create-egregore 0.20.2 → 0.21.0-runtime-mvp.2

package/runtime/claude/.claude/skills/harvest/PROCESS.md

@@ -0,0 +1,395 @@
+# Harvest — Cognitive Protocol
+
+> Harvest elicits latent human or agent context — preferences, positions,
+> tacit knowledge, taste, constraints, role-shaped judgment — and turns
+> the answers into organizationally usable context.
+
+This document is the **cognitive protocol** for `/harvest`. It defines how
+the model conducts the conversation: how it seeds, generates, evaluates,
+checkpoints, cascades, and synthesizes. It is authored fresh from
+[`AUDIT.md`](./AUDIT.md).
+
+For persistence (graph schema, async metadata, local-mode markdown
+contracts), see `AUDIT.md` §9–§11. This document references those
+contracts but does not re-specify them.
+
+For format (synthesis layers, synthesis blocks), see `FORMAT.md`. This
+document points to it but does not duplicate it.
+
+For question intent, answer shape, probe moves, and hard bans, see
+[`QUESTION_PALETTE.md`](./QUESTION_PALETTE.md). It is their canonical
+home; this document controls when they enter the conversational rhythm.
+
+---
+
+## 0. What the model is doing
+
+Harvest is **model-led, not flowchart-led**:
+
+> The model owns the next question. The protocol only requires that it
+> can explain why the question is being asked, preserve the answer, and
+> update its interpretation.
+
+There is no fixed round count, no fixed question count per round, no
+required dimensions. The model reads the situation, generates the next
+question with an articulable purpose, evaluates the answer, and decides
+whether to deepen, pivot, checkpoint, cascade, or synthesize.
+
+---
+
+## 1. Boundary (read once, then never again in this doc)
+
+**Harvest** elicits latent context. **Spiral** stress-tests claims. They
+share craft (good questions, careful evaluation, evidence-bound
+synthesis) but their objectives differ.
+
+Harvest does not pick winners. Harvest does not converge by obligation.
+Divergence between respondents is information, not a problem to resolve.
+**A harvest maps the choices; it need not make them** — for some
+harvests the output is *"Cem prefers X because…"*, *"Oz sees the
+boundary this way…"*, *"Renc is carrying a constraint we haven't
+documented…"* — useful context even when no recommendation is appropriate.
+
+That's the boundary. The rest of this document stays in harvest's own
+register: *elicit, preserve, situate, surface, interpret-with-evidence*.
+
+The six elicitation moves in §3 (seed, generate, evaluate, checkpoint,
+cascade, synthesize) are the durable craft. They were extracted from the
+prior spiral-context harvest draft; the objectives that drifted toward
+claim-stress-testing are not. After this section, that prior draft is
+not referenced again.
+
+---
+
+## 2. Invariants
+
+Every harvest must hold these. They are non-negotiable and apply
+regardless of harvest shape (single / multi, present / async, preference
+/ alignment / extraction).
+
+1. **No hidden role stereotyping.** Every implicit role inference is
+   inspectable in the manifest *and* surfaced into the interaction when
+   confidence is below threshold.
+2. **No survey collapse.** AskUserQuestion (AUQ) options never exhaust
+   the answer space. Every question carries a freeform escape hatch.
+   Options are drawn from context, not prefab taxonomies.
+3. **No unmarked cascade contamination.** When prior respondents'
+   positions are disclosed to a later respondent, that disclosure is
+   explicit, recorded on the manifest, and visible to the respondent.
+4. **No synthesis without evidence.** Every assertion in synthesis
+   traces to a recorded turn or to seed context, marked as such.
+5. **The model can articulate `questionIntent` for every question and
+   explain why the chosen answer shape fits it.** If it can't, the
+   question is not ready to ask. Choose from `QUESTION_PALETTE.md`; when
+   another shape fits better than its closest intent row, record a
+   one-sentence reason.
+
+---
+
+## 3. The six moves
+
+### 3.1 Seed
+
+Every harvest begins with **intent** and **context**.
