{
  "_readme": [
    "A DEVICE PROFILE: what this model of robot can do, what it cannot, what had to",
    "be built, and what will bite the next person. The format and the reasoning are",
    "in docs/superpowers/specs/2026-08-19-device-onboarding-routine.md.",
    "",
    "THE RULE THAT MAKES THIS SAFE: a profile contains declarations and",
    "human-readable procedure. IT NEVER CONTAINS EXECUTABLE CODE. We cannot enforce",
    "that -- a profile is a file, a robot is somebody's machine, and there is no",
    "runtime we control on either -- so the design does not depend on enforcement.",
    "The worst case of a hostile profile is a WRONG CLAIM, which self-declaration",
    "already tolerates and a dissatisfied agent already corrects. A profile that",
    "could ship a script to your robot would be a supply-chain attack with a",
    "friendly name, on a machine with motors.",
    "",
    "So this may NAME a driver and DESCRIBE what to build. It may not carry either.",
    "And the results below are evidence about the unit they were measured on: run",
    "the probes on your own unit before trusting any of it."
  ],
  "identity": {
    "vendor": "Yahboom",
    "model": "Raspbot V2",
    "vendor_name_on_disk": "PI5Car",
    "recognise_by": {
      "/proc/device-tree/model": "Raspberry Pi 5",
      "arch": "aarch64",
      "os": "Debian 12 (bookworm)",
      "i2c": "an expansion board ACKing at 0x2B on bus 1",
      "camera": "a UVC webcam on /dev/video0 offering MJPG 1280x720",
      "vendor_workspaces": [
        "/root/yahboomcar_ws -- THE DRIVER",
        "/root/yahboomcar_ros2_ws -- the tutorials"
      ],
      "audio": "Superior ONLY: a C-Media USB Audio Device, USB 0d8c:0012, appearing as ALSA card 2 with one capture device. Absent on the standard unit.",
      "speech_board": "an offline speech board on /dev/ttyUSB0 (CH340, USB 1a86:7522). It is the voice module, NOT a lidar."
    },
    "note": "The two workspace names differ by four characters. One session opened the tutorials, never opened the driver, and published a conclusion about the whole product from the half it had seen. Read the vendor's own course listing first; it names both in a minute.",
    "variant": "Superior",
    "variants_exist": "THE STANDARD RASPBOT V2 HAS NO AUDIO HARDWARE AT ALL; the Superior does. Same vendor, same model name, different machine -- which is precisely why recognise_by exists and why matching on a model name is unsafe. A profile claiming audio would be a lie on a standard unit, and a standard unit's profile would understate a Superior. Distinguish them by the C-Media USB audio device below, not by what the box said."
  },
  "provenance": {
    "characterised_by": "Claude Opus 5, with the owner operating the robot",
    "characterised_on": "2026-08-18/19",
    "unit": "one robot, one lens, one SD card",
    "firmware": "the expansion board's register map was read from the vendor's Raspbot_Lib.py on this machine; no firmware version is exposed by the board"
  },
  "claims": {
    "_comment": "Exactly the manifest body/stringer_body/adapters/raspbot.py:make_driver declares FOR THE DEPLOYED BUILD (deploy/start.sh exports STRINGER_ROBOT_OCR=1). tests/test_profile.py asserts these agree AND that start.sh really exports it, so this cannot quietly become a description of a robot that no longer exists.",
    "verbs": [
      "observe",
      "inspect"
    ],
    "consent_tiers": [
      0,
      2
    ],
    "free_text_steerable": false,
    "sensing": {
      "classify": "basic",
      "count": "basic",
      "ocr": "basic"
    },
    "expressive": {
      "motion": [
        "spin",
        "strafe",
        "diagonal",
        "shimmy"
      ],
      "light": true,
      "sound": true
    },
    "routines": {
      "observe": {
        "offers": {
          "static": {},
          "sweep": {
            "dwell_s": [
              0.5,
              10
            ],
            "tilt_deg": [
              -20,
              110
            ],
            "pan_deg": [
              -59,
              59
            ]
          },
          "dwell": {
            "duration_s": [
              5,
              300
            ],
            "tilt_deg": [
              -20,
              110
            ],
            "pan_deg": [
              -59,
