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

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

@@ -0,0 +1,436 @@
+# Harvest — Synthesis Format
+
+> Format is a grammar, not a template.
+>
+> The model chooses **which layers exist, which sections within them
+> exist, and which synthesis blocks to use** based on what the
+> harvest produced. Two harvests with different intents and different
+> respondents should not produce the same shape.
+
+This document is the **synthesis runtime contract** for `/harvest`. It
+is the companion to [`PROCESS.md`](./PROCESS.md). Where `PROCESS.md`
+defines how the model conducts the conversation, this defines how the
+model formats what was elicited so it becomes organizationally usable
+context.
+
+For the protocol's invariants, anti-conflation fixtures, and cognitive
+moves, see `PROCESS.md`. This document does not re-state them.
+
+For question intent, answer shape, probe moves, and hard bans, see
+[`QUESTION_PALETTE.md`](./QUESTION_PALETTE.md). This document formats
+what the harvest produced; it does not redefine how questions are asked.
+
+---
+
+## 1. What synthesis is and isn't
+
+Synthesis is a **deliberate act** of reading the running interpretive
+state, all turns across all sessions, and emitting a layered artifact
+that:
+
+- **Interprets with evidence.** Every assertion ties to a recorded turn
+  or marked seed. Quotes are verbatim and attributed. Patterns name the
+  turns they're extracted from.
+- **Layers disclosure.** Readable at multiple depths. The shortest read
+  is one sentence. The longest read is the full transcript.
+- **Stays in harvest's register.** Surfaces what was elicited;
+  organizes it for legibility. Carries no obligation to converge,
+  recommend, or pick a winner. Divergence is information.
+
+Synthesis is **not** a report card, a recommendation memo, a meeting
+minutes file, or a verdict. The boundary stated in `PROCESS.md` §1
+applies here: *a harvest maps the choices; it need not make them.*
+
+---
+
+## 2. The four layers
+
+Synthesis is structured as a progressive disclosure stack. A reader
+should be able to stop at any layer and have learned something
+complete at that depth.
+
+### L0 — Portrait
+
+**Always present.** The one-line finding: one sentence (sometimes two)
+carrying the sharpest finding, preference, or position the harvest
+produced.
+
+L0 carries no obligation to recommend or pick a winner. For different harvest shapes:
+
+- Preference harvest → the strongest preference, with the load-bearing
+  reason in a clause: *"Cem prefers card-first L0 expanding to
+  narrative+rail because the rail keeps decision surfaces visible
+  while the narrative does the work."*
+- Alignment scan → the central pattern across respondents: *"All
+  three converge on free-tier-as-conversion; they diverge on the
+  ceiling — and the divergence tracks how each models acquisition
+  cost."*
+- Knowledge extraction → the single most load-bearing thing the
+  respondent knew that wasn't documented: *"Renc has been carrying
+  three GTM constraints not yet on the canvas, and one of them
+  (legal-coupling) gates the launch sequence."*
+- Decision crystallization → the decision and its rationale: *"Split
+  Hermes into harness + runtime; the harness is vendor-replaceable,
+  the runtime is the differentiator we own."*
+- Tacit position → what the person was carrying: *"Oz reads the moat
+  as substrate-as-convention, not feature ownership — a position the
+  product language hasn't picked up yet."*
+
+L0 is read first. If a reader reads only L0, they have learned
+something true and useful.
