# V2 rollback session · authoring brief

Source pin: `1ceae52c7896a59ac74a4429d410f3218f0b91c6`. This guide develops the
rollback loop and prediction sections of
`V2/architecture/rollback-netcode-and-tag-team.md` against their actual C++
implementations. Snapshot internals, transport, combat mechanics and live
display consumers need their own explanations and evidence.

Visual thesis: preserve the cream, ink and terracotta reading surface; make
frame identity and the difference between replayed history and the next live
result the strongest visual relationship.

Content plan: open with one corrected tick, establish the simulation and
presentation responsibilities, explain input normalization and prediction, trace
confirmation and replay, then teach duplicate confirmations, recovery and
caller-owned input boundaries. End with an investigation route connecting the
returned result to its telemetry and actual consumers.

Interaction thesis: retain the existing slide entrance, source disclosure and
topic navigation with reduced-motion behavior. Exact source annotations and
scrollable comparisons support closer inspection. No ornamental motion or
invented game screenshot is needed to explain the local mechanism.

## Communication jobs

| Slide                  | Job and relationship                                                                      | Emphasis and transition                                                                      |
| ---------------------- | ----------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------- |
| Corrected tick         | A real sequence: restore frame 1, replay 2–4, advance 5.                                  | The returned frame is 5; introduce the owners of this work.                                  |
| Two worlds             | Compare simulation state with presentation records.                                       | Records do not prove sound, particles or a rendered overlay; locate the session entry point. |
| Input identity         | Compare the caller's frame stamp with the session's next-frame stamp.                     | Local input is normalized before advancing; explain missing remote input.                    |
| Prediction             | Compare held controls with pressed/released edges.                                        | Prediction restamps and clears edges; introduce confirmation.                                |
| Confirmation           | Annotate the actual stored-prediction comparison and removal.                             | Only a compared mismatch requests a rewind; explain the equality predicate.                  |
| Equality               | Separate hashed controls and frame/player identity from metadata flags.                   | Recomputed hash equality has a specific scope; follow multiple confirmations.                |
| Earliest correction    | Compare incoming order with the selected replay start.                                    | All confirmations enter the predictor before the earliest correction is replayed.            |
| Worked replay          | Align frames with the exact remote direction used in the native case.                     | Distinguish exact confirmations from repeated latest-confirmed input.                        |
| Result and trace       | Compare the replay's frame-4 hash with the returned frame-5 hash.                         | Different records answer different questions; inspect duplicate input.                       |
| Duplicate confirmation | Show two same-frame confirmations and the final stored payload.                           | A collected mismatch can still cause replay after the payload changes back.                  |
| Recovery               | Compare over-budget freezing, empty correction sets and a successful eligible correction. | Repeating a consumed confirmation does not create a fresh correction.                        |
| Player keys            | Annotate the exact sixteen-bit player-key expression.                                     | Invalid player indices can alias valid storage; identify caller validation needs.            |
| Net graph              | Separate supplied network measurements from confirmed-frame arithmetic.                   | A sample is a diagnostic record, not a measured session correction history.                  |
| Investigation          | Follow a concrete debugging sequence through inputs, replay, live result and consumer.    | Local source characterization leaves cross-platform and live-match acceptance open.          |

Choose sequences only for ordered operations. Tables carry exact frame and field
comparisons; annotated source carries implementation details. Every slide needs
substantive notes, source citations and neutral reader questions. Avoid a
four-point generic opening or a series of interchangeable lists.

## Evidence and review contract

Use the existing explicitly adapted UE 5.5.4 diagnostic host and preserve its
twelve fixture substitutions. Require source integrity, bounded incremental
build actions, discovered tests, terminal results and unchanged dependency
modules. The rollback mechanism remains at the source pin; this is not an
unchanged-pin/full-game build, Blueprint acceptance, strict-FP or cross-platform
proof, a live socket test or an audiovisual consumer test.

Review exact full source sections and their callee boundaries before making
teaching assignments. Read the complete notes and guide sequence, test every
slide at desktop, compact, mobile and embedded sizes, and inspect every PDF
page. Extend semantic/negative-control automation for these particular claims.
Recordings must pass fingerprints, decoding and pause checks after integration.
Normal-speed listening and willing-reader comprehension require separate
evidence; candidate compositions are not accepted benchmarks.

## Initial editorial findings

Direct desktop review exposed two mismatched table headings: the caller does not
supply the completed next-frame input set, and computed restore/depth values are
not confirmed directions. The revised labels identify who builds the input set
and use Record / Value / Role in replay for the mixed example records. Shorter
equality labels preserve the full mask explanation in notes. The net-graph row
now names players 0 and 1 explicitly so their indices cannot be mistaken for
literal frame numbers. These visual-only edits preserve every spoken block.
Final recorded/embedded/PDF checks must use the revised data.

The exact-source PDF check also needed a reflow correction: extracted line
breaks inside an identifier or after a cast are allowed while all non-whitespace
characters and annotation order remain required. Selectable HTML independently
retains the exact original source text, including whitespace. Regression cases
reject changed casts, missing operators and reordered annotations.

## Eight craft requirements and acceptance boundaries

- Benchmark: use the established cream/ink/terracotta pilot as a candidate; no
  accepted benchmark or representative-reader sign-off is inferred.
- Typography and alignment: preserve Georgia/Arial and the shared table rules.
  Review every new compact/mobile slide and PDF page; exact code can wrap while
  its annotations remain aligned with the corresponding source span.
- Quantities: align frame identities and explicit counts in the worked tables.
  The illustrative 4/2/1/3/5 trace is native-tested; it is not match telemetry.
- Continuity: carry that same tick from the opening through identity, earliest
  correction, replay and result; clearly label the separate duplicate/recovery
  cases so they cannot silently change the original input sequence.
- Image direction: source excerpts and frame tables are the evidence. Invented
  game or net-graph images would imply consumers this guide has not established.
- Actual-size readability: cover desktop, compact, 390px and 320px widths,
  landscape, enlarged text, embedded playback and actual 200%/400% table zoom.
  Tall reading captures and deliberate scrolling do not establish viewport fit.
- Comprehension: willing-reader evidence remains outstanding. Ask a reader to
  predict the returned frame/hash, explain why a repeated confirmation cannot
  create a fresh mismatch, identify the condition that releases recovery and
  separate a supplied rollback count from the session's actual replay count.
  Record unaided answers and misconceptions before claiming acceptance.
- Subtraction: four visual labels were corrected or shortened after inspection.
  Keep the source caveats in notes and precise takeaways; avoid decorative
  cards, extra imagery or automatic numbered relationships.