              59
            ]
          },
          "wander": {
            "duration_s": [
              10,
              300
            ]
          },
          "flourish": {
            "move": [
              "nod"
            ],
            "duration_s": [
              1,
              20
            ]
          },
          "blink": {
            "duration_s": [
              2,
              60
            ]
          },
          "chirp": {
            "duration_s": [
              0.5,
              10
            ]
          }
        },
        "default": "static"
      }
    },
    "mobility": {
      "traverse": "free",
      "obstacle_avoidance": "ranged",
      "bounded_by": "none",
      "live_steer": true
    },
    "form": "wheeled",
    "autonomy": {
      "localisation": "none",
      "navigation": "reactive",
      "semantic": "none"
    },
    "name_detection": {
      "level": "none"
    },
    "x_constraints": {
      "kind": "indoor_robot",
      "map_bound": false,
      "allow_stranger": false
    }
  },
  "what_the_claims_do_not_mean": {
    "obstacle_avoidance ranged": "ONE FORWARD-FACING CONE. It sees a wall ahead. It does NOT see a chair leg to the side, a cat, or a step down -- a robot that stops for walls and drives off a landing has satisfied `ranged` exactly as specified. That is why `ranged` and `mapped` are different words. Reversing and rotating are NOT gated, because a forward sensor has nothing to say about them and pretending otherwise would attach the word 'avoidance' to a direction the hardware cannot see.",
    "consent_tiers 2": "Video at ~1 fps with NO AUDIO TRACK. The only ASR on this machine is CLOUD (iFlyTek wss://ws-api.xfyun.cn/v2/iat, and Alibaba Tongyi), so there is no speech redaction that does not send the audio somewhere first. The claim is also stripped at pair time by BodyRuntime unless a delivery path is actually wired -- a driver says what the hardware can do, and whether it is wired is a fact about the installation.",
    "face redaction": "FRONTAL ONLY. A face turned away is not detected and is therefore not blurred. `basis: best_effort` is the accurate word, not a formality.",
    "goto is ABSENT": "This robot does not offer `goto`, and the absence is the claim. `navigate_to_pose` returns NAV_REJECTED unconditionally -- open-loop drive cannot honour a metric goal and nothing on this machine observes whether it arrived -- so the verb would be a promise the body cannot keep. ⚠️ THIS FILE ADVERTISED IT ANYWAY UNTIL 2026-09-01, months after the adapter stopped declaring it: the broker matched the verb, dispatched the mission, and the runtime refused it on arrival. An agent reading this profile chose a body for something it had already been decided could not be done. It can WANDER, which is a routine under `observe`; it cannot GO anywhere, which is a verb.",
    "sensing classify basic": "An 80-class COCO classifier at 0.87 fps. MEASURED, and the measurement is the reason it is not `full`: on the two images every YOLO distribution ships as ground truth it found three of four people and MISSED A BUS fifteen feet wide, and found nothing at all in the second at a 0.4 threshold. `basic` is the ceiling the broker may promise on this model's behalf.",
    "sensing count basic": "A LOWER BOUND, never a census. The count counts what the classifier saw, and the classifier misses things -- so `{\"type\": \"count\", \"label\": \"person\", \"value\": 3}` means AT LEAST three, probably more, and anything the model has no strength for is simply absent. Nothing downstream can tell a confident count from a blind one, which is why the level is the warning. A deployment that needs a trustworthy count needs a different model, not a lower threshold. ⚠️ IT WAS ABSENT UNTIL 2026-09-01 WHILE THE BODY WAS ALREADY EMITTING IT: count events went out on every mission against a manifest that left the modality unknown, so this robot was never matched for work it was already doing. The same drift as `goto`, pointing the other way.",
