S1 / S2 Axis cover, host loop, Hub compile

What the cameras own

Graded Axis map

Live ports (both eyes)

Recordings keep-up — the real break

Before: list_recordings_since returned 1 MQTT stub (size_bytes: 0). /fleet/recorder-inventory had 40 SSH chunks. That is Axis record/list lying.

Fix (no camera UI): host-mcp-resolver.js SSH inventory first — do not mint a one-chunk MQTT list. tools-call-handler.js SSH-merges if the host list is a stub, like recorder inventory.

What this hunt will not do

Continue — Host-only, still no camera UI

People sit parked · 12:45 AM CT · 10:22 AM CT MODEL_MISS · sidecar until IMU glass lands

People-count equal-result sit (12:45 AM)

Same ~30s walk. Host 01 occupancy stays isolated VLM — N/A here. Leftover dets_status.person_count is not PASS.

10:22 CDT walk (pulled via s2-via-jump)

Cloudflare jump only. Hop and tunnel not changed. Device jsonl from CID 65acd4e6-… chunks 1787498384–1787498627. The 15:22:00 stills in s2-1022/ are empty foyer. Those are the wrong covering. Fair subject is s2-010210-20260823T152245Z.jpg.

Overnight prep (what Grok sent / what is live)

Hop is live. iMessage already went to the phone. S2 is up after a runc revive — the overlay restart had left ghost_core/agent down (runc create “container already exists” until $ROOT/$CID was not pre-mkdir’d; leftover also lived under /tmp/runc-root). S1 was never bounced.

Live knobs vs 12:45 knobs

RF-DETR is not slower than YOLO here. Both infer at ~5 Hz. Roboflow’s Pareto vs YOLO11 still stands on paper. The 12:45 miss was recall + sit split + night doorway.

Why 12:45 missed (ranked, keep hunting)

1. Sit split. 30 s across two cameras. S1 got the close door. S2 clip after 12:45:16 is empty foyer. Equal-result needs 15 s planted in S2, then S1.
2. Night silhouette + 544 letterbox. Fewer pixels on a dark person than YOLO sentry on 15 fps encode. Not latency. Next sit: covering S2 JPEG must show the person at T0 or score FOV-miss.
3. Weapons-finetune transit ≠ sentry YOLO. S2 pack already counts large/close people (9:35). It did not fire on this night walk. Domain gap, not “RF-DETR slower.”
4. Sub-threshold was invisible. Core dropped p<0.25 before jsonl. Could not tell “saw at 0.18” vs “never saw.” Overnight: live thresh 0.15. Repo drop log only after next ghost_core bake.
5. camera_profile overwrite. Layer env patched to 0.15, then camera_profile.env still had 0.25 and loads after. Both files now 0.15. Always grep camera_profile after a layer patch.
6. Not leftover MQTT, hop kill, ObjectTracker, or Ultralytics AGPL track.

Equal-result bar (falsifiable)

What shipped overnight (for later discovery)

• Repo: ghost_core.cpp class-0 uses RF_DETR_PERSON_CONF default 0.15; other classes stay RF_DETR_CONF_THRESH. 1/s drop log c,p,thresh.
• Overlay agent: occupancy_from_dets person floor 0.15 class 0. Live on S2 /app-storage/agent/ghost_agent_v2.py.
• Live S2 env (layer + camera_profile): RF_DETR_CONF_THRESH=0.15 + RF_DETR_PERSON_CONF=0.15. Banner prints CONF_THRESH=0.15 (baked core does not print PERSON_CONF yet).
• Poller: scripts/s12-people-sit-poll.py 1 Hz hop + JPEG SHA every 5 s. Self-test PASS. PID docs/grading/s12-sit-morning/poller-arm.pid.
• Tests: overlay source 4/4. Spec docs/RF_DETR_PEOPLE_COUNT_EQUAL.md. Pack docs/grading/s12-sit-0045/.

10:22 AM daylight walk (poller scored after the fact)

The 3 h poller was not empty-aisle. At 15:22:35Z S2 JPEG shows a person in the left opening. At 15:22:45Z the same person is close, large, right of frame. Hop at those seconds stayed S2 rfdetr n_person=0 n_dets=0. A fresh health tick at 15:23:00Z is still 0. That tick can miss a 10 s walk on its own. It did not cause this miss. On-device occupancy never left 0.

On-device jsonl (jump SSH, scorer rerun): closest row to the fair JPEG is 15:22:45.628Z frame_seq 166496, dets=[], person_count=0, 54 ms later. Window ±1 s has 10 rows, 0 class-0 people. The only box in 1519 lines is 15:22:48.460Z class 0, p=0.20, box [0.0, 0.9, 0.1, 0.1] (area 0.01 sliver, jsonl person_count still 0). That is not the large person. Verdict: MODEL_MISS. Not FOV. Not HEALTH_LAG. Lowering 0.15 further would not have boxed the covering subject.

S1 n_dets hit 1 on 5634 rows with n_person still 0. Other-class clutter. Do not read YOLO n_dets as people. S2 n_dets never left 0 in 10761 samples. Health publish is 30 s. A 10 s walk can miss a hop tick even if NN boxed. Next sit must plant 15 s and check on-device dets jsonl, not hop health alone.

Discovery still open on the morning sit (parked)

• 15 s plant is parked with people-count. Score docs/grading/s12-sit-plant/s2-jsonl-sit.jsonl plus a covering still when that sit returns. JPEG person + jsonl dets=[] is MODEL_MISS again.
• After sit, jsonl dets[].p in 0.15–0.25 proves the live thresh helped. Drop-log lines only after the next core bake.
• S2 must stay class 0 for people. Bags at 0.15 may raise n_dets; they must not raise n_person.
• S1 encode 15 fps vs S2 5 fps is unfair for motion, not for a 15 s plant. Do not chase 15 fps on S2 unless thermal stays <78.
• 9:35 S2 two-person still stands as close-range PASS. Morning sit is the night/stand equal-result, not a redo of 9:35.

