roundtrip-pipeline
The groupant beside tools/roundtrip/src/pipeline.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 | model.md |
| ast | partial | syntax.json |
| composition | partial | Composition |
| cost | partial | invariant-and-cost |
| descriptant | present | descriptant |
| invariant | partial | invariant-and-cost |
| tests | 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 |
|---|---|---|---|
behavior_complement | behavior_complement | partial source derived model | behavior_complement |
reverse_reconstruction | reverse_reconstruction | partial source derived model | reverse_reconstruction |
context_partition | context_partition | partial source derived model | context_partition |
damage_precedence | damage_precedence | partial source derived model | damage_precedence |
head_repair | head_repair | partial source derived model | head_repair |
failure_gate | failure_gate | partial source derived model | failure_gate |
Parts
Model text
descriptant
roundtrip-pipeline
2026-10-31 prospective epoch retained externally before edits.
groups partitions sentence indices by Unit. run segments and annotates, reads frames with context groups, realizes/world-encodes with world groups, then exact decodes R4->R3->R2->R1; normal mode independently compares starts. damage edits first eligible residual; drop_class maps surviving heads past removed ancestors.
No annotator/model loading, live corpus, full suite, teacher, GPU or persistent-learning evidence. No executable percentage surrogate: categorical codecs/effects need a useful metric before future caps 1/2/5/10%.
Exact recovery with residual is a diagnostic codec observation. It is not evidence of normal generative equivalence, semantic understanding or persistent learning. Production source is preserved; only isolated tests are appended where needed.
approximant
Forward relation
groups partitions sentence indices by Unit. run segments and annotates, reads frames with context groups, realizes/world-encodes with world groups, then exact decodes R4->R3->R2->R1; normal mode independently compares starts. damage edits first eligible residual; drop_class maps surviving heads past removed ancestors.
Domain: bounded synthetic values, valid source-specific coordinates, nonoverflowing counts, available allocation; fixed environment controls. Relations are source-derived partial models with finite checks, not universal proofs.
Composition
Source children delimit real function scopes, sharing syntax. Real pipeline joins R4 frame reconstruction to Realizer then R3 tree script and R2 token decode; report consumes aggregated tally. Independent path/CLI branches are not fictitious codec callers.
catalog
behavior_complement
Subject: roundtrip-pipeline, source hash 88be3b618dd828411c1e6633432af60ba9fa905429b55cff88fbc31b5ae721da; transition observation.
Exact_ok requires final text equality and no earlier frames/tree/token mismatch; a tree mismatch is not decoded further. Normal text scores do not establish exact recovery, semantic understanding or retained knowledge.
Domain: bounded synthetic values, valid source-specific coordinates, nonoverflowing counts, available allocation; fixed environment controls. Relations are source-derived partial models with finite checks, not universal proofs.
Evidence: external wave32 rt_core receipt includes native-commands.json, exact named outcomes twice debug/twice release, raw status/log files, compiled production controls with byte restoration and paired syntax. Only named synthetic helpers execute; model_path/export and live CLI paths have representation/source evidence.
Status: partial source-derived model.
reverse_reconstruction
Subject: roundtrip-pipeline, source hash 88be3b618dd828411c1e6633432af60ba9fa905429b55cff88fbc31b5ae721da; transition observation.
Exact output with all intermediate residuals and unchanged decoders admits original trees/frames/tokens as reconstruction witnesses. Tally alone collapses volumes and cannot recover them. Unit World and Sentence share singleton worlds but differ reader contexts.
Domain: bounded synthetic values, valid source-specific coordinates, nonoverflowing counts, available allocation; fixed environment controls. Relations are source-derived partial models with finite checks, not universal proofs.
Evidence: external wave32 rt_core receipt includes native-commands.json, exact named outcomes twice debug/twice release, raw status/log files, compiled production controls with byte restoration and paired syntax. Only named synthetic helpers execute; model_path/export and live CLI paths have representation/source evidence.
Status: partial source-derived model.
context_partition
Subject: roundtrip-pipeline, source hash 88be3b618dd828411c1e6633432af60ba9fa905429b55cff88fbc31b5ae721da; transition observation.
Unit Sentence resets reader context each sentence; Block and World share volume reader context. Unit World resets realizer/world only, preserving read frames.
Domain: bounded synthetic values, valid source-specific coordinates, nonoverflowing counts, available allocation; fixed environment controls. Relations are source-derived partial models with finite checks, not universal proofs.