+
+- **Intent** = what the harvest wants to learn — dimensions, not
+  questions. *"Understand each respondent's preferences, constraints,
+  and tensions around pricing"* is intent. *"What price should we
+  charge?"* is a decision, not a harvest intent.
+- **Context** = what is already known: seed documents, prior harvests,
+  RoleSheets for sender and receiver(s), recent sessions/handoffs/quests
+  on the topic.
+
+If intent or context is ambiguous, ask the harvest **initiator** for
+clarification before generating the first questions. A handful of
+clarifying questions upfront produces a dramatically better harvest than
+proceeding on assumption.
+
+**Resolve RoleSheets before the first question.** See `AUDIT.md` §6 for
+the schema. Two sheets per turn: `SenderSheet` (who is asking — usually
+the initiator) and `ReceiverSheet` (who is being asked). The asymmetry
+matters: *cem-asks-renc-about-pricing* differs from
+*renc-asks-cem-about-pricing*.
+
+### 3.2 Generate (situated questions)
+
+Generate questions live from intent + context + RoleSheets + what has
+come back so far. Never from a predetermined list.
+
+Each question carries:
+- An articulable **`questionIntent`** the model can state to itself.
+- An answer shape selected from `QUESTION_PALETTE.md`.
+- A short **header** (≤12 chars) when AskUserQuestion (AUQ) is the
+  chosen interaction.
+- Options drawn from the actual context (seed quotes, prior answers,
+  role-likely lenses) when the chosen shape uses options — never prefab
+  taxonomies.
+- A **freeform escape hatch** for every closed shape (AUQ provides
+  "Other" by default).
+
+Choose the move and answer shape from `QUESTION_PALETTE.md`. Its named
+conversational moves — ladder, critical incident, triad, best-worst,
+swing question, mirrored-inference check, and one-at-a-time forced
+choice — run inside this generate → evaluate rhythm; the palette owns
+their definitions and bans.
+
+**Round shape (affordance, not script)**:
+- One interaction per round. Use AUQ for structured choices and plain
+  dialogue for conversational moves that need no widget.
+- 1–3 questions per round (model decides; usually 2).
+- Round count is not fixed. End when returns diminish or the respondent
+  signals.
+
+**How much context to include with each question is judgment.**
+Considerations the model weighs:
+
+- Showing another respondent's position can provoke valuable reaction —
+  or bias the answer. The disclosure setting (§3.5) settles this per
+  harvest.
+- Extensive recontextualization helps a respondent who lacks
+  background. It patronizes one who has it.
+- Structured options scaffold a low-stakes decision. They flatten a
+  high-stakes preference. Use freeform when nuance matters.
+- A diagram or position map can orient. It can also frame the answer.
+
+**Forbidden in round 1**:
+- Leading questions that presuppose a position.
+- Disclosing other respondents' answers (unless the disclosure setting
+  explicitly permits it; see §3.5).
+- Asking what the seed already answers.
+
+**Always permitted**:
+- Asking the respondent to reframe the harvest's intent.
+- Asking what the model *should* be asking that it isn't.
+- Ending early at any checkpoint.
+
+**Anti-calcification handling.** When a `ReceiverSheet` field is
+inferred at low/medium confidence, surface the inference into the
+question itself: *"I'm reading you as approaching this from the eng-lead
+lens — push back if that's wrong."* The respondent can correct in one
+turn, before the inference shapes a chain of questions.
+
+### 3.3 Evaluate
+
+After each response, the model notices what came back and records an
+**`evaluation`** alongside the answer. The kinds of things a good
+interviewer notices:
+
+- **Novel framing** — the respondent reframed the question in a way
+  that reveals a different mental model. High signal. Explore it.
+- **Unresolved tension** — the answer contradicts something said
+  earlier or something in the seed. Worth probing — but probing is
+  *understanding the tension*, not pressing the respondent toward
+  resolution.
+- **Tacit content** — a preference, taste, or constraint surfaced that
+  the respondent treats as obvious but isn't documented anywhere. High
+  signal for a context/preference harvest.
+- **Shallow signal** — canned or surface-level answer. Try the
+  dimension from a different angle, or accept that the respondent
+  doesn't have more to give on it.
