The briefing is corpus + Recap; edge work carries mandatory intent and a two-tier, mechanical source-feedback

The edge presents a briefing to orient the agent at every dispatch — three projections: Knowledge clusters (← graph), Direction (← log), and a new Recap (← corpus). The corpus is the edge's own created content (a fold of artefato.published); the Recap projects it, synthesized fresh against the mentee's current Atividade. Edge work carries mandatory intent (CONTRACT C3) plus a minimal metadata schema, and source feedback is a two-tier (mechanical hypothesis → grill-curated) relevance signal over Mundo and Atividade. Grounded in a 4-source 2026 survey.

Status

proposed (2026-06-06; Voz ratifies)

Context

• The sweep (ADR-0008) digests the mentee's Atividade but excludes autonomous beats — so the edge

had no leg reading its own work: no self-reference for continuity, and no way to relate what it did to what the mentee is doing. (Voz: "the next leg is summarizing the agent activity.")

• "State digest" (the assemble hand-up) was a mechanical, unnamed brief; briefing.md / exp-009 were the

lineage. Voz reframed it: the briefing is the collection presented to orient the agent, in parts.

• The intent kernel lost its durable home when ADR-0008 dissolved consolidate — and autonomous-beat

transcripts are excluded from the sweep, so the why could vanish entirely.

• Research (4 sources, 2026; X · HN · arXiv · exa): intent is the most-neglected yet highest-value

field (STITCH arXiv:2601.10702 uses an intent tuple as the primary retrieval index; Sayou's SAMB benchmark: decision-reasoning recall 67% structured vs 18.5% embeddings). Source-feedback is best done mechanically (AgentOS RetrievalFeedbackSignal detects used-vs-ignored post-hoc; MemQ arXiv:2605.08374 propagates outcome credit through the provenance DAG). Artifact→human-outcome credit is unbuilt across all systems surveyed.

Considered options

• Keep "State digest" / a flat dump. Rejected: no structure, no self-reference, no corpus.
• Corpus grown from beat logs/blog outside the log. Rejected: a second source of truth vs ADR-0006 —

the corpus is a fold of the log's artefato.published, a new lens, not a new store.

• Source-feedback by agent self-rating. Rejected: cognitive burden and unreliable (the field's

lesson). The agent must carry no scoring load.

• Source-feedback purely mechanical (overlap). Partial — a noisy proxy; adopted **as the hypothesis tier

only**, refined by outcome credit and curated by the grill.

Decision

• Briefing — the orientation presented at every dispatch (and /load): three projections —

Knowledge clusters (← graph) · Direction (← log) · Recap (← corpus). Composed by assemble; supersedes "State digest". The handoff (intent-kernel breadcrumb) is one pragmatic input, not the briefing.

• Corpus — the edge's own created content; a fold of artefato.published events. Per-install (own

group), the reflexive complement to Mundo/Atividade/Voz; autonomous beats land here, not in the sweep.

• Recap — a projection of the corpus: the agent's recent steps + their why, correlated **fresh at

compose-time** to the mentee's current Atividade. Not a stored link (orientation must be current; the stored cites/distills are provenance, not the relation).

• Mandatory intent (C3) — every dispatch that produces an Artefato emits an intent.kernel event at

close: the why the corpus carries and the Recap projects. Closes the kernel-home gap left by ADR-0008.

• Artefato metadata schema (the consensus core, four fields):

{ intent, cites:[{ref, kind: mundo|atividade, relevant}], distills:[cluster:…], proposes:[{body, kind, relates_to?}] }. impact is derived from proposes at compose-time, not stored (the field has not solved outcome-credit; keep minimal).

• Source feedback — two-tier (the edge's universal hypothesis→curated shape; one curation act — the

grill — now governs Knowledge, Direction, and Sources):

• hypothesis (mechanical, no agent/mentee load): intrinsic citation (the agent names sources as part of

writing a cited Artefato) + outcome credit (Voz ratification of a proposes / engagement, propagated back through cites, MemQ-style) + retrieval-use detection (RetrievalFeedbackSignal). Never self-rating.

• curated (grill + Voz): the grill distills the mentee's reasoned source opinion ("values X in

reports because Y"), re-ranks the Source roadmap with curated authority, exempt from passive aging, retirable only by Voz.

• Spans Mundo and Atividade — "this commit was relevant to this report."

Consequences

• The edge gains a reflexive leg (corpus) and self-reference for continuity; the briefing **relates the

edge's work to the mentee's live work** — the mentoring payload, not just a state dump.

• Intent becomes durable and load-bearing (C3) — ahead of the shipped field, where intent is the

most-neglected field.

• Source value is learned mechanically + outcome-grounded (zero cognitive burden), and the mentee's

reasoned preference is curated — closing the artifact→human-outcome loop the 2026 survey found unbuilt.

• Consistent with ADR-0006 (all briefing parts are projections/folds of the log), ADR-0007 (two-tier,

grill-curated — now generalized to a third object), and ADR-0008 (the sweep; intent.kernel closes its gap).

• Cost: a corpus fold + a Recap synthesis at compose-time (one LLM call in assemble); a new intent.kernel

event type; the source-feedback outcome-credit wiring. Build is a follow-up (new issue), extending the spine after #9/#13/#15/#16/#14: intent.kernel emission → corpus fold (corpus_at + state/corpus.md) → Recap composition in assemble → source-feedback credit (citation + outcome) + the grill's curated tier.

• Open: the mechanical hypothesis tier's overlap heuristic is a proxy; the outcome-credit path is the

trustworthy signal, but depends on proposes ratification / engagement actually happening (sparse early).