Evidence: external wave32 rt_core receipt includes native-commands.json, exact named outcomes twice debug/twice release, raw status/log files, compiled production controls with byte restoration and paired syntax. Only named synthetic helpers execute; model_path/export and live CLI paths have representation/source evidence.
Status: partial source-derived model.
damage_precedence
Subject: roundtrip-pipeline, source hash 88be3b618dd828411c1e6633432af60ba9fa905429b55cff88fbc31b5ae721da; transition observation.
Tokens gaps before tails; Tree spell before spacing before order; Frames first Insert/patch; World patches before toggles before inserts; Route false. Empty sentences alone are not damaged.
Domain: bounded synthetic values, valid source-specific coordinates, nonoverflowing counts, available allocation; fixed environment controls. Relations are source-derived partial models with finite checks, not universal proofs.
Evidence: external wave32 rt_core receipt includes native-commands.json, exact named outcomes twice debug/twice release, raw status/log files, compiled production controls with byte restoration and paired syntax. Only named synthetic helpers execute; model_path/export and live CLI paths have representation/source evidence.
Status: partial source-derived model.
head_repair
Subject: roundtrip-pipeline, source hash 88be3b618dd828411c1e6633432af60ba9fa905429b55cff88fbc31b5ae721da; transition observation.
drop_class renumbers survivors and follows removed-head chain with n-step guard. Valid head coordinates required; cycle of removed nodes falls back root.
Domain: bounded synthetic values, valid source-specific coordinates, nonoverflowing counts, available allocation; fixed environment controls. Relations are source-derived partial models with finite checks, not universal proofs.
Evidence: external wave32 rt_core receipt includes native-commands.json, exact named outcomes twice debug/twice release, raw status/log files, compiled production controls with byte restoration and paired syntax. Only named synthetic helpers execute; model_path/export and live CLI paths have representation/source evidence.
Status: partial source-derived model.
failure_gate
Subject: roundtrip-pipeline, source hash 88be3b618dd828411c1e6633432af60ba9fa905429b55cff88fbc31b5ae721da; transition observation.
After R4 comparison frames mismatch is recorded first; R3 tree mismatch prevents Dec calls for that sentence; text mismatch is fallback if no earlier recorded failure.
Domain: bounded synthetic values, valid source-specific coordinates, nonoverflowing counts, available allocation; fixed environment controls. Relations are source-derived partial models with finite checks, not universal proofs.
Evidence: external wave32 rt_core receipt includes native-commands.json, exact named outcomes twice debug/twice release, raw status/log files, compiled production controls with byte restoration and paired syntax. Only named synthetic helpers execute; model_path/export and live CLI paths have representation/source evidence.
Status: partial source-derived model.
tests
Native evidence
pipeline::wave32::grouping_drop_and_damage
17 exact named checks total (16 library, one CLI helper), each twice debug and twice release, numeric exits zero and sorted outcomes identical. Native metadata and raw receipts are in the external wave32 rt_core run directory. No optional model tests execute. lib.rs has compile/representation evidence only.
Four compiled production controls alter class precedence, Block grouping, world boundary XOR and frame lemma patching. Each fails its named runtime assertion (exit101), restores exact source SHA256 and passes restored check (exit0). No assertion-only mutants, survivors or bug repairs. These controls do not establish complete coverage.
Two fresh syntax projections agree for each of nine sources. Production bytes are preserved as exact base prefixes. Format only appended test modules before final native receipts. CLI spec helper errors precede any reader; actual model-loading pipeline is excluded.
invariant-and-cost
Invariant and cost
Exact_ok requires final text equality and no earlier frames/tree/token mismatch; a tree mismatch is not decoded further. Normal text scores do not establish exact recovery, semantic understanding or retained knowledge.
Exact output with all intermediate residuals and unchanged decoders admits original trees/frames/tokens as reconstruction witnesses. Tally alone collapses volumes and cannot recover them. Unit World and Sentence share singleton worlds but differ reader contexts.
groups O(n) singleton storage or O(1) Block. drop_class O(n^2) worst-case ancestor walking, O(n) vectors. run cost sums segmentation, annotation, frame realization, alignments, world passes and text diffs; retained intermediate payload is proportional to all generated representations, not streaming constant memory.
Domain: bounded synthetic values, valid source-specific coordinates, nonoverflowing counts, available allocation; fixed environment controls. Relations are source-derived partial models with finite checks, not universal proofs.
This page in Markdown โ https://md.groupants.elamur.ai/free/en/mova/tools/roundtrip/src/pipeline.rs.groupant/