Document metadata
type: authoring-system-workshop
status: planned
updated: 2026-09-12
authority: workshop plan only; extracted traits require evidence and author review
source: "[[01 Sessions/Daily/2026-09-12 - Voice Key Authorial Decision Fingerprint]]"
Voice Key Guided Transcript Workshop
Purpose
Collect enough varied, guided speech and creative-choice reasoning to improve Voice Key's four functional layers:
- Authorial Decision Fingerprint;
- Prose Topology Profile;
- Residue Filter;
- Pattern Memory.
The workshop does not train the system to reproduce raw speech. It uses conversation to discover how the author prefers information to be selected, arranged, emphasized, demonstrated, left implicit, and concluded.
Status and priority boundary
This workshop is planned for later. It does not replace the active Dynamic Story Workshop or change its author gate. Start it only when the author explicitly says Open the Voice Key workshop.
- Eight modules.
- Approximately 10–15 minutes of dictation per module.
- Modules may be completed separately across several sessions.
- Ask one prompt at a time.
- Use natural follow-ups rather than reading a long questionnaire.
- Do not name the trait being tested before the response.
- Preserve exact dictation and a separate lightly corrected transcript.
- Keep at least one complete response per dimension out of profile extraction for later validation.
- End each module with a short reflection, not a large trait-approval screen.
Estimated complete workshop time: 80–120 minutes across multiple sessions.
Why the workshop mixes tasks
Natural monologues reveal attention, causal reasoning, certainty, and explanation habits. Comparative decisions reveal preferences more directly. Editing reactions reveal prose topology and residue sensitivity. Project-specific tests show whether a general preference improves Seeds rather than merely fitting unrelated conversation.
No single mode is allowed to define the fingerprint.
Module sequence
| Module | Primary evidence | Held-out element |
|---|
| VKW-01 Natural reasoning | unperformed explanation across unrelated topics | final topic response |
| VKW-02 Attention and relevance | what receives space and what gets ignored | one unfamiliar source |
| VKW-03 Causality and constraints | believable consequence and redirection | one counterfactual problem |
| VKW-04 Evidence and judgment | what makes conclusions convincing | one ambiguous case |
| VKW-05 Reveal and surprise | information order and earned non-default choices | one blind reveal comparison |
| VKW-06 Ambiguity and closure | required answers versus productive openness | one unfamiliar ending |
| VKW-07 Prose topology and residue | rhythm, structure, vocabulary, and summary preferences | one unseen prose set |
| VKW-08 Seeds calibration | project-specific selection and blind passage testing | final three-way comparison |
VKW-01 — Natural reasoning
Goal
Collect unperformed speech across unrelated subjects without asking the author to sound literary.
Prompt types
- Explain a system or routine that outsiders misunderstand.
- Describe an idea that changed after encountering practical limits.
- Tell a story whose meaning became clear gradually.
- Defend a judgment that depends on the purpose of the activity.
Follow-up behavior
Use questions such as:
- What changed your interpretation?
- Which part mattered more than it appeared to at first?
- What would somebody misunderstand about that?
- What happened because of that limitation?
- What would count as success in that situation?
Do not suggest a desired answer shape.
Output
- raw transcripts;
- corrected transcripts;
- candidate attention and causal traits;
- certainty contour;
- possible topic-specific contamination;
- one untouched held-out response.
VKW-02 — Attention and relevance
Goal
Learn which information the author expands, compresses, or ignores.
Exercise
Present two short source packets containing more facts than a finished passage needs. Ask the author to dictate:
- what the passage is really about;
- which three details deserve the most space;
- which details can be removed;
- which apparently minor detail may become important;
- how the information should open.
Use one practical subject and one fictional subject. Keep a third packet held out.
Follow-ups
- Why does that detail deserve more space?
- What can the reader safely infer?
- Which fact looks important but is actually decoration?
- If only one detail could remain, which one carries the argument?