+
+### L1 — Surface
+
+**Optional. Emit when the harvest produced anything decision-shaped or
+divergence-shaped.**
+
+L1 organizes findings at the harvest level (not per-respondent,
+not per-question). Sections that may appear, in any combination:
+
+| Section | When it appears |
+|---|---|
+| **Convergence** | Multi-respondent harvests where positions aligned on something worth naming |
+| **Divergence** | Multi-respondent harvests where positions split — presented as information, with each side's reasoning, not as a problem to resolve |
+| **Open questions** | Harvests that surfaced more questions than they answered (this is a valid outcome) |
+| **Action surface** | Optional and only when warranted; harvests with no proposition to act on may have no action surface. Records actions a respondent named or that a divergence implies — not actions the synthesizer invents. |
+| **Tensions** | Internal contradictions a single respondent surfaced — useful for self-harvests and deep elicitations |
+
+A pure preference harvest typically has no L1. A taste elicitation
+typically has no L1. An alignment scan usually has a thick L1 of
+convergence + divergence. An async multi-respondent harvest usually
+has an open-questions section. **The model picks which sections fit.**
+
+### L2 — Per-unit detail
+
+**Optional. Emit when there is per-unit content worth standing alone.**
+
+L2 is where the substance lives. The model chooses **which unit to
+slice on**:
+
+- **Per-respondent (position-unit)** — one section per person, each
+  containing their positions across the harvest's dimensions. Use
+  when the harvest produced person-distinctive positions. Olivier's
+  harvest is the canonical example: `Olivier on scaling`, `Olivier on
+  transmission`, `Olivier on bootstrapping`.
+- **Per-question (question-unit)** — one section per question, each
+  containing what was elicited about that question (one or many
+  respondents). Use when the harvest was structured around concrete
+  asks with concrete answers per question. The Oz canvas harvest is
+  the canonical example: `Q1 — Hermes framing`, `Q2 — sovereign brain
+  cards`, etc.
+- **Per-dimension (theme-unit)** — one section per theme/dimension
+  that emerged across questions and respondents. Use when the harvest
+  surfaced cross-cutting themes the questions didn't directly target.
+  Common in retrospectives.
+
+The model picks one slicing. Mixing slicings within one harvest is
+allowed only when the structure genuinely requires it (e.g.,
+dimension-unit for thematic findings + question-unit for concrete
+decisions). Default to one.
+
+Inside each L2 section, content is built from the **synthesis blocks**
+(§3). Section structure is generative, not templated.
+
+### L3 — Transcript
+
+**Always present (appendix).** The full turn log.
+
+For each turn:
+
+- Question (verbatim)
+- `questionIntent` — the model's articulable purpose
+- Answer (verbatim, or pointer to async file)
+- `evaluation` — the model's read (novel framing, unresolved tension,
+  tacit content, shallow signal, strong conviction, surprising
+  connection, diminishing returns)
+
+L3 is the auditability layer. Anyone reading L0–L2 can drop to L3 to
+verify the chain: which assertion came from which turn, what the
+model intended when it asked, how the model read the answer. This is
+also what makes the synthesis reproducible in fixtures (see §6).
+
+---
+
+## 3. Synthesis blocks
+
+Within L2 (and sometimes L1), content is built from these six blocks —
+the building material of the synthesis document. (Not to be confused
+with answer **modes**, which are the input formats of rendered surfaces
+— `single | multi | rank | weight | spectrum`;
+`QUESTION_PALETTE.md` owns intent-to-shape selection, while SKILL.md
+owns their rendered data fields.) They compose; sections pick whichever
+fit.
+
+| Block | What it is | When it carries weight |
+|---|---|---|
+| **quote** | Verbatim words from a respondent, attributed by handle | Distinctive framing, strong conviction, language worth carrying as-is |
+| **context** | What the synthesizer brings to make the quote legible — background, what was happening before, what assumption the respondent was responding to | When the quote is sharp but its meaning depends on context the reader may not have |
+| **reflection** | Interpretation: what this position *does* to the broader picture; what mental model it implies | When a position has implications beyond what the respondent said directly |
+| **choice** | A concrete choice or commitment a respondent named, or a choice a divergence implies. **Has a rendered form**: when choices are forks with 2–4 distinct options, `/harvest` emits them as an interactive Meridian *decision surface* the respondent decides *on* and pastes back, rather than prose to read | When the harvest produced something actionable; *not* the same as a recommendation — it surfaces the choice without forcing its resolution |
+| **open question** | An unresolved tension, ambiguity, or follow-up worth a future harvest | When the harvest opened more than it closed; explicitly named so it doesn't get lost |
+| **pattern** | A knowledge / action / antipattern the harvest extracted, ready for `/reflect` to consume | When a finding generalizes — not just *"Oz said X"*, but *"the pattern is Y, evidenced by X"* |
+
+**Composition rule**: for each L2 section, a strong default sequence
+is `quote → context → reflection → pattern` (the Olivier shape). The
+Oz shape (`decision → quote → reflection → action`) is also strong.
