# Three complete authored SCRATCH-6 histories

These are independent fictional histories starting from the same before source and empty note. Step labels are sequence markers, not logs. The expected buffer and note are fully supplied in scratch-fixture.md; no real system produced these states.

## A: dictation supplies text; a person edits the source

| Step | Target and action | Expected visible state in the story | Next decision |
| --- | --- | --- | --- |
| A0 | Identify source buffer and separate empty note | Six-line before source; note empty | Keep source separate from note |
| A1 | Put insertion point in note and dictate the requested line | Editable note draft appears | Inspect text before relying on it |
| A2 | Correct or paste exact tokens as necessary | Note exactly `replace freeSeats with availableSeats`; source unchanged | Dictation has completed only the note part |
| A3 | Manually select only L2 declaration token and type/paste `availableSeats` | L2 renamed; L3 still `freeSeats`; L6 unchanged | Source temporarily incomplete; do not treat as finished |
| A4 | Manually select only L3 reference token and type/paste `availableSeats` | L2/L3 match expected source; L6 unchanged | Compare full source and note |
| A5 | Inspect exact two-token diff and preserved string | Expected source plus expected note | Completed text-edit exercise; no execution |

No dictation product is assumed to have performed a rename. A's edit selection comes from the person. The intermediate A3 state is intentionally incomplete and should not be mistaken for the final result.

## B: configured editor actions plus a literal note

Story prerequisites: a hypothetical command system is configured for the chosen editor to select an exact displayed token, replace that selection with exact supplied text, and enter literal text in a separate note. The operator can inspect the target and selection before replacement. These are declared scenario capabilities, not Talon, IDE, operating-system or installed-command facts. This file provides action intentions, not an executable voice grammar.

| Step | Intended operation, not a spoken command recipe | Expected visible state in the story | Next decision |
| --- | --- | --- | --- |
| B0 | Identify source buffer and its before text | Six-line before source; note empty | Establish current target |
| B1 | Use the configured select action for L2's `freeSeats`; inspect selection | Exactly declaration token selected | If mapping/selection cannot be established, take manual fallback |
| B2 | Use configured replacement for that selection with `availableSeats` | Only L2 renamed; source temporarily incomplete | Select the second code token |
| B3 | Select and inspect only L3's `freeSeats`, then replace with `availableSeats` | Expected source; L6 guidance unchanged | Inspect full source |
| B4 | Focus separate note; choose the configuration's literal-text entry route | Note target identified and empty | Do not interpret note sentence as an editing command |
| B5 | Enter exact note line as text and inspect | Expected source plus note `replace freeSeats with availableSeats` | Compare with common target |
| B6 | Inspect two-token diff and preserved L6 | Same completed text artifacts as A | No assertion of real command execution or better performance |

If B1 or B3 has unknown selection, no replacement is assumed. Preserve the current buffer, use manual selection/typing to finish the remaining explicit tokens, then compare against the common target. If B2 already occurred, L2 stays renamed; do not restart with blind global replacement or erase the note. This fallback retains valid task scope; changing input method alone does not require renewed permission.

The N2 Talon source reference documents distinct text-entry/key/mode definitions; it does not establish B's exact-token selection or semantic rename in any installed editor. An adaptation must fill its real capability card before turning the intentions into a command recipe.

## C: an AI task produces a candidate text patch

| Step | Target and action | Expected state in the story | Next decision |
| --- | --- | --- | --- |
| C0 | Keep before source and empty note as the reference | Neither edited | Use complete task packet |
| C1 | Dictate or type the explanation into an AI request draft; paste exact source, names and literal note from delegation-request.txt | Complete inspectable request draft | Check task meaning and exact strings |
| C2 | Submit the comparison-ready request in the intended conversation | Request asks for patch/note as text, no workspace operations | Wait for candidate; no edit inferred |
| C3 | Examine a candidate matching expected-change.patch and expected-note.txt | In this authored story, candidate changes only L2/L3 and preserves L6 | Candidate itself is not an applied change |
| C4 | Compare each diff line and the literal note with the fixed task | Two intended source tokens only; exact note present | Manually adopt into disposable scratch buffers if desired |
| C5 | Compare adopted text to common expected source and note | Same final text artifacts as A/B | Task complete in the story; no runtime/test claim |

C3 is an authored expected candidate, not a captured model response. No real agent was called. In actual use, missing or extra changes return to the same two-token task. A plausible explanation does not replace inspecting the patch.

## Three precise wrong turns

| Wrong turn | Why it misses SCRATCH-6 | Bounded response |
| --- | --- | --- |
| A note contains the phrase but no source token changes | Dictation entered text; rename part is unfinished | Select the two explicit source tokens and edit them manually, or choose a configured action/delegation route |
| A global replace changes L6 to `replace availableSeats with availableSeats` | Literal guidance was not a rename target | Restore L6 from fixed before source; inspect the remaining diff without broad reversal |
| A voice action targets an unknown selection | Intended operation does not identify the receiving token | Leave the uncertain operation unissued; inspect current buffer and use manual fallback |

If an operation has actually been issued but its result is unknown, inspect current text before repeating it. A switch of modality does not establish cancellation or restore the original buffer.