Elamur GroupantsโŒ•
โ˜ฐ

roundtrip-mixed-routing

The groupant beside models/roundtrip-mixed-routing.groupant 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

componentlabelwhere
approximantpartialforward
astnot materialized
compositionpartialcomposition
costsymbolic source derivedcost
descriptantpresentdescriptant
invariantpartialinvariant
testssee source local evidencechecks

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.

idtypelabelwhere
behavior-complementbehavior_complementpartial source derived modelbehavior-complement
reverse-reconstructionreverse_reconstructionpartial source derived modelreverse-reconstruction
coordinate-transportrepresentation_transportpartial source derived modelcoordinate-transport
tagged-routingtagged_routingpartial source derived modeltagged-routing
metric-denominatorsmetric_denominatorspartial source derived modelmetric-denominators
shared-helper-joinsshared_helper_joinspartial source derived modelshared-helper-joins

Model text

descriptant

roundtrip-mixed-routing: wave32

Actual joins: sample::Span -> route/owner in mixed; heb::R1/R2/R3 -> mixed exact decode; link::rename/inverse -> world frame identity; diff::compare -> tally::Diverge; Run::merge -> Tally::merge -> report metrics. compare consumes independent residual/normal schemas, not mixed report::tsv's nine columns. These shared helper joins are not an invented sequential pipeline.

This source-bound partial group organizes existing Mova code. Native evidence and excluded effects are recorded in source sidecars and the root wave32 report. Registration is not a universal proof or whole-system runtime claim.

model

roundtrip-mixed-routing

forward

Synthetic Item streams produce byte-span gold; route splits byte pieces and joint residuals, dispatches tagged En/Heb/Text codecs. Hebrew text tokens retain NFC/gap residual; link creates injective IDs before world encoding, inverse renames Hebrew frames after decoding. Run/Tally aggregate into mixed reports. diff is a shared divergence classifier; compare separately reads run TSV files and is not downstream of mixed TSV without an explicit schema adapter.

invariant

Valid UTF-8 boundaries, item identities, token/scalar/byte coordinate distinction and original residuals are required for diagnostic exact reconstruction. Accepted link pairs have unique Hebrew keys/English values, but rename inverse additionally needs no target/untouched ID collision. Nonnegative finite metric domain and nonoverflowing integer counts are caller premises.

cost

Route filtering O(segmentstokens); Hebrew frame alignment O(targetrealizedlemma bytes); diff LCS O(nm) memory/time. Link scans Hebrew/English piece pairs and entities with map lookup factors. Reports scan cells/items and sort leak rows. Full model inference, corpus IO and retries excluded from measured costs. Sample streams additionally construct ordered verse maps O(N log(V+1)verse-key bytes) and emit verse-pair documents beyond requested D. Report leaks repeat paired item/cell scans per variety and layer/class key; Q is bounded by the fixed enum universe, with explicit leaf bound O(AQ(I log(I+1)+C)+K log(K+1)).

composition

Actual joins: sample::Span -> route/owner in mixed; heb::R1/R2/R3 -> mixed exact decode; link::rename/inverse -> world frame identity; diff::compare -> tally::Diverge; Run::merge -> Tally::merge -> report metrics. compare consumes independent residual/normal schemas, not mixed report::tsv's nine columns. These shared helper joins are not an invented sequential pipeline.

domain

Small in-memory synthetic UTF-8 text, explicit spans, frames, dictionaries, finite residual cells; compare TSVs only under dedicated temp receipt root. All chosen tests read before selection.

limitations

No actual annotator/gparse/model/corpus/teacher/GPU execution. Full run_volume and parser/cache paths source-reviewed only. Allocation/IO failures and hostile malformed public structs not exhaustively tested.

checks

catalog

behavior-complement

On valid residual codec inputs, changing retained original bytes is excluded; token-only Text route cannot establish parser semantics. Mixed reports cannot be assumed compatible with compare TSV parser.

reverse-reconstruction

Retained tokens/scripts/joints reconstruct compatible source text under fixed decoder context; aggregate reports and normalized lexical keys have many antecedents. Exact identity recovery requires injective noncolliding rename.

coordinate-transport

Spans and owner overlap use bytes, sentence cuts use char_indices byte positions, diff preview counts Unicode scalars, LCS and tree Take use token indices, tree heads one-based.

tagged-routing

En/Heb/Text are alternatives; Text stores whole token residual. Language classifier loss and route accuracy use distinct labels.

metric-denominators

Report bits per sampled item differs from parsed sentence chrF and language token accuracy. Meta excluded in summary bits yet present in TSV.

shared-helper-joins

diff updates Diverge consumed by Tally; compare is independent file comparison with before-only set domain. No direct mixed-TSV adapter claimed.

This page in Markdown โ†’ https://md.groupants.elamur.ai/free/en/mova/models/roundtrip-mixed-routing.groupant/