    "sensing ocr basic": "THIS ROBOT READS TEXT, and `basic` is a measurement rather than modesty. RapidOCR on ONNX, composed alongside the Darknet classifier on the one detector seam -- both, not either, which needed `adapters/composite.py` because the seam takes a single implementation and OCR used to REPLACE object detection rather than join it. Measured on the robot 2026-09-01 through the real adapter and the real JPEG decode: construction 0.4 s, first frame 0.96 s, then ~0.91 s per frame. Four synthetic frames read exactly, confidence 0.992-1.0. ⚠️ THAT IS AN UPPER BOUND AND IT IS WHY THE LEVEL IS `basic`: it was machine-rendered black-on-white text at a good size. A photographed sign at an angle in poor light is a different question and nothing has asked it yet. RapidOCR was chosen because it runs without torch on an arm64 machine that was 98% full, explicitly not because it is the most accurate engine available. It is a BUILD OPTION: `STRINGER_ROBOT_OCR=1` in deploy/start.sh turns on the detector AND the declaration together, from one flag, so a host that did not load a reader cannot advertise one. Without it this claim reads `none` -- the question asked, the answer no -- rather than being omitted. `audio` and `self_state` remain OMITTED because nothing has measured them, and unknown is the honest word for that.",
    "name_detection none, even with OCR": "Name detection needs TWO models and this robot loads one. A tagger answers with offsets INSIDE A STRING, so OCR is what makes a string exist for it to read -- but the tagger itself (`--text-detector-module`, spaCy) is still not wired here. Reading a sign and recognising that the sign carries somebody's name are different capabilities, and having the first does not confer the second. Text this robot reads is still screened for PII by pattern, including the pairwise recombination check that catches a plate arriving split across two OCR regions.",
    "routines blink and chirp": "THE TWO A TETHERED ROBOT CAN STILL OFFER, and the reason they are here. Both are `actuates: false` in the central registry, so the tether's routine strip keeps them while it removes `wander` -- which is what lets a plugged-in robot answer a request to express itself with light and sound instead of only panning a camera. ⚠️ THEY WERE A DECLARATION GAP, NOT A NEW FEATURE: `blink` and `chirp` have been in src/capability/routines.ts all along, this body has declared `expressive: {light: true, sound: true}` all along, and the adapter simply never offered them. On 2026-09-02 a tethered robot was asked to dance and the only routine a steering agent could reach was a camera sweep. SOUND IS BOUNDED HARDER THAN LIGHT and that is deliberate: 60 s for blink, 10 s for chirp. A light in a room is a light; a buzzer for the same duration is an imposition on everyone in earshot, including people who never agreed to the mission and cannot see the robot. Both switch OFF in a `finally`, so an abandoned generator -- a stop directive, a live lane breaking out of its loop -- cannot leave fourteen LEDs on or a buzzer sounding in somebody's house."
  },
  "evidence": [
    {
      "claim": "expressive.motion",
      "probe": "body/scripts/dance.py",
      "when": "2026-08-19",
      "result": "12 moves in 12.2 s and 29 in 30.3 s, on the robot. The motion list is drawn from RaspbotPerformer.CAN_EXPRESS at construction, so the declaration cannot drift from what the body does."
    },
    {
      "claim": "mobility.obstacle_avoidance = ranged",
      "probe": "the ultrasonic gate, adapters/obstacle.py, exercised on the robot",
      "when": "2026-08-19",
      "result": "Refused forward at 0.153 m while still allowing reverse. Before this file existed the correct declaration was `none`, which explicitly includes 'has a sensor nobody reads' -- and that is what this robot was, with a working ultrasonic nothing consulted."
    },
    {
      "claim": "sensing.classify = basic",
      "probe": "body/scripts/raspbot_probe.py (DarknetDetector, yolov3-tiny)",
      "when": "2026-08-18",
      "result": "0.87 fps at 1280x720."
    },
    {
      "claim": "consent_tiers includes 2",
      "probe": "body/scripts/media_mission_probe.py against mcp-server/frame-sink-server.mjs",
      "when": "2026-08-19",
      "result": "6 frames of 6 delivered at 1 fps in 6.0 s, gaps 939/951/1134/899/956 ms, 101-120 kB ciphertext each, every one hash=OK by the RECEIVER's own recomputation, and the tip opened with the agent's private key to 75,386 bytes of valid JPEG. That run proves the PATH; the blur was proven separately at 10/10 on a real face."