+The model chooses what the content needs. **Skip what doesn't apply.**
+A section with no `choice` and no `pattern` is fine — the blocks are
+a palette, not a checklist.
+
+**Block usage rules**:
+
+- `quote` — verbatim only. No paraphrasing dressed as quotation.
+  Block-quote format, attributed by handle.
+- `context` — clearly distinct from interpretation. Describe what
+  was, not what it implies.
+- `reflection` — owned by the synthesizer, not the respondent. Use
+  *"this implies..."*, *"the move underneath this is..."* — never
+  attribute interpretation back to the respondent.
+- `choice` — describe the choice, not which side to pick
+  (unless a respondent named the side themselves). **Rendered form:**
+  when ≥3 interrelated forks each have 2–4 distinct options, `/harvest`
+  renders them as an interactive *decision surface* instead of prose.
+  Each option carries a structural *visual*, honest `+/−` tradeoffs,
+  and at most one recommendation; the respondent's pasted choices
+  (`#{slug}-decisions:v1`) are absorbed on `--resume` through the author review gate: a non-author return is pending until accepted, declined with a required attributed reason, or synthesized and approved. Shape + render:
+  harvest's rendered-surface section (SKILL.md) and
+  `packages/egregore-artifacts` (`decision-surface` type).
+- `open question` — phrase as a question or as *"unresolved: …"*.
+- `pattern` — name the pattern type (knowledge / action /
+  antipattern) and cite the turn(s) it was extracted from.
+
+---
+
+## 4. Generative emission — how the model picks shape
+
+Synthesis shape is decided once the harvest is done, not in advance.
+The model reads the running interpretive state and answers four
+questions:
+
+1. **Did the harvest produce decision-shaped or divergence-shaped
+   findings?** If yes → emit L1. If no → skip L1.
+2. **What's the natural unit?** Respondent? Question? Theme? → that's
+   how L2 slices.
+3. **For each L2 section, which blocks does the content need?**
+   Pull only those.
+4. **Where would a reader want to go next?** That's where to put
+   `open question` blocks and `choice` blocks — at
+   the points the document hands off to action.
+
+A cleanly emitted harvest:
+
+- L0 always.
+- L1 if and only if the harvest produced something decision- or
+  divergence-shaped that's worth surfacing at the harvest level.
+- L2 if and only if there's per-unit content that stands alone.
+- L3 always.
+
+A bad emission emits all four layers regardless of content. Empty L1
+sections, padded L2 sections, and L0 sentences that don't actually
+hold the most are the failure mode.
+
+---
+
+## 5. Orientations by harvest shape
+
+These are orientations, not templates. The model adapts.
+
+### Preference / taste
+
+Likely shape: **L0 (the strongest preference + reason) · no L1 · L2
+per-respondent or per-question with `quote → context → reflection`,
+no `choice`, no `pattern` for the harvest itself (patterns
+emerge later when many preferences accumulate) · L3.**
+
+Example L0: *"Cem prefers card-first L0 expanding to a narrative+rail
+workspace, because the rail keeps decision surfaces visible while the
+narrative does the work."*
+
+### Alignment scan (multi-respondent)
+
+Likely shape: **L0 (the central pattern across respondents) · L1
+(convergence + divergence, divergence presented as information) · L2
+per-respondent showing each position's reasoning · L3.**
+
+Divergence sections in L1 lay positions side-by-side. Each position
+gets its own load-bearing reason. The synthesis does not pick a
+winner — but it can name what each position is optimizing for, which
+is often the most useful read.
