to

ASSEMBLE, PDF, QA

Training

Not a trained component. No model, no weights, no training corpus. Behaviour is fixed by code and configuration, and identical inputs must produce identical outputs.

Active tool

Invocation

python tools/mechanical_pipeline.py

Input

original pixels + CAP masks + protection manifest

Output

swept substrate + removed-ink mask + residual report

Dependencies

Python 3.12 (NOT installed on this machine) · OpenCV · Frozen CAP polygons · Protection masks — see step 20, incomplete

Why this version

The V4 mechanical sweep is the code-bearing base of the sweep lineage (V4 to V5 swept substrate to V6 alignment to V7 solver). Later sweep work exists (V32 SWEEP_ENGINE_REWRITE, V33 HEADER_BOUNDARY_COMPLEX_SWEEP) and is newer by date, but those packages contain evidence and reports only — their code is not in the archive. PARTIAL for that reason, and because the owner's own correction records that the current system fails at this step: line-local fills producing pale strips, and removal of whole rectangular line boxes rather than glyphs.

Tool receipt

4 file(s) copied, all hash-verified

Known limitations

• The owner's own correction records that the current system fails here: line-local fills produce pale strips, and whole line boxes are removed instead of glyphs.
• Code is present only up to the V7 generation. Two later sweep packages (V32 engine rewrite, V33 header-boundary) contain evidence and reports but no code in the archive.
• Sweep cannot be trusted while step 20 emits no protection manifest.
• The swept substrate is a renderer derivative only. It must never feed an evidence detector, and there is no feedback edge back into MAP, CAP, ROLE, TYPE, PIC or TABLE.

Archived versions (2)

How it works

Performer and AI boundary

Performer

Deterministic ID-based assembler.

AI

None.

Starts

After required page/content/reconstruction and instrument states are available.

Hands off to

QA, PDF, DGTL, ARCH

Training

Not a trained component. No model, no weights, no training corpus. Behaviour is fixed by code and configuration, and identical inputs must produce identical outputs.

Active tool

Invocation

python BUILDER_STAGE_B.py

Input

artifacts bound to one sealed run

Output

review_data.json — canonical page and issue package

Dependencies

Python 3.12 (NOT installed on this machine) · Frozen upstream artifacts bound to one run

Why this version

The builder joins by stable content-bound IDs and records INPUT_HASHES and SOURCE_HASHES, which is the V8 gate — mixed run, version or parent hashes are rejected. Evidenced at scale: 146 groups across 68 sequential pages in the issue-4547 package, with 13,815 anchors each owned exactly once.

Tool receipt

3 file(s) copied, all hash-verified

Known limitations

• Assembly is only as complete as its inputs: with no media and no table cells, the assembled package cannot represent those layers at all.
• Proven on one issue.

Archived versions (0)

No superseded versions recorded.

How it works

Performer and AI boundary

Performer

Deterministic validators plus required human visual inspection and independent audit where policy requires.

AI

Auditor may assist, but evidence remains reproducible.

Starts

After assembly; repeated after each compatible correction.

Hands off to

PDF, DGTL/ARCH publication gate, Owner acceptance

Training

Not a trained component. No model, no weights, no training corpus. Behaviour is fixed by code and configuration, and identical inputs must produce identical outputs.

Active tool

Invocation

python tools/qa_full_issue_v34.py / python tools/audit_v7.py

Input

assembled page packages

Output

audit.json + REPORT.md with an explicit verdict

Dependencies

Python 3.12 (NOT installed on this machine)

Why this version

Mechanical QA tools exist and run, and an independent pixel audit was performed as recently as 2026-09-04. But the standing verdicts are HOLD, not PASS: the V46 architecture re-audit returned HOLD_FOR_GENERIC_FIXES, and the V004 spell audit returned HOLD as a list of real spelling errors. Several QA gates in the V8 contract cannot be evaluated at all because their upstream roles — PIC, TABLE-PHYSICAL, TABLE-ASSIGN — have no implementation.

