Whisper

Identity

• Canonical ID: whisper
• Primary alias: Whisper
• Candidate aliases: muse, scribe, echoform, voiceforge, utterance, quill
• Scope: library

Purpose

Whisper turns author intent into meaningful text through a governed composition lifecycle. It extracts a small text intent substrate, selects a balanced combination of narrative primitives and techniques, proposes a construction plan, drafts the text, validates the result, and preserves learning residue for future writing.

The first proof target is a Substack post for an Arcanum research group. Fundraising copy is the next planned transport, so the initial substrate must preserve an extension boundary for persuasion, proof, trust, and ask mechanics without building that transport in L0.

Trigger Conditions

• The user wants to create or revise meaningful text, not just summarize existing material.
• The text needs a target public, tone, style, narrative anchor, structure, and validation criteria.
• The output can range from short copy to a post, article, slide narrative, fundraising page, or book section.
• The operator wants an interactive lifecycle with decisions, drafts, review, and iteration.

Required Sigils

Sigil	Role In Spell	Required Mode
structured-interview-kits	Ask one high-discrimination question when intent, transport, audience, or validation is ambiguous.	auto / interrogation
distill	Reduce broad author intent into the smallest coherent text intent substrate and prove recomposition into the full artifact.	standard or tournament

Optional Sigils

Sigil	Use When	Notes
decision-gate	A consequential choice blocks the writing direction, such as audience priority, ethical stance, evidence standard, or call to action.	Route only blocker-level decisions.
feature-glossary	A domain vocabulary or recurring research language needs stabilization before drafting.	Useful for research group posts and technical articles.
context-builder	Source material, notes, or prior artifacts need compact selection before composition.	Keeps drafting grounded without ingesting everything.
task-session	A long text needs SWUs, review tasks, or multi-step development.	Use for books, series, large articles, or fundraising campaigns.
experiment-harness	The reusable spell needs validation examples across transports.	Required before broad promotion.

Prerequisites

• A raw author intent, topic, rough thesis, or source material exists.
• A target transport is selected or can be selected through one focused question.
• The operator accepts that L0 optimizes for substack_research_post; other transports remain schema extensions until proven.

Shared State

State	Owner	Updated By	Consumed By
text_intent_substrate	Whisper	intake, distill	candidate tournament, composition plan, validation
transport_schema	Whisper	transport selection	composition plan, validation
scu_candidate_set	Whisper	distill tournament	Pareto consensus
composition_plan	Whisper	plan phase	draft phase, review
draft_artifact	Whisper	draft phase	validation, revision
learning_residue	Whisper	validation, reflection	future composition runs

Artifact Lifecycle Contract

Whisper treats composition as an artifact state machine, not as a loose sequence of writing tasks. Each artifact has a phase owner, a gate, and downstream consumers.

Artifact	Category	Produced In	Feeds	Gate
text_intent_substrate	intent substrate	intake, distill	candidate tournament, composition plan, validation	transport, objective, audience, SCU cores, and constraints are explicit
transport_schema	transport contract	transport selection	composition plan, validation	required body parts, length, evidence, introduction, ending, and CTA policy are named
scu_candidate_set	primitive/technique tournament	distill tournament	Pareto consensus	each candidate combines resonance, relevance, and trajectory rather than optimizing one category repeatedly
pareto_consensus	selection decision	candidate tournament	composition plan	selected set is non-dominated across resonance, relevance, trajectory, and cost
composition_plan	construction plan	plan phase	draft phase, review	body parts, sequence, anchor, examples, and validation checklist are ready
draft_artifact	text artifact	draft phase	validation, revision	draft exists, names its schema source, and preserves known citation gaps
review_html	review surface	review phase	revision, learning residue	every draft can be rendered into stable comment blocks whose agent payload preserves block_id, part_id, selected text, and requested change mode
validation_report	quality gate	validation	revision, learning residue	checks pass, flag, or block with actionable reasons
learning_residue	reusable memory	validation, reflection	future composition runs	distinguishes observed lessons from canonical author voice or transport rules

task-session is used only when one artifact needs bounded execution, such as drafting the article, verifying a source gap, or revising from review notes. The spell itself remains responsible for the lifecycle, gates, and artifact transitions.