    },
    {
      "claim": "pixel redaction is affordable at the delivery rate",
      "probe": "timed on the robot at 1280x720",
      "when": "2026-08-19",
      "result": "decode 6.3 ms + Haar faces 37.6 ms + hyperlpr3 plates 55.8 ms + blur 7.2 ms + encode 5.7 ms = 112.6 ms, 11.3% of a 1 fps budget. A SLOW DELIVERY RATE BUYS PRIVACY: at 1 fps every frame is detected on its own with nothing cached, where the phone at 4 Hz caches boxes and blurs up to 250 ms of frames against stale positions."
    },
    {
      "claim": "the camera can be aimed from code",
      "probe": "body/scripts/servo_repeatability.py",
      "when": "2026-08-19",
      "result": "3.03 px (~0.3 deg) repeatability at a 5 s settle; 112 px at 2 s. An earlier 'the servos do not repeat' finding was WRONG and was retracted: the run moved two axes at once, the robot was physically picked up mid-test, and phaseCorrelate's confidence value was not being printed."
    },
    {
      "claim": "a fiducial fix is usable for indoor position",
      "probe": "body/scripts/fiducial_probe.py --live, against a tape-measured room",
      "when": "2026-08-19",
      "result": "3-5 cm, heading -0.0 deg against a known east. Needs STRINGER_ROBOT_FX/CX/CY -- see hazards."
    },
    {
      "claim": "the camera can be aimed at the room (CameraMount default tilt 95)",
      "probe": "body/scripts/aim_probe.py --axis {tilt,pan}",
      "when": "2026-08-19",
      "result": "Found by LOOKING at the frames, which is the only instrument that settles it. reservo.py sets tilt 25 at boot (ceiling) and the first default here was 50 (also ceiling); 95 shows wall, window and furniture. Both axes swept before concluding anything -- sweeping only channel 2 produced 'the mount is mechanically stuck', which was wrong. Contention ruled out by experiment, not argument: killing the only other holder of /dev/i2c-1 changed travel by 0.2 px and the pose held for 15 s after one command."
    },
    {
      "claim": "a sidecar on another machine can serve this device over the LAN",
      "probe": "body/scripts/sidecar_host.py on an NVIDIA Spark (aarch64, native docker build) + remote_inference from the robot",
      "when": "2026-08-20",
      "result": "Camera frames: 68 kB, 360-380 ms round trip over wifi. Positive control with rendered text: three regions read, TWO returned -- 'FIRE' and 'EXIT' came back and 'Nigel Cannings' did not. The name tagger ran on the sidecar and the name never crossed back to the device."
    },
    {
      "claim": "power guard reads the rail",
      "probe": "stringer_body.power.read_supply(), against vcgencmd",
      "when": "2026-08-26",
      "result": "Supply(state='ok', volts=5.12818, undervoltage=False) against 'EXT5V_V volt(24)=5.13220000V' and get_throttled=0x0 -- agreement to 4 mV. Checked because the guard has NEVER FIRED, and 'power has been fine' and 'it cannot read at all' look identical from the log; absent is not no. ⚠️ IT DOES NOT SHOW THE ROBOT IS ON MAINS: EXT5V_V is the INPUT rail and reads ~5.1 V from a healthy battery exactly as from a wall adapter. This machine can notice it is DYING and can never notice it is PLUGGED IN, which is why the tether is a switch a person sets -- a 5.13 V reading must not be argued into 'it is tethered'.\n⚠️ UPGRADED FROM ONE READING TO TWELVE, because one reading is not a pattern -- the broker session retracted a six-file diagnosis built on a single sweep the same evening, and the same error was sitting in this entry. 12 samples over 48 s: all 12 readable (so the sensor is reliable, not a lucky instant), 5.1175-5.1523 V, spread 35 mV, every state 'ok', undervoltage never set. 317 mV clear of the WARN floor and 467 mV clear of CRITICAL.\nThe single reading proved the PARSE (it agreed with vcgencmd to 4 mV). The twelve prove the SUPPLY IS STEADY, which is the different question somebody asks when they want to know whether the robot will still be up in the morning."