+
+### Knowledge extraction
+
+Likely shape: **L0 (the most load-bearing undocumented thing) · L1
+optional (often emits as `open questions` if extraction surfaced
+gaps) · L2 per-dimension with `quote → context → pattern` heavy ·
+L3.**
+
+The `pattern` block earns its keep here. Knowledge extraction's whole
+point is producing reusable knowledge nodes.
+
+### Decision crystallization (the one shape that may end with an action surface or commitment)
+
+Likely shape: **L0 (the decision + load-bearing rationale) · L1
+(considerations weighed, commitments) · L2 per-question with
+`decision → quote → reflection → action` (the Oz shape) · L3.**
+
+This is the shape whose intent was to crystallize a decision, so an
+action surface or named commitment is appropriate. Even here, the
+model records the choice as the respondent(s) made it — it does not
+manufacture an action the respondents didn't name.
+
+### Tacit position elicitation
+
+Likely shape: **L0 (what the person is carrying that the org hasn't
+articulated) · no L1 · L2 per-dimension with `quote → context →
+reflection → pattern` · L3.**
+
+The Olivier harvest is the canonical example. The output is a portrait
+of someone's position, not a verdict on whether they're right.
+
+### Retrospective
+
+Likely shape: **L0 (the shared narrative or its absence) · L1 (shared
+narrative section, divergent interpretations section, lessons,
+commitments) · L2 per-theme · L3.**
+
+Useful pattern: when respondents agree on what happened but diverge
+on what it meant, L1 names this shape directly — *"shared event,
+divergent reading"* — and L2 explores each reading.
+
+---
+
+## 6. Fixtures
+
+Two harvests in the repo are gold examples. The rebuild's synthesis
+must reproduce their shapes (within minor variation) given the same
+intent + RoleSheets + seed.
+
+### Olivier harvest — position-unit slicing
+
+`memory/knowledge/harvests/2026-03-23-olivier-harvest-analysis.md`
+
+What it demonstrates:
+- L0 *"It makes it possible to embody the overall organization in a
+  variety of personal ways, while sharing a common ground with
+  others"* — single sentence that holds the most.
+- L2 sliced per-position: 7 positions, each as `Context → Olivier's
+  Position (with quotes) → Reflection → Pattern Extraction`.
+- Patterns named explicitly with type prefix (*Knowledge pattern —
+  Junction Primacy*, *Action — Reframe positioning language*).
+- L3 as `Full Exchange` with 7 turns, question + answer per turn.
+- No `choice` block used — this was a tacit-position
+  harvest, no decisions were on the table.
+
+### Oz canvas harvest — question-unit slicing
+
+`memory/handoffs/2026-05/06-cem-harvest-oz-canvas-tech-and-capabilities.md`
+
+What it demonstrates:
+- L2 sliced per-question: 7 questions, each as `1. Decision · 2.
+  Positions (block quotes) · 3. Why · 4. Action space`.
+- The L2 section's own internal table of contents (lines 142–159
+  describe how to read the document) is a generative move worth
+  copying when the harvest has many questions and a synthesis-ready
+  summary section.
+- The `choice` block is heavy here — this was a
+  decision-crystallization harvest with concrete asks.
+- Cross-question consolidation in a `Canvas Changes & Additions`
+  section is a domain-specific L1 emission. Generalizable: when L2
+  produces many small actions, an L1 consolidation section that
+  collects them is a useful surface.
+
+**Both are valid synthesis shapes.** The model picks the slicing that
+fits what the harvest produced.
+
+---
+
+## 7. Evidence binding (mandatory)
+
+Every assertion in synthesis traces back. The mechanism:
+
+- **Quotes** carry the respondent handle: `> "..." — oz` or block-quote
+  with attribution line.
+- **Patterns** name the turn(s) they were extracted from. Either inline
+  (*"extracted from turn 3"*) or implicit when the L2 section's quote
+  block is the source.
+- **Reflections** that draw on multiple turns name the turns they
+  synthesize.
+- **Seed-context references** are marked: *"per `whitepaper-v0.1.md`
+  §3"* — so a reader can distinguish between *what the harvest
+  produced* and *what was already known going in*.
