Elamur Groupantsocrfix-text-command
The groupant beside tools/ocrfix/src/main.rs in the Mova repository: what the code computes, what it never does, how to go back from its outputs, its invariants, cost and checks.
Facts
Implementation material. A Mova code model describes this code at one source revision: partial models, finite checks and their stated scope. It is not a reviewed general exposition, and its labels are shown as written, never raised.
Components
| component | label | where |
|---|---|---|
| approximant | partial | approximant |
| ast | partial | syntax.json |
| composition | partial | composition |
| cost | symbolic source derived | invariant-and-cost |
| descriptant | present | descriptant |
| invariant | partial | invariant-and-cost |
| tests | executed bounded checks | tests |
Obernants
The typed relations around the forward model: the behaviour complement (what the code never does), the reverse reconstruction (every input compatible with an output) and specialised views.
| id | type | label | where |
|---|---|---|---|
ocrfix-text-command-behavior_complement | behavior_complement | partial source derived model | behavior_complement |
ocrfix-text-command-reverse_reconstruction | reverse_reconstruction | partial source derived model | reverse_reconstruction |
ocrfix-text-command-channel_budget | channel_budget | partial source derived model | channel_budget |
ocrfix-text-command-score_margin | score_margin | partial source derived model | score_margin |
ocrfix-text-command-mask_coordinate_join | mask_coordinate_join | partial source derived model | mask_coordinate_join |
ocrfix-text-command-effect_cleanup_order | effect_cleanup_order | partial source derived model | effect_cleanup_order |
Parts
Model text
descriptant
ocrfix-text-command
Epoch 2026-10-31
Full source read preceded changes. Prospective model predicted branch/order, resource and information-loss behavior. Tests exercise actual helpers on bounded synthetic inputs; command workflows remain unexecuted. Production source prefix is preserved; only inline tests added.
Independent commands compose normalization/lexicon/channel correction, CER alignment, OCR masks, witness voting and line export. Corrector applies two-edit candidates and strict log-score improvement; CER normalizes letter streams then diagonal-first DP. External OCR renders then recognizes/masks then writes text, removes PNG after recognition closure.
Executable percentage approximation is unsupported: categorical parsing/effects have no useful certified metric here. Partial syntax excludes macro expansion and name resolution.
approximant
Forward relation
Independent commands compose normalization/lexicon/channel correction, CER alignment, OCR masks, witness voting and line export. Corrector applies two-edit candidates and strict log-score improvement; CER normalizes letter streams then diagonal-first DP. External OCR renders then recognizes/masks then writes text, removes PNG after recognition closure.
Domain
UTF-8 Hebrew text, finite positive lexicon total/oov and nonnegative finite costs for score interpretation; lengths/boxes bounded against u32/i32 overflow. Raw parsers do not enforce all premises.
composition
Local composition
The scoped alignment fragment models its declared source interval. Parent relation: Independent commands compose normalization/lexicon/channel correction, CER alignment, OCR masks, witness voting and line export. Corrector applies two-edit candidates and strict log-score improvement; CER normalizes letter streams then diagonal-first DP. External OCR renders then recognizes/masks then writes text, removes PNG after recognition closure.
Shared representation/invariant: Skeleton correction retains punctuation and unchanged original token if no improvement; altered skeleton drops points. Final-letter gate rejects internal final forms only. Alignment counts sum to DP distance under bounded u32 lengths. Command output is nontransactional. Child costs are included in parent costs, not added a second time. Other command paths are tagged alternatives; the child alone does not establish whole-file behavior.
models
Typed source models
All models are partial source-derived relations at this manifest snapshot. Scope is one call/trace under the stated domain; no universal semantic proof.
behavior_complement
Short words below min_len cannot change; unattested/nonpositive candidate cannot win; mixed core scripts/brackets leave original token. OCR network/model/process effects are outside tests.
Domain: UTF-8 Hebrew text, finite positive lexicon total/oov and nonnegative finite costs for score interpretation; lengths/boxes bounded against u32/i32 overflow. Raw parsers do not enforce all premises. Observation: Skeleton correction retains punctuation and unchanged original token if no improvement; altered skeleton drops points. Final-letter gate rejects internal final forms only. Alignment counts sum to DP distance under bounded u32 lengths. Command output is nontransactional. Assumptions: bounded resources and source branch semantics. Oracle: selected actual-helper tests in tests.md constrain only their finite fixtures; remaining branches are source-derived.
reverse_reconstruction
Skeleton maps many pointed/editorial texts to one consonant sequence. Corrected output admits unchanged originals or channel predecessors plus compatible lexicon scores; logs do not recover stripped marks.
Domain: UTF-8 Hebrew text, finite positive lexicon total/oov and nonnegative finite costs for score interpretation; lengths/boxes bounded against u32/i32 overflow. Raw parsers do not enforce all premises. Observation: Skeleton correction retains punctuation and unchanged original token if no improvement; altered skeleton drops points. Final-letter gate rejects internal final forms only. Alignment counts sum to DP distance under bounded u32 lengths. Command output is nontransactional. Assumptions: bounded resources and source branch semantics. Oracle: selected actual-helper tests in tests.md constrain only their finite fixtures; remaining branches are source-derived.
channel_budget
Depth bounds edit path length2; cost gate c<=max+1e-9; best map stores strictly cheaper discovered paths; original removed even if reachable via cycle.