    },
    {
      "claim": "both directive paths reach the real broker's handler (the screen and the ACTIVE gate still need a mission)",
      "probe": "client.read_directives against production, with a wrong-route control",
      "when": "2026-08-26",
      "result": "wrong route  -> 404 '404 Not Found'      (plain text, the framework)\nreal route   -> 404 {'error':'not_found'} (the handler's own JSON)\nwith since=  -> 404 {'error':'not_found'} (parsed, not a 400)\nThe CONTROL is the point: both answers are 404, so the status code alone proves nothing about whether the path reached a handler. The BODY shapes differ, and that is what shows the route matched. Checked because read_directives and post_directive had never sent a real packet -- and client.py's own header records a 403 error-code-1010 that was invisible to every fake and surfaced on the first real one.\n⚠️ NARROWED 2026-08-26 BY AN AUDIT OF MY OWN CLAIMS. This says 'the directive transport' and only `read_directives` was exercised. `post_directive` HAS STILL NEVER SENT A REAL PACKET -- it is the path that carries every refusal, every clamp footnote and every 'switched to X' back to the agent, and it is the one that meets the broker's cleanText screen and its ACTIVE gate. A refusal that the broker rejects is a silence, which is the failure the whole channel exists to remove.\nIt needs an ACTIVE live_interactive mission, so it needs the owner. Until then: the read path reaches the handler; the write path is proven only against a fake I wrote.\n✅ CLOSED IN PART, SAME EVENING. post_directive HAS now sent real packets, with the same wrong-route control:\n  wrong POST route  -> 404 '404 Not Found'       (plain text, framework)\n  real POST route   -> 404 {'error':'not_found'}  (the handler's JSON)\n  via client.post_directive -> the same, so the method's URL, body and headers are right.\nSafe because a POST to a mission that cannot exist is refused at the ownership query before anything is written -- nothing created, nothing screened.\n⚠️ STILL UNPROVEN AND IT NEEDS A MISSION: the cleanText screen on the reply text, and the ACTIVE gate that makes a reply on a settled mission a 409. Those are the two that turn a refusal into a silence, and neither can be reached without an ACTIVE live_interactive mission. The transport is proven; the POLICY is not."
    },
    {
      "claim": "the kill-switch PREDICATE flips on real hardware (the daemon acting on it is untested)",
      "probe": "rename /etc/stringer/robot-enabled, read enabled(), restore",
      "when": "2026-08-26",
      "result": "switch REMOVED  -> enabled() False\nswitch RESTORED -> enabled() True\nfile back with its original Aug 19 02:25 timestamp, no leftover probe file, service still active, NRestarts 0, broker still 'listening'.\nChecked because this is THE ONLY E-STOP THIS ROBOT HAS -- nothing exposes a stop as a topic, a GPIO line or an I2C register -- and it had only ever been proven against a fake filesystem. The file has existed untouched since 2026-08-19, so 'it is there' was never evidence that removing it does anything.\nThe restore ran under a trap on EXIT INT TERM: a dropped connection mid-probe would otherwise leave the robot silently refusing every mission -- the fail-safe direction, and still wrong.\n⚠️ NARROWED 2026-08-26. What was measured is `os.access` -- the predicate the driver reads. What was NOT measured is the DAEMON acting on it: that `available()` withdraws presence and that `_abort_check` terminates a mission in flight. Those need a mission, and the framing above ('the only e-stop this robot has') invites reading this as an end-to-end proof that it is not.\nThe unit tests cover the consequence and cost 86 red tests to get right; the hardware check covers only the sensor. Both halves exist and neither has met the other on this machine."
