# Agentic Studio · applied source corrections, not guide acceptance

Audience: a product or engineering reader who has read orientation and Isis and
needs to distinguish declarations, admission, attempted work and observed
effects. The source pin remains `1ceae52c7896a59ac74a4429d410f3218f0b91c6`.

Visual thesis: keep the cream, ink and terracotta reading plane; give each local
check, execution order and recovery question its own explanatory structure.
Content plan: establish scope and vocabulary, distinguish validation from
authority, examine the run record and controls, follow one eligible attempt,
then compare outcomes and return to delegation, isolation and delivery evidence.
Interaction thesis: retain slide entrance, source disclosure and sidebar/drawer
transitions with reduced motion. Static HTML and PDF must retain every qualified
relationship. No decorative media or new interactive dependency is introduced.

## Reading sequence and communication jobs

| Slide                           | Job and relationship                                                                                                                 | Transition                                                    |
| ------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------- |
| `agentic-studio-boundary`       | Establish shared automation versus Eve's builder runtime. Existing opening still needs the full source/craft pass.                   | Identify the declared tools.                                  |
| `agents-catalog`                | Orient the tool/version inventory; not an exhaustive implementation walkthrough.                                                     | Ask what preflight can establish.                             |
| `agents-preflight`              | Compare four supplied declarations and their actual validation limits.                                                               | A valid declaration still needs a run record and enforcement. |
| `agents-envelope`               | Explain attempt telemetry, usage, raw/redacted payload boundaries and evidence limits.                                               | Distinguish records from control actions.                     |
| `agents-lifecycle`              | Explain individual status guards, ordinary resume, fork and array-suffix retry.                                                      | A checkpoint is a different recovery mechanism.               |
| `agents-checkpoints`            | Compare identity, selected fields, recorded budget and returned state.                                                               | A queued result is not dispatch.                              |
| `agents-admission`              | Compare the helper, legacy plan and governed plan; retire the misleading universal rate-check graph.                                 | Follow the route's governed path.                             |
| `agents-execution-path`         | Trace the eligible success path in actual order, with conditional network preparation and failed-await/refusal scope.                | Interpret the outcome and charge.                             |
| `agents-attempt-accounting`     | Compare denied/error/OK/partial outcomes; notes work the 15/10/1 and last-call 25-unit examples.                                     | A charge or record need not capture every external effect.    |
| `agents-execution-interruption` | Compare last saved evidence, an actual callback and the recovery decision. Independent columns have no step numbers.                 | Reconcile effects before retrying; revisit authority.         |
| `agents-grants`                 | Retain the specificity table while distinguishing the simple resolver from the broker and caller-owned renewal/revocation duties.    | Delegation requires additional checks.                        |
| `agents-handoff`                | Compare history, authority and budget independently, without ordinal numbers; explain permitted re-entry and non-atomic deductions.  | Local validation is not transport enforcement.                |
| `agents-isolation`              | Existing network/secret boundary overview; deeper source and real-transport teaching remains open.                                   | Select the operating mode and human gates.                    |
| `agents-modes`                  | Existing mode comparison; actual approval and consumer integration still need review.                                                | Apply the pipeline contract.                                  |
| `agents-pipelines`              | Existing pipeline overview; registry, stages, failures and real execution still need granular teaching.                              | Follow the existing delivery example.                         |
| `agentic-studio-worked`         | Existing artifact/delivery example; not accepted as complete product evidence.                                                       | Test the recovery claim.                                      |
| `agentic-studio-failure`        | Compare checkpoint mismatch, same-ID content change, queued recovery and delegated/egress work with their additional evidence needs. | Return to source detail and actual deployment evidence.       |

The application retains the reviewed proposal text. Three bounded visual
refinements suppress ordinal markers on the envelope, handoff and interruption
comparisons; the one superseded admission diagram override is removed. Exact
before/after slide fingerprints and these refinements are recorded in
[the application record](agentic-source-application.json). No other authored
slide changes. Historical proposal tests remain bound to `7167e08ce5`; separate
live tests reject missing application, reintroduced overrides, unrelated changes
or malformed numbering options. Narration text is unchanged by these
refinements.

## Viewing and acceptance boundary

Check the full guide at 1440×900, compact 1280×720 including active narration,
390×844 initial and full reading views, every mobile table column, embedded
desktop/mobile, and every PDF page at page-fit. Tables retain row identity while
scrolling. The conditional preparation wording and completion-write distinction
must remain legible. Existing typography and shared table geometry are retained;
this is not a new accepted typography benchmark.

Source probes and browser assertions are not human comprehension, production
durability, exactly-once dispatch or whole-guide acceptance. No new
section-level teaching assignment is claimed simply because a source is now
cited. The recovery, authority and execution review maps retain the missing
full-unit explanations.

Use the [reader protocol](reader-review.md) with a willing participant and
freeze the final stimulus hashes before asking these questions without coaching:

1. A checkpoint comparison succeeds while an artifact keeps its identifier but
   changes its stored content. What do you know, and what would you check next?
2. Describe what has happened when the execution flow reaches the tool call.
   Which conditions could have stopped it earlier?
3. The last saved call says pending. What can you conclude about its external
   effect, and how would you decide whether to retry?
4. A plan is completed with one partial call. What does that tell you about the
   requested deliverable?

Assess answers against the cited implementation after recording the unaided
interpretation. Record viewing accommodations, uncertainty and any assistance;
do not invent participant results, quotes or a comprehension score. Final
normal-speed listening and actual reader sessions remain separate required work.