Tool receipt

6 file(s) copied, all hash-verified

Known limitations

• Standing verdicts are HOLD, not PASS. The V46 architecture re-audit returned HOLD_FOR_GENERIC_FIXES; the 2026-09-04 spell pixel audit returned HOLD.
• Several gates cannot be evaluated at all because their upstream roles — PIC, TABLE-PHYSICAL, TABLE-ASSIGN — have no implementation.
• Sample accuracy is not page accuracy. Earlier figures of 93.9 percent strict and 96.1 percent owner-relaxed were measured over 558 sampled words and must never be quoted as full-page accuracy.
• A previous dashboard claim was withdrawn after direct image review found mislabelled pages and a classifier false positive.

Archived versions (0)

No superseded versions recorded.

How it works

Performer and AI boundary

Performer

Deterministic pinned HTML/browser/PDF renderer.

AI

None.

Starts

Only from QA-eligible PagePackages.

Hands off to

Owner/reviewer, Authorized publication

Training

Not a trained component. No model, no weights, no training corpus. Behaviour is fixed by code and configuration, and identical inputs must produce identical outputs.

Active tool

Invocation

python tools/export_original_rebuilt_pdf.py

Input

QA-passed page packages in issue page order

Output

original-then-rebuilt QA PDF

Dependencies

Python 3.12 (NOT installed on this machine) · Pinned font, shaping engine, browser/PDF engine, locale and colour profile

Why this version

This exporter produces the original-then-rebuilt QA PDF, which is one of the two PDFs the owner's corrected pipeline requires. The separate production PDF — rebuilt pages only — is not produced by any tool. Only QA-passed page packages may enter either, and step 23 currently returns HOLD, so neither PDF is release-eligible today.

Tool receipt

2 file(s) copied, all hash-verified

Known limitations

• Only the original-then-rebuilt QA PDF is produced. The production PDF of rebuilt pages alone is not implemented.
• Nothing is release-eligible today because step 23 returns HOLD.
• Byte-level determinism is not proven and must not be claimed; only raster comparison across repeated builds has been considered.

Archived versions (0)

No superseded versions recorded.

How it works

Performer and AI boundary

Performer

Deterministic digital renderer with standardized templates.

AI

None.

Starts

After complete ORDER/ISSUE-LINK/TITLE/TAG and supported media/table bindings.

Hands off to

ARCH, REV, Authorized public site

Training

Not a trained component. No model, no weights, no training corpus. Behaviour is fixed by code and configuration, and identical inputs must produce identical outputs.

Active tool

Invocation

python build_viewer.py — emits a static read-only viewer bound to a specific review package

Input

review_data.js — frozen package

Output

static RTL HTML viewer

Dependencies

Python 3.12 (NOT installed on this machine) · A frozen review package

Why this version

V117 is the highest-numbered viewer package and its README binds it explicitly to the independently verified V002 review package. It renders page navigation, zoom, TOC-group overlays, frozen SPELL candidates and instrument-date overlays. It is read-only by construction, which matches the contract that DGTL renders and never authors.

Tool receipt

4 file(s) copied, all hash-verified

Known limitations

• The current implementation is a read-only page viewer bound to one frozen review package, not a general digital article renderer.
• It presents candidate evidence. Nothing in it is accepted, released, or production content.

Archived versions (0)

No superseded versions recorded.

How it works

Performer and AI boundary

Performer

Deterministic archive/index service.

AI

None.

Starts

After permanent article IDs, verified tags and date provenance exist.

Hands off to

Search/navigation, Authorized public archive

Training

Not a trained component. No model, no weights, no training corpus. Behaviour is fixed by code and configuration, and identical inputs must produce identical outputs.

Active tool

Invocation

TOC groups are embedded in the viewer data by build_viewer.py

Input

issue metadata + article IDs + verified tags

Output

TOC group overlay data only

Dependencies

Step 25 DGTL

Why this version

A TOC and group index exists inside the viewer bundle. The rest of the ARCH contract does not: no archive index, no tag or date index, no search records, no permanent article IDs and no permanent URLs. The native-versus-inherited date distinction that the gate requires is not implemented anywhere.

Tool receipt

1 file(s) copied, all hash-verified

Known limitations

• Only a table-of-contents group overlay exists, inside the viewer bundle.
• No archive index, no tag or date index, no search records, no permanent article ids and no permanent URLs.
• The native-versus-inherited date distinction, which is the gate for this step, is not implemented anywhere.
• Only exact printed primary titles may enter the table of contents as source titles; derived labels must be marked as derived.

Archived versions (0)

No superseded versions recorded.