Output
- attention weights;
- omission profile;
- specificity rhythm;
- high-resolution-detail preference;
- compression boundaries.
VKW-03 — Causality and constraints
Goal
Learn what makes change, failure, escalation, and surprise feel earned.
Exercise
Present a goal, available resources, system rules, and a new obstacle. Offer several possible next developments without labeling them. Ask the author to choose, combine, or replace them and explain the causal chain aloud.
Run three versions:
- an everyday planning problem;
- a technology or institutional problem;
- a fictional conflict unrelated to Seeds.
Follow-ups
- Which consequence feels convenient?
- What cost should the workaround create?
- Which limitation would genuinely redirect the plan?
- What would make the surprising option feel inevitable afterward?
Output
- constraint sensitivity;
- preferred causal-chain depth;
- coincidence tolerance;
- workaround-cost preference;
- surprise legitimacy rules.
VKW-04 — Evidence and judgment
Goal
Learn what must be shown before the author accepts a conclusion and how the author evaluates competing interpretations.
Exercise
Present several ambiguous situations containing testimony, behavior, records, incentives, and system outputs. Ask the author what can be concluded, what remains uncertain, and what additional evidence would change the judgment.
Do not use real criminal allegations or private individuals. Use invented or ordinary low-stakes situations.
Follow-ups
- Which evidence is strongest, and why?
- What are we being asked to assume?
- What would make this conclusion premature?
- Are the two sides measuring success differently?
- What fact would reverse your judgment?
Output
- evidence hierarchy;
- certainty thresholds;
- correction pattern;
- competing-goal sensitivity;
- qualification behavior.
VKW-05 — Reveal and surprise
Goal
Learn preferred information order independently of prose wording.
Exercise
For the same approved facts, present several reveal maps:
- conclusion before evidence;
- evidence before interpretation;
- consequence before cause;
- focal character misunderstanding before correction;
- ordinary expectation before a system exception.
Ask the author to select or combine a route and explain when the reader should understand more than the focal character.
Follow-ups
- What should the reader know first?
- Which fact changes the meaning of what came before?
- Is this suspense or merely missing information?
- Where does surprise become manipulation?
- Which version would remain interesting on rereading?
Output
- reveal-order distribution;
- delayed-recognition preference;
- reread-value preference;
- suspense versus confusion boundary;
- earned-surprise profile.
VKW-06 — Ambiguity and closure
Goal
Learn what a passage, scene, or ending owes the audience and what may remain open.
Exercise
Present several short ending summaries with different levels of resolution. Ask the author to identify:
- the primary promise;
- the questions that must close;
- the questions that may remain open;
- the difference between future possibility and unfinished work;
- the emotional consequence that should be demonstrated rather than explained.
Follow-ups
- What would make this feel evasive?
- Which unanswered question expands the story?
- What should change permanently by the end?
- Does the ending flow from the story or impose an effect on it?
Output
- closure hierarchy;
- open-perimeter tolerance;
- required-state-change preference;
- emotional-resolution preference;
- ambiguity limits.
VKW-07 — Prose topology and residue
Goal
Learn the shape of writing the author finds effective without asking the author to produce polished prose.
Exercise
Present unlabeled versions of the same passage that vary only in selected dimensions:
- uniform versus varied sentence lengths;
- full sentences versus purposeful fragments;
- balanced versus asymmetrical attention;
- early versus delayed thesis;
- explicit transitions versus logical proximity;
- three-item versus irregular lists;
- summarized versus pressure-changing paragraph endings;
- abstract dramatic language versus concrete verbs and nouns.
Ask for spoken reactions, combination requests, and exact reasons. Do not identify one version as human or AI.
Follow-ups
- Where did your attention drop?
- Which sentence feels too written by a system?
- Which irregularity improves the passage, and which feels performed?
- What should be deleted rather than rewritten?
- Does this ending conclude the thought or merely repeat it?
