EPIC-08 — Exploration first, governed by the fly

[superposition]

Steps

Phase 8 — Exploration first, governed by the fly (steps 23–24)

Step 23. Make exploration the default mission.
In runners/explore/src/main.rs, the run loop already asks QUALIA_COMPUTE_SOCKET to plan. Change
its default objective so the first mission opened at stack start is a frontier sweep: publish
MissionOpened { mission_id: "explore-frontier" } through the braid, request plans as today, and
include the fly prior's type-level in-strength in the planner request's existing belief_risk field —
uncertainty_weight from the prior's total applied weight, semantic_novelty from the fraction of
types in the chosen frontier that carry dimorphism = male-specific. Both are clamped to 0.0..=1.0
as belief_risk_penalty in runners/cuda-service/src/planner.rs already assumes.

Step 24. Keep leash's full capability set, governed.
Leash stays the motion authority: nothing in Phases 3–8 issues motor output. The harness reaches it
through QUALIA_LEASH_BASE_URL (already read by runners/agent). Add to
docs/contracts/planner-service-v1.md a one-paragraph note, and to crates/sync-types/src/mission.rs
a pub fly_governed: bool field defaulting to false (serde default, so older envelopes still
decode) which the agent sets to true when the prior is loaded. Test:
crates/sync-types/tests/mission.rs::fly_governed_defaults_false_and_round_trips.

Step 24b. Keep the sensing runners, and make time a belief input.
These runners stay and are not rewritten: runners/lidar, runners/camera, runners/vision
(object recognition; it fills WorldModel/voxels/lore and answers QuestionSlots),
runners/floor, runners/vslam, runners/pose, runners/map, runners/health. They are what gives
the harness leash's sensing capability set; this plan edits only pose and map (Step 21).
Time perception: qualia-jepa-model's RuntimeInput already carries delta_seconds: f32, so the
belief tick is already timed. Surface it — add pub fn decision_latency_ns(&self) -> u64 to
crates/braid/src/lib.rs (newest BeliefSlot timestamp in crates/types/src/lib.rs minus the newest
ledger entry's), and have Step 27's TUI Braid line render it as belief lag: {ms} ms. If a pose filter
(Kalman or otherwise) is expected, find it by reading runners/pose/src/main.rs before editing; if it
holds no filter, do not add one — pacing (Step 21) is the only change this plan makes to the
estimation path.

Clean-room decision in effect. The public repository is a from-scratch rewrite under
Apache-2.0. No file is copied from any private repository; the private engine is read only for
its interfaces, and the connectome dataset only for its graph. See docs/decisions.md.

Tickets

• T23 — T23 Exploration default mission + risk fields #38 Exploration default mission + risk fields
• T24 — T24 fly_governed in missions #39 fly_governed in missions
• T24b — T24b Sensing runners stay; time becomes a belief input #40 Sensing runners stay; time becomes a belief input

Journal

Entry slug: the-agent-in-the-loop — URL filled in when the entry is published.

[superposition]

All three tickets closed: #38, #39, #40. No entry is dedicated to this epic; its explored path is measured inside the-fly-in-the-belief-matrices (governance at the coupling bounds) and the-agent-in-the-loop (the pace and stale-publication windows: 250 ms × 8, 10,000–11,000 ms).