+- **Strong conviction** — clear, defended position. Clarify once if
+  useful; respect it. Strong preferences are data, not resistance.
+- **Surprising connection** — the respondent linked two dimensions the
+  model didn't anticipate. Follow the thread.
+- **Diminishing returns** — confirms what's known. Move on.
+- **Satisficing watch** — straight-lining, instant first-card picks,
+  one-word notes, or all-selected multis suggest shallow processing.
+  Do not treat the pattern as reliable signal; switch shape or probe
+  before synthesizing it.
+- **Sycophancy guard** — evaluations must record material disagreement
+  or unresolved tension when present. An evaluation stream that only
+  affirms the respondent is incomplete.
+
+These are orientations, not a checklist. They feed the next question
+generation and they accumulate into a **running interpretive state** —
+the model's evolving read of what this harvest is producing. By
+synthesis time, the model is assembling and interpreting from the
+running state, not starting from scratch.
+
+### 3.4 Checkpoint
+
+At natural transition points — when a cluster of dimensions feels
+explored, when returns diminish, when the respondent signals — offer a
+checkpoint:
+
+> "We've covered [dimensions]. Threads still worth pulling on:
+> [remaining]. Continue, or synthesize what we have?"
+
+The respondent co-steers. They might say *"skip X, go deeper on Y"* or
+*"I'm done"* or *"actually, there's something we haven't touched."* All
+productive responses.
+
+Checkpoints respect the respondent's time and agency. **Do not impose
+arbitrary limits** — no fixed round count, no fixed turn count. Offer
+checkpoints when a genuine transition arrives.
+
+### 3.5 Multi-respondent disclosure (cascade)
+
+Cascade = running the harvest across multiple respondents in sequence;
+the design question is what each later respondent sees. The harvest
+declares a **disclosure setting** (the `context_mode` field) at seed
+time and records it on the manifest:
+
+| Setting | Behavior | Use when | Default |
+|---|---|---|---|
+| **blind** | Each respondent answers independently; other respondents' answer content neither appears nor shapes their questions | Preference, estimate, or position intents where uncontaminated signal matters | **Yes** for preference, estimate, and position intents |
+| **disclosed** | Later respondents see attributed, crystallized prior positions | Reaction is the point; later respondents add value by responding to what already surfaced | Only when the initiator explicitly chooses disclosure |
+| **comparative** | Later respondents are explicitly asked to compare or contrast their position with attributed priors | The comparison itself is the elicitation intent | Only when the initiator explicitly chooses comparison |
+
+The setting is **explicit on the manifest, visible to respondents**, and
+shapes what context the model includes per question (§3.2). An unmarked
+cascade is an invariant violation (§2 invariant 3).
+
+In **disclosed** and **comparative** modes, attribute prior positions
+when surfacing them: *"Oz said X. Where does that land for you?"* —
+never *"Some on the team think X."* Anonymized contamination is worse
+than disclosed contamination.
+
+In **blind** mode, another respondent's answer content must not shape a
+later person's questions, even invisibly. Adaptive interviews may use
+the seed, RoleSheets, and that respondent's own prior turns. A blind
+shared-artifact round freezes the source version and question set before
+the first response so every respondent receives the same artifact.
+
+Never turn attributed positions into an anonymous or aggregate claim
+such as *"most of the team thinks X."* Name the respondents and preserve
+their reasons. A majority is not collective alignment; 2/3 may be
+reported as an attributed majority, never as a shared team position.
+
+**Corrections and sealing.** While collection is open, a respondent may
+correct their own response. Store each correction append-only; the latest
+finalized response is canonical, and the completion count is the number
+of distinct finalized respondents, not the number of submissions. A
+shared-artifact round seals only when its declared completion condition
+is met (by default, one finalized response from every declared
+respondent). After sealing, the evidence set is immutable. A changed
+position starts a new round that references the sealed one. There is no
+automatic second pass after disclosure.
+
+### 3.6 Synthesize
+
+Synthesis is a deliberate act when all respondents finish (or a solo
+harvest reaches diminishing returns). The model reads everything
+through the running interpretive state and produces a layered artifact
+per `FORMAT.md`.
