The third chapter of the How Eve works track, rewritten in the Eve edition on
10 September 2026. It follows the plan agreed on 9 September and recorded in
EVE-V1-REDESIGN-TODOS.md. The receipt for the delivery is under
verification/; this file records what the chapter is for and why it is shaped
this way.
The job#
Chapter eight ended with four kinds of information arriving in one turn. This chapter is about the boundary that decides whether any of it may cause anything: who is asking, what a piece of text is allowed to do, and which rule wins when two of them disagree.
The mistake it is written against is treating persuasive text as permission. A retrieved page can explain a workflow accurately and, in the same paragraph, ask Eve to run an operation nobody authorised.
Running example. One support thread the builder is working: a retrieved page carrying somebody else's instruction, and a message carrying a person in distress. The same boundary has to handle both, and it handles them differently.
Three learning promises (the cover)#
- Content is not authority — where a sentence came from decides what it may cause, and that label survives assembly.
- Protection outranks completion — a documented precedence, an interruption that changes the whole journey, and a continuation the person controls.
- Permission and evidence expire — a grant specific enough to refuse, a budget that is not permission, and a measurement bound to its model and date.
Three things to remember (the close)#
- Content is not authority. Only one origin may carry conduct and tool rules; a summary keeps the least-trusted label of its inputs.
- The more protective rule wins. Approval does not reach past authorization, and an interruption changes the whole journey.
- Permission and evidence expire. A grant is specific enough to refuse with; a measurement belongs to its model, date and revision.
Bridge to chapter ten. What happens after an action is admitted: how accepted work executes, and what makes a result verified rather than reported.
Sections#
Four sections, opened by a section-divider that names all four.
| Section | Slides |
|---|---|
| Where authority comes from | trust labels, the worked attack, the precedence stack, the conflicting-rule decision, the crisis interruption, four outcomes, adversarial review |
| Grants, budgets and records | grants against budgets, one grant record, three independent checks, four data records, disclosure against receipt, threat coverage |
| The dated runtime | the evidence chain, what a turn is granted, high-impact recovery, the injection measurement |
| The surfaces, end to end | speech bindings, page capture, reply details, the end-to-end review |
Treatment decisions#
- The chapter is pinned at two revisions, on purpose. Eighteen slides read
the documented architecture at
1ceae52c; the seveneve-current-*slides read the implementation at6c3186d2, where the files they cite first exist. A chapter file carries exactly one revision, so the section lives in its own manuscript,content/68-eve-trust-runtime.json, and the scope line at the top of each of its slides says which snapshot it was read from. That difference is the section's first lesson rather than a formatting artefact. - Three stat panels, because three slides have real numbers. Threat coverage
(75 applicable cells, 49 controlled, 20 partial, 6 gaps), what one turn is
granted (11 authority dimensions, 1 explicitly absent, 0 filesystem roots, 0
command roots) and the injection measurement (120 attack runs and 80 control
runs per leg, 2 turn-leg successes, 0 escalation-leg successes). Every number
cites the file it was read from, and two new tests in
test_supplemental_sources.pyre-derive them from the pinned source so a changed constant fails rather than quietly ages. - No capture. The plan asked for captures of the reply-detail and voice
surfaces.
assets/eve-product/holds no crop of either, and this host's Chromium 148 against the recorded 152 would makeassets/eve-product/provenance.json's environment false if the harness ran here. The "at least two diagrams and one real capture" checkbox is therefore left unchecked, as it was for chapters five, seven and eight. The chapter has seven diagrams. - One record specimen.
eve-detail-grant-contractis the chapter's onlyrecord-anatomy: a grant is a record, and the point of the slide is that each field is one a service can refuse with. - Twenty-seven slides, not twenty-five. Five pairs of inherited slides
merged. Two more pairs would have removed
eve-audit-and-data-rightsandeve-prompt-trust-labels, which are reviewed foundation slides and may not leave the library, so the id of the surviving slide is theirs:eve-prompt-trust-labelscarries the merged trust-boundary teaching, andeve-audit-and-data-rightsstayed a slide of its own beside the disclosure panel. The dated runtime then split into two sections rather than running eight slides without a marker, which is the longest stretch the redesigned library would have had.
What was rebound#
- Twenty slides now live in
content/67-eve-trust.jsonand seven incontent/68-eve-trust-runtime.json; they came from02-eve.json,03-y-current-eve.json,21-eve-deepdives.json,53-eve-safety-detail.jsonand56-eve-current-runtime.json, the last two of which are now empty and removed. - Five assignment fingerprints re-reviewed rather than re-pointed, across two
assignment files, each with a dated note. Two of those re-reads found a
dropped claim and put it back: the client-supplied grant context on
eve-detail-grant-contract, and the four spending kinds a budget covers oneve-detail-impact-budget. - The course-shape test now expects the Lilith handoff as the chapter's last content slide, with the chapter's own close after it.