Elamur Groupantsmova-eel-receipt-evidence
The groupant beside tools/eel/src/evidence.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 source derived | Forward |
| ast | partial | syntax.json |
| composition | partial source derived | Composition |
| cost | partial source derived | Cost |
| descriptant | partial source derived | Descriptant |
| invariant | partial source derived | Invariant |
| tests | finite evidence | 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 finite checks | behavior_complement |
reverse_reconstruction | reverse_reconstruction | partial source derived finite checks | reverse_reconstruction |
actual_parent_exclusion | partial_exclusion | partial source derived finite checks | actual_parent_exclusion |
outcome_error_partition | error_partition | partial source derived finite checks | outcome_error_partition |
Model text
model
Declared terminal evidence
Descriptant
Typed interpretation of historical trainer receipts and Eel summaries. This is an observation model of declared JSON, not a verifier of model files or a claim that a subprocess executed. Only installed, rejected, below_threshold and deferred are admitted outcomes. No filesystem/provider/GPU effects originate in this file.
Forward
Let O be the outcome, G a valid generation digest, P an actual-parent digest, C a candidate digest and I an installed digest. Digest validity is exactly 64 lowercase ASCII hexadecimal characters. Generation and actual-parent fields accept both old trainer names and Eel summary aliases. Proposed-parent is never P. The classifier is E(O,G,P,C,I): deferred/below_threshold -> NoTrainingRecorded; rejected and G&P&C -> CandidateRejectionRecorded; installed and G&P&I -> InstallationRecorded; otherwise -> IncompleteLineage. Unknown/failed outcomes error. No numeric approximation or probabilistic error rate is admitted.
Invariant
Changing only proposed-parent, round count, rejection streak or wall time cannot promote a receipt to recorded training. Missing/invalid digest fields stay None. Returned lineage is a projection of declared fields, never reconstructed bytes. Historical partial installed/rejected receipts stay readable as incomplete.
Cost
Fixed field count and bounded 64-byte digest validation; output clones at most four valid digest strings. Input JSON parsing/allocation belongs to the caller. No measured speed, real execution or freshness guarantee follows.
Composition
Native trainer terminal receipt -> Eel alias summary -> this classification. Full legacy digest tuples retain outcome-specific interpretation across aliases. The projection loses the original error text, metrics, config and artifact existence. It does not certify that a summary and trainer receipt agree on disk.
behavior_complement
Domain: JSON terminal outcome/digest tuples. Observation: TrainingEvidence. Assumptions: this is interpretation of claims. Relation: deferred and threshold outcomes exclude CandidateRejectionRecorded and InstallationRecorded regardless of counter/streak/proposed-parent fields. Counterexample: deferred round1709 with reject_streak9 did not itself train or reject. Evidence: three finite Rust controls; source-derived model, no formal proof or artifact-authentication claim.
reverse_reconstruction
Domain: JSON receipts. Observation: one classified ReceiptEvidence. Assumptions: same alias and digest admission rules. Relation: preimages preserve all variations in omitted metrics/errors/counters/proposed-parent and invalid/missing lineage. Counterexample: two deferred receipts with different proposed parent weights classify identically; original lineage intent cannot be uniquely reconstructed. Evidence: source inspection and finite non-promotion/alias controls, not a proof.
actual_parent_exclusion
Domain: partial receipt lineage. Observation: P presence and classified evidence. Assumptions: polling and acquisition may race; proposed weights precede trainer lease. Relation: proposed-parent alone cannot inhabit actual-parent fiber; require actual_parent_weights_sha256 or old_weights_sha256. Counterexample: another trainer may install between probe and actual lease. Evidence: explicit proposed-only test. This partial exclusion does not prove that a supplied actual-parent digest is true.
outcome_error_partition
Domain: known outcomes plus malformed/missing outcome. Observation: Result and E. Assumptions: JSON values remain caller-owned. Relation: unknown/failed -> Err; known partial -> Ok(IncompleteLineage), rather than inventing training or dropping historical records. Counterexample: valid-generation installed without installed weights stays incomplete. Evidence: missing-output and uppercase/partial controls.
tests
Three unit controls cover proposed-only/deferred non-promotion, outcome-specific actual lineage, and historical partial/failed/uppercase partitions. Public native entrypoint additionally checks full installed receipt interpretation. Finite JSON/script fixtures do not authenticate operational artifacts.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/tools/eel/src/evidence.rs.groupant/