+
+**Two principles bind synthesis**:
+
+1. **Interpret with evidence.** Every assertion ties to a recorded turn
+   or marked seed (§2 invariant 4). Quotes are verbatim and attributed.
+   Patterns name the turns they're extracted from.
+2. **Generative emission.** Which layers and which sections within them
+   exist depends on what the harvest produced. A pure preference harvest
+   may emit L0 + L2 + L3 with no L1; an alignment scan may emit a thick
+   L1 with thin per-person L2; a knowledge extraction may emit L0 +
+   structured-knowledge L2 + L3. The model chooses based on what's
+   present, not what the template "expects."
+
+Synthesis carries **no obligation to converge or pick a winner**. It
+surfaces what was elicited, organized for legibility. For some
+harvests, *"Cem prefers X because Y; Oz prefers Z because W; they don't
+disagree, they're optimizing for different constraints"* is the entire
+output.
+
+The synthesis artifact is `/reflect`-able: patterns, preferences, and
+positions emit in shapes the org's pattern/decision graph can ingest.
+See `FORMAT.md` §8.
+
+---
+
+## 4. The full rhythm
+
+```
+seed → generate → present → evaluate → ((generate → ...))* → checkpoint
+     → ((continue / pivot / cascade / end))* → synthesize
+```
+
+This is a rhythm, not a script. The model inhabits it the way a good
+interviewer inhabits conversation — alert, present, willing to follow
+signal, willing to end early, willing to be surprised.
+
+---
+
+## 5. Self-harvest (single-respondent case)
+
+When `sender == receiver` — *"I want to think through my own position
+on X"* — cascade is moot, the disclosure setting is moot, RoleSheet
+asymmetry collapses. Everything else applies: seed, generate situated
+questions, evaluate, checkpoint, synthesize. The model interviews the
+user and produces the same layered artifact. No special path; just the
+single-respondent case of the multi-respondent flow.
+
+---
+
+## 6. Async harvest (respondent not in session)
+
+When a respondent isn't present:
+
+1. The model generates the next round's questions for that respondent
+   based on intent + context + that respondent's RoleSheet + context
+   allowed by the disclosure setting. In blind mode, that means the
+   respondent's own prior turns only, never another respondent's
+   answers. A blind shared-artifact round dispatches its frozen question
+   set unchanged.
+2. Questions are emitted into a question file with **harvest metadata
+   in the frontmatter** — see `AUDIT.md` §9 for the schema. The async
+   bridge extends `/ask`'s existing markdown shape with `harvest_id`,
+   `harvest_session_id`, `turn`, `question_intent`, `context_mode`, and
+   `status`.
+3. The session's `HarvestSession` status is `pending` until the
+   respondent engages.
+4. When the respondent answers via `/ask`, status moves *pending →
+   answered*. The next time `/harvest --resume` runs, it incorporates
+   the answers, evaluates them under the running interpretive state,
+   and continues toward the next round or synthesis.
+
+The model treats async answers the same way as in-session answers for
+evaluation purposes. The latency is operational, not cognitive.
+
+---
+
+## 7. Resume
+
+Harvests span days. Re-entry is first-class.
+
+- `/harvest --resume {harvest_id}` — explicit resume.
+- Implicit resume: a `/harvest <topic>` that matches an active harvest
+  surfaces it and asks: *continuation or fresh start?*
+- On resume, the model reloads the manifest, all sessions, and the
+  running interpretive state from recorded `evaluation` fields. It does
+  **not** re-do already-answered turns. It picks up where it left off
+  or, if all sessions are `complete` / `incorporated`, proceeds to
+  synthesize.
+
+---
+
+## 8. The four anti-conflation fixtures
+
+Before adding any new rule to this protocol, run it through these
+fixtures. If a rule does not apply cleanly to all four, it has drifted
+into forced-adjudication territory and does not belong here.
+
+1. *"What's Cem's design taste for the artifact viewer?"* — pure
+   preference, no proposition being evaluated, no alignment target.
