Eve library

Events, jobs and discovery · One finished session, and everyone who wants to know · receipt · 12 September 2026

Guide

Revision 1

On this page

Chapter three of Shared systems, delivered in the Eve edition on 12 September 2026. Two inherited guides merge into one forty-four-slide chapter under a new events-jobs-discovery id. The brief is authoring/events-jobs-discovery-chapter.md. Machine-readable state is in state.json.

What was delivered#

Forty-four slides, all authored here. Thirty-four came from v1-events-jobs-discovery and ten from communication-discovery; ten inherited slides merge away and ten are new — a cover, six section dividers, a close, and two teaching slides.

The merge direction is set by the reviewed foundation. Seven of the communication-discovery slides are foundation slides and must keep their identifiers, so the rewrite carries the foundation id and the V1 slide that duplicated it merges into it instead:

Merged away Into
v1-event-mode-choice communication-modes
v1-event-bus-topology event-bus-reality
v1-job-envelope durable-jobs
v1-integration-boundary integration-boundaries
v1-search-pools search-live-path
v1-recommendation-candidates recommendations-and-graph
v1-search-curation curation-ownership
communication-discovery-boundary communication-modes
communication-discovery-worked ejd-obligations
communication-discovery-failure v1-event-review

content/70-v1-events-jobs-discovery.json held those thirty-four slides and nothing else, so it empties and is retired. content/05-shared-architecture.json loses seven slides and content/23-shared-deepdives.json three, all of which move into this chapter or merge within it. Both guide ids retire; the chapter takes communication-discovery's place in every ordering, so the v1-platform-foundations route and the sophia and psyche-lilith prerequisites follow it.

Evidence#

No capture. The chapter's subject is a fan-out and a request path rather than a surface, so every claim is executed or counted at the pin.

  • The live search ranker, executed here. Four lexical weights — forty for an exact title phrase, twenty-four for an exact summary phrase, eight per title token, three per summary token. Nineteen named kinds carrying six distinct boosts from fourteen down to seven, with anything unrecognised scoring seven. Domain intent adds ten to the named domain and subtracts four from the rest. Two guards exist and neither can fire: the score is clamped at zero and candidates scoring zero or less are filtered out, but the lowest reachable total is three. The ranker returns every candidate it is given. An empty query scores every candidate a lexical one, so a blank search returns everything the caller is authorised to see, ordered by the kind-boost table.
  • The recommendation scorer, executed here. Five weights — 0.34, 0.18, 0.18, 0.18 and 0.12 — summing to exactly one. Nine reasons, all nine present in both of the scorer's lookup tables, producing composites between sixty-five (time_based) and fifty-three (popular) with everything else held equal. A lone candidate from an under-represented domain ranked first of nine against eight otherwise identical siblings, fifty-nine against fifty-three, on the diversity bonus alone.
  • The shipped topic registry, executed here. Seventy-six topics across seven owning domains, one of which owns fifty-one. Three hundred and ninety-two declared fields, two hundred and forty-two of them required. Every topic is active, every schema version is backward-compatible, every descriptor is json-schema, and not one topic carries a second version — the deprecation and compatibility machinery is written and unexercised. The payload check is required-field presence only: an undeclared extra field validates.
  • The two event envelopes, executed and compared here. The contract schema has nine fields and the bus interface ten; they differ on five. The contract refuses the string 1.0, which is the default schema version every one of the seventy-six registered topics declares. The contract also accepts unknown top-level keys, so a bus-shaped envelope loses its extra fields silently.
  • Outbound signing, executed here. HMAC-SHA256 over the timestamp, the event id and the body, returned as six headers with a v1= prefix and compared in constant time. It refuses a changed body, a different event id and a signature made with another secret. The timestamp is divided to whole seconds before signing, so a signature verifies across any timestamp inside the same second and fails one second later. Backoff measured for all three strategies.
  • The durable queue, executed here. Five named priorities converting to the backend by subtracting from one hundred; all five round-trip, and the resulting order is critical, high, normal, low, background. An unnamed priority returns undefined rather than falling back to normal. Six SLA metrics each produce a breach on their own against the shipped editorial policy, and the warning ratio raises an at-risk condition before the threshold. Six job classes ship policies whose waiting limits span twenty-five to one thousand, and every one of the six sets its poison tolerance to zero.
  • Counted at the pin. The event bus opens two Redis connections and issues twenty-one distinct commands. None is from the Streams family, and there is no MULTI, EXEC or EVAL anywhere in the file. The recommendation route fans out to six domain adapters — Tara, Veritas, Nyx, Arete, Nisaba and Metis — matching the scorer's literal one-sixth diversity target, while search's canonical domain order lists four.
  • Read at the pin. The search route can answer with five statuses before any candidate is gathered, and refuses no query for being short or empty. The worker defaults to concurrency one, a thirty-second lock, a thirty-second stalled check, one stalled recovery and one attempt. The offline-eval gate computes six per-slice metrics and three drift detectors, and imports its summary builder from the retired search library — which is the sense in which that library is "adopted".
  • A comment that describes something absent. The scheduler's comment says a Lua transaction keeps the range read and the removal from racing. There is no Lua in the file. The scheduler reads due entries, removes each one, then publishes; the exclusion is real, but it comes from the removal's return value rather than from the atomicity the comment names. The slide says so.

