Direction is a two-tier addressable projection: grill-`proposed` and Voz-`set` threads, folded from the log

The Direction standing page holds two tiers — proposed (the grill's strategic achados, hypothesis) and set (Voz-ratified, curated) — as addressable items with an explicit propose/set/drop lifecycle, both rendered, folded from the event log (ADR-0006). Artefatos declare candidate steers; the grill consolidates them into proposed. This refines ADR-0006's single direction.set blob and the CONTEXT.md Direction term.

Status

proposed (2026-06-06; Voz ratifies)

Context

ADR-0006 declared direction.set events folded to a page, but only one tier (Voz) and

last-write-wins on a single {plan} blob (eventlog.py:direction_at/project_direction).

The grill's second, load-bearing output is orientation — the strategic read, the next bet, the

decision not yet made (grill/SKILL.md). Today it dies in the transient handoff: the edge persists facts (clusters/curated) but not strategy. Voz named the gap: "estamos montando os fatos mas não estamos montando isso ainda."

A wholesale blob re-introduces evaporation-by-omission — a thread survives only if the grill

re-states it every consolidation; miss it once and it is silently gone. That is the exact failure the change exists to remove, reproduced one layer up.

Artefatos exist to move or confirm Direction (the house "decision not yet made" section), but they

are read-only on state (CONTRACT C1) — they cannot write Direction.

Considered options

Single-tier, Voz-only Direction; grill achados live elsewhere. Rejected: fragments one concept

into two homes; strategy stays homeless.

Two tiers as wholesale-rewritten blobs (proposed blob + set blob). Simpler, matches the

current tool. Rejected: evaporation-by-omission returns — the property we are trying to kill.

Two tiers as addressable items with explicit lifecycle. Chosen: a thread persists until

deliberately dropped; promotion and provenance are first-class.

Artefato writes direction.proposed directly. Rejected: two consolidators, and it skips the

harm-potential funnel — every report's open question would auto-become a steer.

Decision

Two tiers, mirroring hypothesis/curated. proposed (grill, hypothesis) and set (Voz, curated).

curado > hipótese: set outranks proposed. "A correção sempre ganha" — set is always Voz-originated.

Addressable items. Each thread has a stable id and an optional `kind ∈ {phase, priority,

constraint, thread}`. Events (asserted per ADR-0006 — direct Cypher, no LLM, replayable):

direction.proposed {id, body, kind, from_artefato?, relates_to?}
direction.set {id, body, supersedes?}
direction.dropped {id, reason}
Persist-until-dropped. A thread survives until an explicit direction.dropped. Never lost by

omission — the anti-evaporation property, the whole point.

Fold/replay per id. Last event per id wins; set outranks proposed for the same id;

dropped removes it. direction_at(t) folds to the as-of-then page (both tiers). Strategic versioning = replay.

The page renders both tiers — a Set section (ratified) and a Proposed section (open threads).
Artefato → Direction. An Artefato declares candidate steers in its own artefato.published

event: proposes: [{body, kind, relates_to?}] — open a fresh thread, or confirm/challenge an existing one. It never writes Direction itself. The digestion sweep (ADR-0008) auto-writes each candidate into the non-curated proposed tier — eagerly, even when ambiguous — so the tier is current at every dispatch entry. The candidate is optional — a knowledge-only Artefato proposes nothing and only deepens a cluster.

Who writes which tier. The proposed tier is populated permissively — by the sweep (artefato

candidates) and by the grill's live orientation; ambiguity is allowed to exist there, flagged. The set tier is the grill's alone: on in-session Voz ratification it promotes proposed → set (the same promotion it runs on knowledge). The grill's harm-potential funnel applies to curation, not entry: it promotes the high-harm and drops/merges the noise — "don't drain the hypothesis tier; consolidate what carries harm potential."

Obsoletes eventlog.py's last-write-wins {plan} blob. direction_at/project_direction rework

to a per-id fold and a two-tier render.

Consequences

The grill's strategic achados become durable, replayable, versioned — the gap closed.
Strategy leaves the transient handoff and lives in Direction (a fold); killing the handoff no

longer loses the steer (see ADR-0008).

Provenance is explicit (from_artefato): a steer traces to the Artefato that proposed it, even

after that blob is pruned.

Cost: id management and a richer event taxonomy than ADR-0006 sketched — kept to append-only JSONL +

per-id pure-function fold, no framework ("shed scaffolding toward a legible core").

Build (tracer-bullet, ADR-0006 — Direction projection first): rework eventlog.py to the per-id fold;

render the Proposed tier; add proposes to artefato.published and to grill_lint's agenda.

Consistent with ADR-0005 (render unchanged — it now projects a two-tier page) and ADR-0006 (asserted