+- **Async answers** are quoted with their source noted: *"from
+  `memory/knowledge/questions/2026-05-07-pricing.md`, answered
+  2026-05-08"*.
+
+If an assertion in synthesis cannot be traced, it does not belong in
+the artifact. (`PROCESS.md` §2 invariant 4.)
+
+---
+
+## 8. Linkability — `/reflect`-able output
+
+Synthesis is not a leaf. Patterns, preferences, positions, and
+(where present) decisions emit in shapes the org's pattern/decision
+graph can ingest:
+
+| Block | Becomes a |
+|---|---|
+| `pattern` (knowledge type) | Knowledge pattern node, linkable to topics + people |
+| `pattern` (action type) | Action node, linkable to quests / decisions |
+| `pattern` (antipattern type) | Antipattern node |
+| `choice` (where a respondent named a side) | Decision node, attributed to the respondent |
+| Synthesized position from L2 (e.g., *"cem prefers X because Y"*) | Position / Preference node, attributed to the respondent, with quote-block(s) attached as `Evidence` nodes — never the quote alone as the position |
+| `quote` block (verbatim, attributed) | `Evidence` node — raw, uninterpreted, links into the Position / Preference / Pattern it supports |
+| `open question` | Open-question node, surfacable in `/activity` and seedable for follow-up harvests |
+
+**The quote-vs-position distinction is load-bearing.** A quote is
+evidence; the synthesized position is the node `/reflect` consumes.
+This keeps the pattern/decision graph from treating every verbatim
+utterance as already-interpreted knowledge, and it makes the
+synthesizer's interpretive work explicit (and contestable) rather
+than disguising it as the respondent's own words.
+
+Concretely: when the synthesis is written, a `/reflect` pass can read
+the artifact and emit graph/memory nodes from each load-bearing
+section without re-interpreting. The artifact's structure carries the
+graph's structure.
+
+For a pure preference harvest, this means the org gains a
+tacit-knowledge node (*"cem prefers X because Y"*) other agents can
+reason from. For an alignment scan, it gains a divergence-map node.
+The shape varies; the legibility doesn't.
+
+---
+
+## 9. Negative examples — what synthesis does not do
+
+| Anti-pattern | Why it's banned |
+|---|---|
+| L1 with empty `Action surface` because the template said so | Forces an action the harvest didn't produce; violates §1 |
+| L2 sections with all six blocks padded in regardless of content | The blocks are a palette, not a checklist |
+| L0 that summarizes the dimensions covered (*"this harvest explored A, B, and C"*) | L0 is the sharpest finding, not a table of contents |
+| Paraphrased "quotes" | The `quote` block is verbatim only |
+| Reflections attributed back to the respondent | Reflections are owned by the synthesizer; respondents own quotes |
+| Divergence framed as "the team disagrees" | Divergence is information, not failure; name what each side is optimizing for |
+| "Recommended direction: X" when no respondent named X | Synthesis surfaces what was elicited; it does not invent direction |
+| Synthesis without L3 | Removes auditability; breaks `PROCESS.md` §2 invariant 4 |
+| Mixing per-respondent and per-question slicing in L2 without reason | Pick one; mix only when structure genuinely requires it |
+
+---
+
+## 10. What this document does not cover
+
+- **Cognitive protocol** — how the model conducts the conversation,
+  evaluates answers, runs multi-respondent disclosure: see `PROCESS.md`.
+- **Question palette** — intent-to-shape selection, probe moves, and
+  hard bans: see `QUESTION_PALETTE.md`.
+- **Persistence** — graph schema, ID conventions, async question
+  metadata, local-mode markdown contract: see `AUDIT.md` §9–§11.
+- **Orchestration** — argument parsing, mode detection, ID minting,
+  file writes: belongs in `SKILL.md` once it's rewritten as a thin
+  entry point.
+
+The synthesis runtime contract is complete here. The model has what
+it needs to produce a layered, evidence-bound, generative artifact
+that preserves what the harvest elicited without forcing it into a
+shape the harvest didn't produce.