+2. *"What is Oz carrying about the moat question?"* — tacit position,
+   no decision being forced.
+3. *"What does Renc know about go-to-market constraints we haven't
+   documented?"* — knowledge extraction, nothing to recommend.
+4. *"What do cem, renc, and oz each prefer for pricing?"* — alignment
+   scan; divergence is information, not a problem to resolve.
+
+This is the test the protocol must pass.
+
+---
+
+## 9. What this document does not cover
+
+- **Persistence** — graph schema, ID conventions, async question
+  metadata, local-mode markdown contract: see `AUDIT.md` §9–§11.
+- **Format** — synthesis layers, synthesis blocks, layered disclosure
+  rules: see `FORMAT.md`.
+- **Question palette** — intent-to-shape selection, conversational-move
+  definitions, and hard bans: see `QUESTION_PALETTE.md`.
+- **Orchestration** — argument parsing, mode detection, file
+  read/write, ID minting: belongs in `SKILL.md` once it's rewritten as
+  a thin entry point.

package/runtime/claude/.claude/skills/harvest/QUESTION_PALETTE.md

@@ -0,0 +1,76 @@
+# Harvest — Question Palette
+
+> Ask for the kind of judgment the harvest actually needs.
+
+This document is the single canonical home for choosing a question move
+and answer shape from `questionIntent`. `PROCESS.md` owns the conversational
+rhythm and evaluation; `SKILL.md` owns rendered-surface wiring; scrolls
+consume this palette directly. They point here instead of restating it.
+The byte-identical `.codex/skills/harvest/QUESTION_PALETTE.md` file is a
+packaging mirror for Codex and Pi; tests reject drift from this source.
+
+The palette is a decision aid, not a checklist. It never requires a mix or
+minimum number of shapes. An all-`single` surface is correct when every
+question is truly an enumerable fork; it is wrong when rankings, subsets,
+graded positions, or unmapped spaces were flattened into option cards.
+
+## Selection rule
+
+For every question:
+
+1. State the `questionIntent`: what the answer must reveal or change.
+2. Choose the closest intent row below.
+3. Use its move and answer shape when available.
+4. If another shape fits better, record one sentence explaining why.
+5. In a new rendered surface, declare `mode` explicitly. An absent mode
+   falls back to `single` only for backward compatibility.
+
+The author must be able to explain both why the question is being asked
+and why its answer shape fits. There is no shape quota or preset flow.
+
+## Intent → move → answer shape
+
+| Elicitation intent | Question move / answer shape | Guard |
+|---|---|---|
+| Decide among enumerable, mutually exclusive alternatives | Run the **enumerability test**: the author can state the full space. Use option cards, mode `single`, with 2–4 options, an “other, in your own words” escape, honest ± tradeoffs, and at most one argued lean. | If the full space cannot be stated, it is not a single-pick fork. |
+| Place a graded position or intensity between two honest poles | Name the evaluated dimension and both poles concretely. Use mode `spectrum` with 5–7 discrete, fully labeled stops; follow a midpoint with a clarifying probe. | Nothing is preselected. Untouched, midpoint, and UNDECIDED remain distinct. Never use a continuous slider. |
+| Set priority or sequence over a small set | Use mode `rank`; allow ties or top-k when a total order would invent precision. | Keep to at most 5 items. |
+| Compare importance across a longer list | Run **best-worst** as dialogue: show 3–4 items, ask most and least, then rotate subsets. | Do not turn a long list into a rating battery. |
+| Choose a non-exclusive subset | In an artifact, use mode `multi` with an explicit `max`. In live dialogue, walk items one at a time as explicit include / leave-out judgments. | Cap the subset and keep the candidate set small enough to inspect. “Select all that apply” without a cap invites shallow selection. |
+| Allocate relative magnitude or resources | Use mode `weight` with coarse 5–10% steps and an explicit untouched state. | Use at most 5 buckets; do not imply false precision or prefill an even split. |
+| Weight decision dimensions with different ranges | Ask a **swing question**: with every dimension at its worst, which swing to best would the respondent buy first? Continue against the concrete ranges. | Do not substitute context-free importance ratings. |
+| Elicit a quantity plus confidence | Ask, in order: lowest plausible, highest plausible, best guess, then sureness. Prefer frequency framing. | This needs the four-field interval card; until that renderer addition exists, run it conversationally. |
