Sigil: Distill

• an idea, architecture, model, plan, or workflow is too broad to implement responsibly,
• the user wants planning before committing to one solution,
• a solution may be overbuilt, underbuilt, or prematurely optimized,
• the work needs a clear relationship between broad concept layers and the first coherent unit,
• multiple possible designs should compete before convergence,
• the output needs a navigable next route such as implementation-layering, robot-talks, decision-gate, invoke design, invoke plan, or task-session.

• the user already has one small deterministic edit,
• direct implementation is safer than conceptual decomposition,
• the request only needs a summary or explanation,
• there is no meaningful output artifact to optimize toward,
• the user needs factual discovery as the primary work rather than planning optimization.

• seed point: starting concept, model, architecture, design, plan, problem, or workflow,
• target context: the size and purpose the result must serve,
• output artifact: data model, architecture design, implementation plan, code structure, decision record, research map, technique spec, or other concrete artifact,
• optimization goal: clarity, scope, architecture, planning depth, model quality, implementation readiness, or another stated goal,
• constraints: time, cost, quality, governance, risk, implementation, audience, or domain limits,
• existing artifacts: specs, notes, code, diagrams, plans, or decisions that should count as evidence,
• budget preference: Compact, Standard, Tournament, Deep, or Validate.
• optional evidence context: accepted event schema, append-only event ledger, and observed-ledger digest when downstream handoff is evidence-gated.

I understand the design intent as: <seed point>.
Target context: <context size and purpose>.
Expected output artifact: <artifact shape>.
Optimization goal: <goal>.
Recommended budget: Standard - one proposal track, Proposer and Balancer roles,
two recursive rounds, then reconciliation.
Do you want Compact, Standard, Tournament, Deep, or Validate?

If the user does not choose, proceed with Standard and record that assumption.

All modes must keep finite proposal tracks, finite recursive rounds, cycle guards, and skipped-technique reasons.

Both paths must preserve the same role trace:

• Proposer claim,
• evidence or assumption,
• Balancer objection category,
• reconciliation decision,
• stable disagreement, if one remains.

Technique activation trace must include:

• technique id,
• hook,
• activation reason,
• inspected state,
• emitted output,
• decision,
• readiness effect.

• the recursive round budget is reached,
• the same split reappears with new names but no new structure,
• a round adds terminology without improving closure, recomposition, or risk handling,
• Proposer and Balancer keep trading the same tension without new evidence,
• a smaller concept fails closure but is repeatedly reintroduced,
• further reduction would damage the user-selected context.

When a guard triggers, record the reason and either choose the current best optimization point or ask for one blocker decision.

Complexity is allowed only when the current context has:

• a named tension,
• a concrete failure mode, or
• a confirmed evolution pressure that the simpler unit cannot responsibly handle.

Before deferring future scale, ask what kind of evolution the system, solution, or plan is likely to have. Natural evolution pressure can justify a small extension boundary when it is concrete: expected variants, repeated integrations, growing policy rules, multiple actors, scaling volume, migration needs, or governance review.

When the evolution profile is unknown, preserve a clear boundary and defer the heavier mechanism.

• confirm objective and output artifact before recursive decomposition,
• select or infer a finite mode budget,
• build a discovery baseline before proposing layers,
• preserve Proposer/Balancer trace,
• classify Balancer objections by named category,
• run always-on techniques or record why readiness is downgraded,
• activate conditional techniques only when their trigger is present,
• stop infinite reduction through cycle guards,
• select a smallest coherent unit with closure and recomposition proof,
• defer complexity that lacks a named tension or confirmed evolution profile,
• return a navigable result with start-here guidance, unresolved tensions, and next route.

• reducing concepts into tiny fragments that no longer carry behavior,
• adding abstractions for elegance, reuse, or hypothetical scale,
• letting the Balancer object vaguely without a concrete category,
• treating a concept claim as knowledge-backed when evidence is weak,
• skipping the output artifact because the concept map feels complete,
• running Tournament mode without assumptions and elimination conditions,
• letting true-subagent and role-simulation paths produce different trace contracts,
• folding an invoked Distill execution only into caller telemetry,
• emitting child telemetry for a skipped or not-required Distill route,
• appending more than one Distill signal for the same child run,
• claiming complete evidence emission when the accepted resolver cannot close the full role/process boundary sequence,
• allowing runtime-event or telemetry evidence to set the Distill verdict or mutation readiness,
• claiming pass readiness when the result has no navigation guide,
• routing to implementation when a blocker decision or cross-layer tension remains.

Recommended signal fields:

• child run id and direct or invoked-by lineage,
• parent run id plus caller capability and mode when invoked by another capability,
• objective-output confirmation,
• target context,
• selected mode,
• proposal tracks,
• recursive rounds,
• role execution path: true subagents or role simulation,
• techniques triggered and skipped,
• verdict,
• objective-output drift,
• navigation closeout status,
• execution-evidence status and available request, runtime-event, receipt, and validator references,
• evidence-emission status: complete, partial, failed, not-required, or not-configured,
• telemetry append status,
• next route.

For a meaningful direct Distill execution, Distill owns the post-run append through scripts/observe-direct-invocation.sh. A direct envelope has invocation_source: direct, no caller lineage, and records evidence-emission status under evidence.emission_status. Signal Observer deduplicates the row by the direct Distill run ID.

When another capability invokes Distill, Distill must prepare a child invocation envelope with its own run ID. The caller owns the post-run append through signal-observer and must link the row to the caller run with lineage.relation: invoked-by. A completed, partial, blocked, or failed meaningful Distill run produces one Distill telemetry row in addition to the caller's row. A skipped or not-required Distill route produces no child row; the caller records the skip rationale instead.

Telemetry and runtime-event evidence remain non-authoritative. Missing or failed telemetry is visible residue and cannot change the Distill verdict or authorize downstream mutation.

## Distill Result

- Target context: <context summary>
- Objective and output artifact: <objective; artifact shape>
- Mode and budget: <compact | standard | tournament | deep | validate>
- Proposal tracks: <count and role summary>
- Recursive rounds: <count completed / budget>
- Verdict: pass | flag | block
- Role conversation trace: <Proposer claims, Balancer objections, reconciliation decisions>
- Current smallest coherent unit: <unit name and responsibility>
- Optimization point: <why this unit is the best size for the target context>
- Concept layer map: <broad layer to selected unit>
- Technique pack trace: <techniques run, skipped, triggered, and outcomes>
- Closure and recomposition proof: <how the unit closes and recomposes upward>
- Evolution profile: <expected evolution and smallest extension boundary>
- Deferred complexity: <what was deferred and why>
- Tension ledger: <resolved and unresolved tensions>
- Premortem: <likely failure reason and guardrail | skipped with reason>
- Frame-expiry note: <context change that invalidates this optimization point>
- Navigation guide: <where to start, what changed, what remains unresolved, and how to use the result>
- Evidence emission: <complete | partial | failed | not-required | not-configured; evidence references or diagnostic>
- Telemetry: <recorded | failed with residue | not configured; direct or caller-owned>
- Next route: implementation-layering | robot-talks | decision-gate | invoke design | invoke plan | task-session | deferred