    },
    {
      "claim": "the service is configured to come back after a reboot",
      "probe": "resolved unit properties + whether network-online.target is genuinely reached",
      "when": "2026-08-26",
      "result": "WantedBy resolves to multi-user.target and After resolves to network-online.target, so systemd will start it at boot and order it after the network.\n⚠️ AND THE ORDERING IS NOT INERT, WHICH WAS THE THING WORTH CHECKING. On many Pi installs network-online.target is NEVER REACHED because no wait-online service is enabled -- and an `After=` on a target nothing pulls in is a directive that does nothing, exactly like StartLimitIntervalSec in the wrong section. Here NetworkManager-wait-online.service is ENABLED, the target is active, and stringer-body.service appears in its reverse dependency list. So the daemon genuinely waits for the network rather than racing it and masking a failed first declare as a transient.\n⚠️ THIS IS STILL NOT PROOF THAT IT COMES BACK. Nothing here reboots the machine. It rules out the specific ways the unit could be wrong; only a real boot rules out the rest, and that is the owner's to authorise."
    },
    {
      "claim": "the service recovers from a crash without a human",
      "probe": "SIGKILL the daemon, wait out RestartSec, check the broker",
      "when": "2026-08-26",
      "result": "before  active, NRestarts 0, pid 3372\nkill -9 (SIGKILL, not SIGTERM -- a clean stop is not a crash, and Restart=on-failure is specifically the branch that had never run)\nafter   active, NRestarts 1, pid 6314, RECOVERED\nand it RE-DECLARED: measured_at moved 20:59:48Z -> 23:35:12Z, status listening, live_steer true, all three routines offered, stranger work not blocked.\n⚠️ THE SECOND HALF IS THE POINT. 'The process exists again' is not 'the daemon works again' -- a restarted process that fails its declare looks identical to systemd and is invisible to the broker. The moved measured_at is the only thing that separates them.\nRestart=on-failure was configured tonight and only ever observed NOT failing, which is the fifth-flavour shape: a mechanism whose branch has never executed. This is the unattended recovery the service was installed for, short of an actual reboot."
    },
    {
      "claim": "tonight's code runs on the robot's own interpreter, not just the dev machine's",
      "probe": "stdlib smoke test of the new paths, executed on the Pi",
      "when": "2026-08-26",
      "result": "19/19 on Python 3.11.2 / aarch64. The suite runs on 3.13 / x86-64, so every test of the directive channel, the clamp footnote and the refusal bounding had only ever executed on the WRONG INTERPRETER AND THE WRONG ARCHITECTURE. An import check proves syntax; this exercises behaviour.\nCovered: used->applied translation, asked==applied refused, bool and infinity refused, footnote surviving the sanitiser, empty footnote omitted, takes_directives both ways, steer reaching the chassis, stop reaching the mission, an unserveable control refused in words, the device not reading its own replies as commands, read derived from acted+refused, the high-water mark, hostile input described rather than quoted, and the terminate vocabulary including operator_stopped and the two accept refusals.\n⚠️ BOUNDARY: a smoke test of pure logic, NOT the suite. pytest is not installed and the tests are not deployed -- deliberately, because the card sits at 98% and `pip install` into this Python is what CLAUDE.md warns about. Nothing here touches hardware or the daemon. The probe was written into the package directory, run, and removed; absence verified."
    },
    {
      "claim": "the crash-loop limiter stops a loop, and the recovery is not `systemctl start`",
      "probe": "SIGKILL the daemon six times faster than RestartSec",
      "when": "2026-08-26",
      "result": "kills 1-5 -> restarted each time, NRestarts 2..6\nkill 6    -> FAILED. Still failed 20 s later, not thrashing.\nChecked because I had told the owner this limiter would bound a disk-full failure, and that was a REASONING claim about a mechanism I had only ever seen configured. Restart=on-failure was proven by one kill; the LIMIT never had been.\n⚠️ AND THE RECOVERY IS NOT WHAT ANYBODY WOULD TRY FIRST. A plain `sudo systemctl start` after a limit hit FAILS: 'Job for stringer-body.service failed because a fatal signal was delivered to the control process' -- a message about signals, pointing nowhere near the cause. `sudo systemctl reset-failed stringer-body` first, THEN start. That is a morning wasted if you meet it cold.\nRestored and verified working rather than merely active: status listening, measured_at moved to 01:43:02Z, live_steer true, stranger_blockers empty, NRestarts back to 0."