Overnight hunt finished at 6:03 AM CDT

Sight held. Host COTA soak still does not pass. Hop untouched — no oakctl, PoE, firmware, or kickstart.

Rail walk (once at 1:13 AM)

Failed 7 icons. Four of those (live-s1/s2, Shadow Play, Marks) were the harness demanding edge=live before Watcher had ticked — that check is already fixed. Still real: Watcher click did not spawn a Watcher agent; Dual DET toast lost to Library; Playlist MCP showed LOOK suppressed · watcher fail · edge_s2 · backoff-circuit.

Do not

Hub compile / DLC · then compiled hunt picture · 2026-08-23

Compile / DLC — official vs live vs refuse

Hop and fleet score the operator Qualcomm AI Hub compile — not leftover dets_status person. Overnight S2 hop nn_kind=rfdetr is the lab transit lane, not jp4me 544 weapons MET. S1 stays COUNT_NN / yolo_count. Do not Engage from this page.

Classifier: node scripts/snpe-dlc-loadability.mjs. Head+tail scan: mayEngage = dlc.metadata2.0.1 && backend.metadata0. SNPE 2.41 already failed Could not load the container. Letterbox 1280×720 encode → square 544 if jp4me ever loads. Overnight soak did not Engage. S1 never RF-DETR.

Two compile tracks (do not mix)

How a loadable recook is supposed to look

1. ONNX fixed square 544, outputs named dets, labels. Do not Hub-compile 672 as SoT (j5qzj2l75 is sibling only).
2. Hub job: qnn_dlc + HTP context graph. Prefer the recipe that produced transit’s dlc.metadata2.0.1 + backend.metadata0, not SNPE 2.41.
3. node scripts/snpe-dlc-loadability.mjs job_….dlc must print mayEngage: true.
4. S2-only copy. Point RF_DETR_MODEL_PATH and camera_profile.env (profile loads after sanitize and will reload qcs8550 if left alone).
5. One S2 runc. Hop: nn_kind=rfdetr + in-frame n. Then decide if official can flip — not before.

People-count equal-result does not require a new Hub compile. Transit lab pack + 0.15 conf + 15 s plant is the morning bar. New compile is the weapons-SoT lane.

Compile discovery still open

• Does a 544 recook of jp4me with metadata 2.0.1 + HTP load on this OS, or is 2.41-only Hub output a hard wall?
• Transit sha 1c86082e vs Hub 5ed0ac14 — same family, different SNPE metadata. A people sit on transit does not promote jp4me.
• After any Engage bounce, grep both layer env and camera_profile. Profile still loves qcs8550.
• Do not compile YOLO packs named “transit.” Do not git-add .dlc. Do not Engage 4287bb36 / 2.41.
• Until ghost_core is rebaked, jsonl p in 0.15–0.25 is the only proof the overnight thresh helped; drop-log lines will not exist yet.

Cite: docs/RF_DETR_OFFICIAL_S1_S2.md · docs/HANDOFF_S2_RFDETR_2026-08-22.md · docs/HANDOFF_RFDETR_WEAPONS_FINETUNE_2026-08-22.md · docs/RF_DETR_PEOPLE_COUNT_EQUAL.md.

Compiled hunts — one page

Specter is still the only glass. S1 and S2 have no camera webapp. Overnight they kept unique Sight for 4h 53m. Host COTA is still not production-ready. Compiling Axis + people sit + soak + Hub:

Compile — what is true this morning

Sources: this Axis hunt · docs/grading/rail-soak-2026-08-23.md · ticks 2026-08-23T06-10-55-759Z · IMU PR #82. Hop still LAN :8788. Jump s2-via-jump.

IMU / accelerometer on Specter glass · live mqtt · isolated fixture kept · 2026-08-23

What the operator sees

MotionHud already drew a 2D horizon and g-sensor from fixture roll_deg / forward_g. Live eyes publish RDF accel [x,y,z] m/s², not degrees. Glass now maps that packet at the telemetry boundary and paints the same HUD.

Path (smallest that works)

What we did not do

Aligned with prior hunts

• Specter is the only glass. No webapp on S1/S2. No VAPIX. No :8765 bind.
• Do not oakctl unless LED is blue / ghost_core down. Hop not kicked for IMU.
• People-count equal-result stays sidecar. S2 10:22 MODEL_MISS is not overwritten.
• Overlay LUKS format stays rejected. FIFO dated-glob stays code-only until camera copy.
• Jump s1-via-jump / s2-via-jump is the remote camera path. Tunnel not changed.

Tests: canvas IMU 21 pass. Hop resolver 12/12 including dets_status IMU without SSH.

LAN 2026-08-24 (this Mac owns :8788)

Operator back on LAN 2026-08-24 02:20 CDT. Hop /health ready, mqtt connected, snapshots 200 unique both eyes via LAN IP. Alias shadow1 503 stays honesty. Soak already PASS 6615 s. Probe lab_hil_ready true, host_production_ready true, blockers []. The 15:30–15:51 CDT plant window was NO_PLANT (hop mqtt 4 h stale, JPEGs 503). Night foyer stills now are empty. 3 h hop poller armed docs/grading/s12-sit-lan/. People-count stay sidecar. Do not oakctl. Do not a second soak.