What this chapter's evidence tops out at#

Executing a pinned module proves what that code does with the inputs it was given. Counting at the pin proves how often a pattern appears in the repository, not how often it runs. Reading a route proves what an author wrote. None of the three establishes that a published event reached a subscriber, that a queued job ran, or that a member saw a result — and no claim here is that. The SLA monitor was driven through a doubled substrate: the policy evaluation is real, the queue statistics were declared.

Checks#

Check Result
Build 1,623 slides, 704 sources; freshness check clean
Python suite 481 tests pass
Node suite 489 tests pass
Browser sweep 44 slides × 1440, 1280 with audio chrome, 390 and print — 0 findings
PDF 44 pages, exact authored visual text; the chapter has no table
Narration 44 tracks, 62.1 min; full library --check verifies 1,623 tracks
Bindings 58 reviewed source units, regenerated and re-fingerprinted

The first sweep reported nine findings across five slides, all of one kind: content tall enough to push the narration chrome off screen. It took four passes to clear, because trimming text only helped where the text was the cause. Two things actually were:

  • Two card grids held five and six cards, where every one of the two hundred and thirty-six existing Eve card grids holds three or four. v1-event-context became a record-anatomy and v1-event-review a graph-diagram — both node counts the library already uses.
  • Two diagrams were authored on a one-based grid where the renderer expects zero-based, which made each of them a row taller than intended. Shortening their labels changed the overflow by nothing at all, which is what pointed at the grid.

The rest was trimming: a case study to three nodes, which is what every other case study in the library carries.

Ten narration tracks were pruned after the merge. A merged-away slide keeps its mp3 and its manifest entry, and the full-library --check fails on the inventory mismatch while --pending stays clean — so the ten mp3s and their manifest entries were removed with the slides.

Binding work#

Twenty-three of this chapter's slides are cited in assignments/v1-platform-depth.json, which is generated. The generator targeted v1-events-jobs-discovery by name and was repointed. The regenerated file holds fifty-eight units — the same count as before the merge, so nothing came unbound.

Two things made that possible. The assembler restores each slide's original heading citations before writing the manuscript, moving a merged slide's headings to the slide that absorbed it. And it folds duplicate citations so each path appears once: authoring a path twice, once per heading, reads identically on the slide but hides the second set from anything that resolves the path, which is how the review would have silently lost a unit.

assignments/shared-curation-surface.json is hand-maintained and cited curation-ownership; that slide survives the merge, so only its guide id changed.

Tests updated#

Five, each with the reason written beside it: the platform-depth assignment test (guide id, slide count 34 → 44, the sequence, and both chapters now Eve rather than one); the rollback guard (the merge recorded, and the ten merged ids — of which only three appear in the published baseline, the other seven having been added after it); the tracked-chapter count, 1,107 → 1,151; and two pinned Node tests, pinned-curation-surface and pinned-event-bus-boundaries, whose guarded teaching was restored to the rewritten slides before the tests were repointed at the new manuscript.

Treatment deviations#

Four, each deliberate:

  • No capture, as in the two sibling chapters: the subject has no surface, and the plan's "one real capture per chapter" is answered by executed evidence.
  • v1-job-dead-letter became a state-machine, which is what the plan asked for. An earlier pass used recovery-map because a first search of the visual module did not turn state-machine up; the library has five other state-machine slides and it is the better fit for six named statuses.
  • v1-event-context: card grid → record-anatomy. Five continuations is more cards than any Eve card grid in the library carries, and the five actions are naturally a record of what each one writes.
  • v1-event-review: card grid → graph-diagram. Six ordered questions, likewise more than a card grid holds; the graph carries the order explicitly.

What this chapter does not establish#

Nothing here touches Redis, a deployed service, a real member account or a live queue. Every module was executed in isolation against declared inputs, and every count was taken across the repository at the pin.

Correction (12 September 2026)#

state.json claimed that check-center-coverage.py --check reported no stale bindings. That claim was wrong. The generated assignment was clean, but rewriting curation-ownership left the hand-authored reviewed binding in assignments/shared-curation-surface.json pointing at the slide's old fingerprint, and the rewrite also left a duplicated heading citation on that slide. Both were fixed later the same day with the Iris chapter: the citation was de-duplicated and the unit re-reviewed against the rewritten slide, which still carries the package inventory the unit teaches.