Output
- sentence-length distribution preference;
- fragment triggers;
- paragraph-shape profile;
- transition density;
- list and symmetry tolerance;
- thesis-position range;
- project residue lexicon;
- delete-versus-rewrite preference.
VKW-08 — Seeds calibration
Goal
Determine whether the extracted fingerprint improves actual Seeds of the Throne writing.
Required preparation
Use one approved scene contract and lock its facts, POV knowledge, revelation ceiling, entering state, and resulting state. The Book One outcome-presentation vertical slice remains the preferred first test when structurally ready.
Blind comparison
Create three unlabeled passages:
- strong project-aware baseline;
- surface speech-profile revision;
- Voice Key authorial-selection and prose-topology version.
Randomize their order. Keep length, facts, and scene purpose comparable.
Evaluation questions
- Which passage is best?
- Which feels most specific to Seeds?
- Which makes the strongest presentation decisions?
- Which cause-and-effect chain is most convincing?
- Which contains an earned surprise?
- Which feels polished but generic?
- What should be combined?
- Which exact decision should Voice Key remember?
Output
- blind preference;
- project-fit evidence;
- accepted and rejected modifiers;
- passage-level corrections;
- updated confidence;
- new held-out benchmark.
Capture model
Each response preserves three separate layers:
| Layer | Contents |
|---|
| Source | exact prompt, raw dictation, corrections, and situational context |
| Understanding | candidate trait, evidence, confidence, competing interpretation, and possible topic bias |
| Use | proposed general preference, project modifier, topology range, residue rule, or no promotion |
No inferred trait becomes active merely because it appears once.
Proposed evidence lifecycle
- Observed: appears in one response.
- Candidate: appears across two unrelated contexts or one explicit comparison.
- Working: survives a counterexample and improves a blind output.
- Author-approved: the author confirms the useful decision in plain language.
- Project-active: has a defined scope, trigger, and recent-use control.
- Revised or retired: later feedback weakens, narrows, contradicts, or removes it.
These stages prevent a single anecdote or speaking habit from becoming a permanent writing rule.
Training and validation split
- Split by complete response or exercise, never random sentences from the same answer.
- Do not use held-out responses to extract or phrase the profile before evaluation.
- Keep general-profile validation separate from Seeds-specific validation.
- Preserve at least one negative example showing when a normally preferred technique should not be used.
- Re-run blind tests after major profile changes.
Pattern-memory capture
During every comparison, record whether the author rejects a passage because a device is intrinsically weak or because it appeared too recently. Store those as different findings.
The workshop should identify acceptable repetition windows for:
- fragments;
- short punchline conclusions;
- delayed theses;
- evidence-first reveals;
- three-part lists;
- contrast constructions;
- abrupt paragraph endings;
- behavioral emotion;
- open questions.
Exact numerical windows remain an implementation test rather than an interview demand.
Success criteria
The workshop succeeds when it produces:
- evidence across unrelated topics and actual project material;
- at least one supported trait in each core fingerprint dimension or an explicit finding that the dimension should remain neutral;
- a project-specific topology profile that does not imitate raw dictation;
- negative preferences and counterexamples, not only positive rules;
- a held-out validation set;
- one blind comparison preferred for quality and project specificity;
- no canon changes hidden inside writing preferences;
- a clear list of traits still too uncertain to activate.
Stop and resume behavior
At the end of each module, record:
- module completed;
- raw sources captured;
- held-out material protected;
- candidate findings;
- unresolved contradictions;
- exact next prompt.
The author may stop after any response. Resume with a one-paragraph recap and the next single question. Do not repeat completed prompts merely because a new chat begins.
Start command
When the author says Open the Voice Key workshop:
- identify the earliest incomplete module;
- briefly explain that module's purpose without naming the hidden trait being tested;
- ask one natural-language prompt;
- wait for dictation;
- continue with adaptive follow-ups until the module has sufficient evidence or the author stops.