+| Explore an unmapped answer space, rationale, or discovery question | Use an open prompt anchored in the respondent’s own words. Acknowledge the answer, then probe only when it is insufficient. | Ask at most 1–2 probes per question; do not use a bare “why?” when a specific clarifier is available. |
+| Discover tacit taste dimensions before minting a fork | Run a **triad**: present three concrete elements and ask which two are alike and how the third differs. The named difference becomes a candidate axis. | Do not supply the taste taxonomy in advance. |
+| Reach tacit expertise behind a general claim | Run a **critical-incident sequence**: one concrete incident, a co-built timeline, deepening probes, then “what if X had been absent?” | Generalities are elicited through a real case, not requested directly. |
+| Deepen a thin why-note | **Ladder up** with “why does that matter to you?” or **ladder down** with “give me a concrete example.” | Stop after at most 3 consecutive ladder steps. |
+| Verify the interviewer’s interpretation | Run a **mirrored-inference check**: “My read of your position is X — what is wrong in that?” End a harvest with a member check of each respondent’s own synthesis when useful. | Mirror the sharpest reading, including tensions; do not smooth disagreement into affirmation. |
+| Get concrete feedback after abstract forks run dry, or when the respondent asks to see the work | Show only the relevant draft or excerpt in the decisions view and ask an open prompt naming what the answer will change. | This is selective authoring, not a mode: no default per-section controls, no generic ready/not-ready verdict, and no interaction added to the readable face. |
+| Elicit multi-respondent preferences, positions, or estimates | Default to blind independent collection. Freeze one shared artifact and question set when everyone should answer the same inquiry; otherwise adapt only from the seed, RoleSheet, and that respondent’s own prior turns. | Attribute every position and preserve its reasoning. Append pre-seal corrections and let the latest supersede; after sealing, a changed position starts a new round. Never report an anonymous aggregate such as “most of the team thinks.” |
+| Preserve honest nonresponse | Offer explicit `UNDECIDED`, distinct from untouched. | Never force an answer. A scroll may carry it once, then must reframe or demote the question. |
+| Handle an escape-hatch answer that names a missing option | At review, treat it as a candidate option or fork for the next version. | The creator decides whether it enters the pool; never silently coerce it into an existing option. |
+
+## Conversational moves
+
+These moves need no widget. They run inside PROCESS’s generate → evaluate
+rhythm:
+
+- **Ladder up / down** moves between value and concrete example.
+- **Critical incident** reconstructs one real episode before generalizing.
+- **Triad** lets the respondent name the comparison axis.
+- **Best-worst** extracts discrimination from rotating small subsets.
+- **Swing question** ties importance to the actual range of change.
+- **Mirrored-inference check** asks the respondent to correct the model’s
+  strongest interpretation.
+- **One-at-a-time forced choice** replaces an unbounded check-all with
+  inspectable include / leave-out judgments.
+
+## Hard bans
+
+These are lintable rules:
+
+- No agree/disagree, true/false, or yes/no stems — rewrite as explicit alternatives or an item-specific scale.
+- No draggable sliders with a preset thumb; no numbers-only or endpoint-only scales.
+- No batteries of stacked identical gauges; one question per fold-point, adjacent to its motivating prose.
+- Nothing pre-selected, ever — including spectrum position and weight splits; the lean argues, it never pre-picks.
+- No dropdowns in surfaces; >4 genuine options means an unsplit fork — split it.

package/runtime/claude/.claude/skills/harvest/SKILL.md

@@ -0,0 +1,100 @@
+---
+name: harvest
+description: "Run an adaptive harvest — directed elicitation that extracts, deepens, and synthesizes what people think about a topic. Use for 'harvest', 'run a harvest', 'align the team on', or a structured interview; not for a casual question or fixed survey."
+---
+
+# Harvest
+
+Elicit tacit preferences, positions, constraints, and knowledge, then preserve
+an attributed synthesis. A harvest maps choices; it does not force convergence.
