Elamur Groupantsmova-whale-pod
The groupant beside tools/whale/src/pod.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 | Forward approximant |
| ast | partial | syntax.json |
| composition | partial | Composition |
| cost | partial | Cost |
| descriptant | present | Descriptant |
| invariant | partial | Invariant |
| tests | finite evidence | tests.md |
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 |
Model text
Descriptant
A pod's shared knowledge: one store for all Whales of a batch pod, keyed by normalized term (knowledge::norm_term), with provenance per Whale, paper, item and answer. A term one Whale asked is not asked by another; what one Whale learned, every Whale explains with. Bare-category facts go to review instead of knowledge. Source: tools/whale/src/pod.rs.
Forward approximant
open(dir): terms.jsonl (entries), claims.jsonl (first claim per term wins), review.jsonl, answers/. publish(whale, store): under pod.lock, every term item of the store not yet published -> non-bare facts as an entry, bare facts to review, the answer file copied; every asked term key -> a claim. claim(term, whale, paper): under pod.lock, reload; true and recorded only when no claim and no facts exist for the normalized term. items(exclude): entries of other Whales as knowledge items pod:<whale>:<item>, ordered by (term, whale, item). answers(): every copied answer text (for the overlap gate).
Invariant
Appends only, fsynced, under a blocking flock; a (whale, item) pair is published at most once; a normalized term has at most one effective claim; items() depends on the file's content, not its order; answer texts never leave the pod except into the overlap gate.
Cost
Each claim/publish rereads the three files (linear in pod size: hundreds of lines per batch); items() sorts O(n log n).
Composition
Parent: whale-workers. Written and read by learn.rs (publish at start and after each round, claim before asking a term, items for Known::of and the gates' item lookup, answers for the Guard); configured by batch.rs (pod).
behavior_complement
Not done: retracting or correcting a fact, choosing between conflicting facts of two Whales beyond «first per relation in the reader's order», sharing sentence or section readings (paper-specific), locking across machines (flock is local), deduplicating identical facts from different Whales. Parallel learners make append order timing-dependent; a Whale's view in a given round therefore depends on what the others had published by then. Evidence status: source-derived model with finite checks, not a proof.
reverse_reconstruction
Every entry names its Whale, paper, item id and answer hash; with the Whale's own store and the copied answer the source answer is recovered. A claim names who asked first; the facts do not determine the answer text.
Epoch 2026-11-04 (created)
Measured on a fresh pod with today's real answers replayed (10 papers, 5 Whales, sequential, two identical runs): 81 term entries, 118 claims, 26 review records; no term asked twice across the pod.
tests.md
pod::tests::one_term_one_question_across_the_pod (2026-11-04): plural and case variants are one term; a claimed or known term is not claimed again; bare facts go to review; other Whales read the entry, the publisher does not; a second open sees the same items. Negative control: has ignoring claims turns the test red (restored).
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/tools/whale/src/pod.rs.groupant/