← E2T System V1 portal

17 — SPELL

Nominate suspicious printed spelling/anomaly candidates without altering the capture ledger.

Scope

Stage 1 runs a project-trained token classifier over units tagged instrument or instrument_candidate, flagging suspicious spans above a frozen threshold. Stage 2 applies a deterministic exact-glyph verifier that can only suppress a stage-1 candidate. Every surviving flag carries the exact value, its offsets, the page, group, token, bbox and crop linkage, and the model identity that produced it. Nothing is corrected.

This role is allowed to

• Flag words, numbers and dates for review
• Return no-decision/HOLD

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.

Non-scope — what this role does NOT own

• Rewriting CAP
• Claiming a correction is truth
• Deciding a flag is a real error
• Authorising training or release

Explicitly forbidden

• Rewrite CAP
• Claim a correction is truth
• Use untraceable language-model guesses

Dependencies and position

Starts

After TAG; for instrument and instrument_candidate units.

Previous step

16 CRITICAL-NOMINATE

Next step

18 REV — HUMAN REVIEW

Hands off to

REV

Exact inputs

• Exact CAP tokens and full context
• Structural role, semantic tag and source crop linkage

Exact outputs

• Suspicious token/span
• Anomaly type, score, reason and model/tool provenance

Performer and AI/ML boundary

Performer

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

Class

Project-trained

AI boundary

Model detection is allowed; correction authority is not.

Training information

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.

Deterministic validation and acceptance gates

• Every flag preserves exact value and offsets
• Zero-output on a substantial eligible set is NOT_MEASURED until coverage is proven

HOLD and failure behaviour

Failure codes this step may emit:

• SPELL_PROVENANCE_MISSING
• SPELL_SPAN_UNBOUND
• SPELL_NOT_MEASURED
• SPELL_OUTPUT_MUTATION

Downstream handoff

REV

Active tool

Status

Active tool

Run / inspect

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 is the active 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.

Copied files — source receipt

15 file(s), all hash-verified against the original. Full detail in tool/SOURCE_RECEIPT.json. Copy-only: the historical source is never modified.

Archived versions

3 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 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.

Security and privacy

Operates on frozen local text. No credentials. The 414 MB weight file is carried by Git LFS with its SHA-256 recorded, so the artifact stays verifiable.

Cost behaviour

Zero. Training cost USD 0.00 with no provider calls; inference is local CPU.

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

SPELL

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.

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.