    },
    {
      "claim": "the camera path resolves and OPENS -- capture is not proven",
      "probe": "by-id path in start.sh, daemon restart, /proc environ",
      "when": "2026-08-26",
      "result": "The daemon opens /dev/v4l/by-id/usb-DHZJ-240229-XH_Integrated_Webcam_HD-video-index0 instead of bare index 0, because an index is assigned at ENUMERATION and this Pi carries video19-24 for its own ISP and codec blocks. Verified: service active after restart, the by-id path in the daemon's /proc environ, broker listening, measured_at moved.\n⚠️ AND THE CLAIM STOPS THERE. `V4L2Camera.__init__` raises if the device will not open, so a running service proves the path RESOLVED AND OPENED. It takes five warmup frames and DISCARDS their results, and isOpened() can be true for a device that delivers nothing -- so this does NOT prove the camera will serve a mission.\nThat step cannot be taken from outside: the daemon holds the device, so capturing means stopping it or waiting for a mission. THE FIRST OBSERVE IS THE PROOF. If it returns 'examined 0 frames', the camera path is the first place to look -- it changed on 2026-08-26 and nothing since has captured through it."
    },
    {
      "claim": "the camera SENSOR STREAMS for the daemon's whole lifetime, not per mission",
      "probe": "open fds and fuser on /dev/video0 with no mission running, against the code path",
      "when": "2026-08-26",
      "result": "With the daemon idle -- no mission, nothing in the log since its declaration -- /proc/<pid>/fd shows fd 3 -> /dev/video0 and `fuser -v` reports the python process holding it with the mmap flag, alongside pipewire and wireplumber. V4L2Camera.__init__ calls cv2.VideoCapture and then reads 5 warmup frames, which issues STREAMON; close() would release it, and the daemon builds the driver ONCE at startup and never closes it while running. So the sensor is powered and streaming from daemon start until daemon stop. NOT a claim that frames are captured while idle: run_once calls only report_presence and poll_missions, and never touches the camera without a mission. The steady ~19 read syscalls/s this process makes were NOT attributed and are not claimed to be frames. Recorded because an owner reading 'consent tier 0' would reasonably expect the camera to be off between missions, and it is not; and because continuous streaming is a power cost on a battery robot. Whether to open lazily is an owner decision, not a defect: the 5 warmup frames exist because the first frames are unusable, so a lazy open moves that latency into every mission."
    },
    {
      "claim": "sensing.ocr = basic",
      "probe": "body/scripts/ocr_probe.py, on the robot",
      "when": "2026-09-01",
      "result": "RapidOCR on ONNX, in the venv, through the real adapter and the real JPEG decode. Construction 0.4 s; first frame 0.96 s; then ~0.91 s per frame (about 1.1 fps). Four frames read EXACTLY: 'Costa Coffee'; 'Nigel Cannings' + 'Costa Coffee'; 'LM67' + 'HGG'; 'Parking 2hr max'. Confidence 0.992-1.0.\nONE DETECTION PER LINE, which is the half that matters downstream: joined, the name tagger reads 'Nigel Cannings Costa Coffee' as ONE PERSON -- an entity nobody wrote, which would then be masked as somebody's name -- and a plate split across regions could not be recombined and caught. The engine returns regions, so both rules still work.\n⚠️ AN UPPER BOUND, AND WHY THE LEVEL IS basic AND NOT full: this was machine-rendered black-on-white text at a good size, the easiest input OCR ever gets. A photographed sign at an angle in poor light is a different question and nothing has asked it yet. The camera could not be used for this because the daemon holds it.\nDeployed and live: start.sh exports STRINGER_ROBOT_OCR=1, the daemon restarted clean (NRestarts 0) and declared sensing {classify: basic, count: basic, ocr: basic}. The SAME flag composes the detector and sets the level, so a host that did not load a reader cannot advertise one."