Domain: UTF-8 Hebrew text, finite positive lexicon total/oov and nonnegative finite costs for score interpretation; lengths/boxes bounded against u32/i32 overflow. Raw parsers do not enforce all premises. Observation: Skeleton correction retains punctuation and unchanged original token if no improvement; altered skeleton drops points. Final-letter gate rejects internal final forms only. Alignment counts sum to DP distance under bounded u32 lengths. Command output is nontransactional. Assumptions: bounded resources and source branch semantics. Oracle: selected actual-helper tests in tests.md constrain only their finite fixtures; remaining branches are source-derived.
score_margin
Candidate wins only ln(count)-ln(total)-cost > incumbent+1e-12; lexical candidate iteration resolves near-equal ties by first encountered incumbent.
Domain: UTF-8 Hebrew text, finite positive lexicon total/oov and nonnegative finite costs for score interpretation; lengths/boxes bounded against u32/i32 overflow. Raw parsers do not enforce all premises. Observation: Skeleton correction retains punctuation and unchanged original token if no improvement; altered skeleton drops points. Final-letter gate rejects internal final forms only. Alignment counts sum to DP distance under bounded u32 lengths. Command output is nontransactional. Assumptions: bounded resources and source branch semantics. Oracle: selected actual-helper tests in tests.md constrain only their finite fixtures; remaining branches are source-derived.
mask_coordinate_join
Each Hebrew box uses maximal positive intersection mask box; ties use last max iterator occurrence. Contiguous line keys preserve physical grouping; empty keep still newline.
Domain: UTF-8 Hebrew text, finite positive lexicon total/oov and nonnegative finite costs for score interpretation; lengths/boxes bounded against u32/i32 overflow. Raw parsers do not enforce all premises. Observation: Skeleton correction retains punctuation and unchanged original token if no improvement; altered skeleton drops points. Final-letter gate rejects internal final forms only. Alignment counts sum to DP distance under bounded u32 lengths. Command output is nontransactional. Assumptions: bounded resources and source branch semantics. Oracle: selected actual-helper tests in tests.md constrain only their finite fixtures; remaining branches are source-derived.
effect_cleanup_order
Correct/vote creates output dir before lex reads; OCR PNG removal occurs after recognition closure, but render failure returns first. Multi-file writes have no rollback.
Domain: UTF-8 Hebrew text, finite positive lexicon total/oov and nonnegative finite costs for score interpretation; lengths/boxes bounded against u32/i32 overflow. Raw parsers do not enforce all premises. Observation: Skeleton correction retains punctuation and unchanged original token if no improvement; altered skeleton drops points. Final-letter gate rejects internal final forms only. Alignment counts sum to DP distance under bounded u32 lengths. Command output is nontransactional. Assumptions: bounded resources and source branch semantics. Oracle: selected actual-helper tests in tests.md constrain only their finite fixtures; remaining branches are source-derived.
tests
Native evidence — 2026-10-31
Package/bin: ocrfix. Exact selected names:
tests::brackets_rejoinedtests::stream_and_certests::channel_corrects_and_keepstests::negative_control_known_word_stayswave33::channel_depth_alignment_and_masks
All 5 selected checks passed twice in debug and twice in release, one exact name per invocation with one test thread. Numeric exits are zero; sorted named outcomes match across all four runs. Added test blocks were formatted before these receipts; original source prefix (including earlier tests) remains byte-identical.
One compiled production mutant changed for _ in 0..depth to for _ in 0..0. The new wave33 test compiled and failed its runtime assertion (exit 101). Exact original/restored SHA-256 equality and a restored positive exact test (exit 0) were verified. No assertion-only mutant is counted; existing negative assertions remain ordinary tests. No surviving compiled mutant is omitted.
The external wave33 text receipts retain native-commands.json, verified-outcomes.json, mutation-report.json, raw logs/statuses, paired syntax projections and prospective.md. No private filesystem path is part of this public model.
Scope: actual helpers on synthetic in-memory data only. UTF-8 Hebrew text, finite positive lexicon total/oov and nonnegative finite costs for score interpretation; lengths/boxes bounded against u32/i32 overflow. Raw parsers do not enforce all premises. The scoped fragment reuses these tests where its helper is reached; other declared paths remain source-derived. Macro expansion, full command/effect execution, corpus accuracy and universal correctness are not established.
invariant-and-cost
Invariant and cost
Invariant
Skeleton correction retains punctuation and unchanged original token if no improvement; altered skeleton drops points. Final-letter gate rejects internal final forms only. Alignment counts sum to DP distance under bounded u32 lengths. Command output is nontransactional.
Cost
Normalization O(B); two-depth channel frontier may expand O(branching^2) with string/BTreeMap costs. CER O((n+1)(m+1)) time/space; word alignment O((N+1)(M+1)(L+1)^2) time plus O((N+1)(M+1)) matrix. Mask merge O(HM*textlen+H), process latency unmeasured.
Arithmetic and allocation premises are conditional. No timing or percentage surrogate claimed.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/tools/ocrfix/src/main.rs.groupant/