+
+## When to invoke
+
+Run an adaptive harvest — directed elicitation that extracts, deepens, and synthesizes what people think about a topic. Use for 'harvest', 'run a harvest', 'align the team on', or a structured interview; not for a casual question or fixed survey.
+
+## Load the cognitive contracts
+
+Read these completely before questioning:
+
+- `.claude/skills/harvest/PROCESS.md` — adaptive elicitation rhythm
+- `.claude/skills/harvest/QUESTION_PALETTE.md` — intent, move, answer shape
+- `.claude/skills/harvest/FORMAT.md` — evidence-bound synthesis grammar
+
+Use `AUDIT.md` §14 only when a pasted decision-surface return requires its
+review grammar. Its historical graph/file persistence recipes are not an
+execution contract; Runtime adapters below own persistence.
+
+## Seed
+
+Parse topic plus optional `--respondents`, `--seed`, `--resume`, and
+`--mode blind|disclosed|comparative`. Reuse attached
+`EGREGORE_ORG_CONTEXT_V1`; do not repeat retrieval or reopen supplied sources.
+If a named gap remains, make one `bin/search.sh query` episode and open only
+the missing canonical evidence. Do not query infrastructure directly.
+
+Make low-confidence RoleSheet assumptions visible. In blind mode, other
+respondents' answer content neither appears nor shapes their questions. For a
+blind shared-artifact round, dispatch the frozen question set unchanged.
+
+## Elicit
+
+Apply PROCESS.md's seed → generate → evaluate → checkpoint → cascade →
+synthesize rhythm:
+
+- Ask one interaction at a time and always allow freeform correction.
+- Record a `questionIntent` and evidence-bound evaluation for each answer.
+- Watch for satisficing and sycophancy; name real tensions.
+- Offer checkpoints at natural transitions and respect “I'm done.”
+- Attribute disclosed positions. Never turn a majority into anonymous team
+  alignment.
+- While collection is open, a respondent may correct their answer. After the
+  declared completion condition seals the evidence, a changed position starts
+  a new round; there is no automatic second pass.
+
+For an absent respondent, create one authorized canonical question per turn:
+
+```bash
+bash bin/question.sh create \
+  --from "{initiator}" --to "{respondent}" \
+  --topic "{topic}" --question "{question}" \
+  --harvest-id "{harvest_id}" \
+  --harvest-session-id "{harvest_session_id}" \
+  --turn "{turn}" --question-intent "{intent}" \
+  --context-mode "{blind|disclosed|comparative}"
+```
+
+This records the pending question; it does not notify anyone. Delivery requires
+a separate action permission and exact recipient/message consent.
+
+## Decision-shaped rounds
+
+For three or more genuine interrelated forks, use the existing Meridian
+decision-surface renderer and QUESTION_PALETTE modes. Self surfaces may remain
+local. Do not publish, share, or deliver a directed surface implicitly. A
+separate permitted, explicitly consented action is required. Absorb returned
+answers into the evidence set; keep `UNDECIDED` open. Do not write graph or
+event-log projections directly.
+
+## Synthesize and write once
+
+Build the attributed layered synthesis from FORMAT.md. Every assertion must
+trace to a recorded answer or marked seed. Show the synthesis and wait for
+acceptance or edits. Then render one complete Markdown body to a temporary file
+and make exactly one canonical synthesis write:
+
+```bash
+bash bin/knowledge.sh create \
+  --type finding --subtype harvest-synthesis \
+  --title "{topic} — harvest synthesis" \
+  --input "$TMP_SYNTHESIS" \
+  --topic harvest --topic "{topic}" \
+  --workstream "{workstream}"
+```
+
+The Runtime owns ActorContext authorization, stable provenance, Markdown/Git,
+retrieval refresh, background embedding, and content-free telemetry. Do not
+run a second save, index, graph, publish, share, or notification action.
+
+Render the compact `⊙ HARVEST` completion with topic, respondents, disclosure
+mode, canonical path, and writeback status. Canonical Markdown wins over every
+optional projection.