Review HTML Contract

Every reviewable draft should be convertible into a Whisper review HTML page using spells/whisper/templates/draft-review-base.html and spells/whisper/tools/build-whisper-review-html.py.

The review page is the operator-facing surface for comments. It anchors each paragraph to a stable block_id, maps the block to a Whisper composition part_id, stores comments in browser localStorage, and exposes window.WhisperReview.getAgentPayload() for Playwright extraction.

Agent revision requests should consume the exported payload rather than vague prose-only feedback. Each comment carries:

• block_id and part_id,
• paragraph index and source line,
• selected text when available,
• issue type,
• requested change mode,
• priority,
• comment text,
• original block text.

This keeps review feedback addressable enough for the next Whisper revision pass to offer alternatives, apply a targeted change, or route a part-level mini-tournament when the comment says the local section failed.

Execution Phases

Phase	Sigil	Input	Output	Gate	Failure Policy
1. Transport and intent intake	structured-interview-kits	raw author intent	selected transport and blocker decisions	transport, objective, target public, and success signal are named	Ask one focused question; block only when target text cannot be identified.
2. Substrate distillation	distill	intake record	text_intent_substrate with resonance, relevance, and trajectory cores	each core has named values and recomposes into the target artifact	Flag unsupported assumptions; route consequential uncertainty to decision-gate.
3. SCU candidate tournament	distill	substrate and transport schema	candidate primitive/technique sets	one balanced candidate preserves Pareto trade-offs across all three cores	Keep stable disagreement in the ledger; do not optimize only tone/style.
4. Composition plan	Whisper	selected candidate set	body-part plan, template, validation checklist	plan includes introduction policy, narrative anchor, sections, ending, and constraints	Revise plan if it violates transport schema.
5. Draft and review	Whisper	composition plan	draft text and review notes	draft passes constraint, audience, resonance, and structure checks	Produce revision tasks or return flag when quality is below target.
6. Learning residue	Whisper	final or flagged draft	reusable lessons, technique results, unresolved gaps	residue distinguishes observed result from canonical truth	Defer promotion to inventory or glossary owners when durable knowledge appears.

SCU Core Model

Whisper uses three Smallest Coherent Unit cores:

Core	Responsibility	Typical Fields
resonance_core	Define the felt meaning the text should carry.	tone, voice, style register, emotional residue, value signal, forbidden feels
relevance_core	Define why this text belongs to this audience and domain.	target public, reader state, domain, authority mode, assumptions, objections
trajectory_core	Define the movement the text performs.	objective, transport type, narrative anchor, structure, introduction, body parts, ending, length

Each viable candidate must include one coherent selection from all three cores. Advanced mode may test multiple candidates per core, but L0 prevents three tone variants from crowding out audience fit or structure.

Transport Sequence

Order	Transport	Status	Reason
1	substack_research_post	selected L0 proof	Forces clarity, voice, research-group relevance, argument structure, and reflective ending without immediate conversion pressure.
2	fundraising_copy	next extension	Adds trust, proof, urgency, ask mechanics, and donor psychology after the base substrate works.

Local Customization

• Spell root: spells/whisper/ for this reusable library spell.
• Development artifacts: spells/whisper/development/.
• Gate strictness: standard for L0, strict for blocker audience/objective decisions.
• Interaction mode: guided-auto, with one-question interrogation when a decision changes the draft.
• Runtime execution: long model-backed work should use native skill/subagent execution with a receipt. When needed, a deterministic handoff wrapper can prepare receipt metadata; explicit legacy adapters are opt-in.

Observability

Record spell-level telemetry when .arcanum/observability/ exists:

• spell name and transport,
• phases attempted,
• sigils invoked,
• SCU cores selected,
• candidate set count,
• gates passed, flagged, or blocked,
• draft artifact status,
• validation result,
• user corrections,
• learning residue and next transport pressure.

Output Contract

Return:

## Whisper Result

- Spell: whisper
- Transport: <transport id>
- Objective: <author objective>
- Target public: <audience>
- SCU cores: <resonance | relevance | trajectory summary>
- Candidate selected: <candidate id and rationale>
- Composition plan: <sections or body parts>
- Draft status: pass | flag | block
- Validation: <checks and result>
- Learning residue: <durable lessons or none>
- Next route: revise | publish-prep | task-session | decision-gate | deferred