📐mova-eel-autoencoder-round

The groupant beside tools/eel/src/autoencoder.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

Components

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.

Model text

Descriptant

The Eel's second round kind: an autoencoder round on CPU between parser rounds. The schedule eel_autoencoder.every (≥ 2) makes round n an autoencoder round when n mod every = every − 1; the kind is chosen in step at the round boundary, after recovery, cooldown, retention and the storage guard, and before the GPU probe. The round spawns the pinned native roundtrip wiki-round, checks its files itself, and installs the candidate tables (en-lin.bin, domain.tsv) only when the dev residual bits go down and every dev volume still comes back byte for byte. No GPU lease is taken, no teacher or large model is called.

Forward

Let S be the state directory with current → gen-<m> (or none: the base model B), k = ⌊n / every⌋. The child writes train T_k, dev D, test E, candidate C and measurements (b_D, b'_D, b_E, b'_E, x'_D, v). The Eel computes: sizes |T_k| = train_sentences, |D| = |E| = dev_sentences; articles(T_k) ∩ articles(D ∪ E) = ∅; hash(D), hash(E) equal the pin in S/dev.json (written by the first round); the measured current model path is S/current/en-lin.bin or B. Decision: installed iff x'_D = v > 0 and b'_D < b_D; otherwise rejected with autoencoder_exact_gate or autoencoder_dev_bits_gate. Any check failure, a timeout or a child failure is an error: failure.json, PAUSE, no retry.

Invariant and cost

The installed tables change only through install: under S/.eel-ae-install.lock, the parent hash is rechecked, the candidate is copied to gen-<n> and hash-verified, then one rename of a symlink switches current; the previous generation stays. The parser reject streak is carried unchanged. state_dir must lie outside run_root. Cost: one child process (sample, two measures of D and E, one Model::adapt over ≤ domain_max sentences), file hashes of the base model (≈150 MB) and the roundtrip executable each round, two copies of ≈160 MB on install, retention of three older generations. Measured cost of a real round: see History.

Composition

lib::step → kind → (autoencoder) round → spawn (roundtrip wiki-round) → check_sample, pin_heldout → gate → install / refusal → lib::finish (receipt, counter, retention). The summary keeps the parser receipt's lineage aliases (trainer_generation = hash of round.json, actual_parent_weights_sha256, candidate_weights_sha256, installed_weights_sha256) so evidence::inspect classifies it unchanged.

Behavior complement

Domain: rounds of a configured Eel. Observation: which work runs and what changes. Assumptions: the schedule is validated at start. Relation: an autoencoder round never calls the parser trainer, never probes or takes the GPU lease, never writes the parser model directory; a parser round never spawns roundtrip; a refused candidate never changes current. Counterexample: a crash between the symlink switch and the receipt leaves new tables without a receipt (the round directory then pauses the controller for manual reconciliation). Evidence: finite fake-executable tests below.

Reverse reconstruction

Domain: receipts and gen-* manifests. Observation: an installed table hash. Relation: the manifest of gen-<n> names its parent hash, base hash, dev/test/train hashes and the dev/test bits before and after, so the chain of installs can be walked back to the base model; the window itself is recovered from k and the corpus (deterministic sample). Not unique: the domain file is capped at domain_max lines, so older windows are not recoverable from the tables alone.

Heldout exclusion

Domain: sentence ids wiki:<title>:<k>. Relation: a train id whose article is a dev or test article fails the round before any install; a dev or test file whose hash differs from the pin fails it too. This is an independent recheck of the child's own split, not a proof that the corpus split function is used.

Tests

schedule_alternates_kinds_at_round_boundaries, worse_equal_or_inexact_candidates_are_refused_and_tables_stay, dev_article_in_training_or_a_moved_dev_sample_pauses_without_install, autoencoder_round_runs_without_the_gpu_lease_and_parser_round_still_waits, a_switch_request_lets_the_parser_round_finish_then_stops, unpinned_executable_or_base_pauses_before_any_autoencoder_work (in lib.rs). Red controls: admitting equal bits, dropping the dev or test overlap check and ignoring the pin each turned one test red (see History).

This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/tools/eel/src/autoencoder.rs.groupant/