    }
  ],
  "prerequisites": {
    "_comment": "Per-CAPABILITY, never a blanket list: a robot that only dances needs nothing installed, and tying each package to what it unlocks is what lets an owner decline a capability and skip its dependency. THE PROFILE NAMES A PREREQUISITE; THE OWNER INSTALLS IT. Never a script the profile runs.",
    "venv": {
      "path": "/home/pi/stringer/.venv",
      "created_with": "--system-site-packages, so the vendor's cv2, numpy and mediapipe stay visible without being reinstalled or paid for twice in disk",
      "why_not_system_python": "Debian 12 refuses a bare pip install (PEP 668), and --break-system-packages is exactly the wrong answer. A named venv is reversible with one rm -rf, which matters because the next session inherits this machine and cannot see what a previous one did to it.",
      "disk_free_before": "1.1 GB",
      "disk_free_after": "1.0 GB"
    },
    "by_capability": [
      {
        "capability": "expressive / dance",
        "package": "none -- stdlib and smbus",
        "source": "present on the vendor image",
        "disk_mb": 0,
        "without_it": "nothing; this needs no installs at all"
      },
      {
        "capability": "observe, sensing.classify",
        "package": "cv2, numpy, yolov3-tiny weights",
        "source": "vendor image / ~/yolov3",
        "disk_mb": 0,
        "without_it": "no object observations"
      },
      {
        "capability": "face redaction (face_blur)",
        "package": "YuNet: face_detection_yunet_2023mar.onnx (232 KB, OpenCV Zoo), plus OpenCV >= 4.5.4 for cv2.FaceDetectorYN. The Haar cascade at ~/project_demo/raspbot/haarcascade_frontalface_default.xml remains as a NAMED FALLBACK only.",
        "DO_NOT_SHIP_THE_CASCADE_ALONE": "On 2026-09-03 the owner's own face was published UNBLURRED over a live stream declaring face_blur. Seated, three-quarter view, glasses, camera tilted up 49 degrees -- an ordinary way to be in a room. Measured against that exact frame: haarcascade_frontalface_default at its paranoid redaction tuning found 0 faces in 531 ms; so did sf 1.10/1.05/1.30, so did the equalised image, so did twelve rotations, so did frontalface_alt2 and profileface. YuNet found it at confidence 0.863 in 86 ms. The cascade did not miss by a little -- it could not see the face at all, at any setting. It is also SIX TIMES more expensive to fail.",
        "and_the_body_must_say_which_one_it_loaded": "A body whose model file is absent falls back to the cascade and redacts materially worse than its declaration implies. Expose a redaction_basis() returning 'yunet' or 'haar_frontal'. Until that existed the two were indistinguishable from outside, which is why this shipped for weeks.",
        "source": "OpenCV Zoo, committed to the repo and carried in the deploy payload -- a model is not code, so nothing else notices it missing",
        "disk_mb": 0,
        "without_it": "make_media_delivery raises at construction; no tier-2 mission can be served"
      },
      {
        "capability": "plate redaction (plate_blur)",
        "package": "hyperlpr3",
        "source": "vendor image",
        "disk_mb": 0,
        "without_it": "no plate blur; faces are unaffected"
      },
      {
        "capability": "tier-2 sealed delivery",
        "package": "pynacl 1.6.2",
        "source": "INSTALLED into the venv, 2026-08-19",
        "disk_mb": 29,
        "without_it": "NaClFrameSealer raises at construction, deliberately, so a deployment without it fails at wiring time rather than at the first frame of a live mission"
      },
      {
        "capability": "text redaction (text_blur)",
        "package": "an EAST text-detection model",
        "source": "NOT INSTALLED",
        "disk_mb": 96,
        "without_it": "text_blur cannot be claimed at all -- there is no engine. Also blocks name_detection: a tagger answers with offsets inside a STRING, so with no OCR there is nothing to hand it"
      },
      {
        "capability": "speech redaction",
        "package": "a LOCAL ASR (vosk-small or whisper-tiny)",
        "source": "NOT INSTALLED -- the only ASR here is cloud",
        "disk_mb": 75,
        "without_it": "no audio track may be offered; tier 2 is video-only"
      },
      {
        "capability": "name_detection (personal names in read text)",
        "package": "an OCR engine AND a name tagger -- BOTH, they are one capability",
        "source": "Dockerfile.sidecar, BUILT AND VERIFIED on linux/arm64. 745 MB on disk (docker image inspect: 710 MB; 273 MB compressed to transfer). CORRECTED 2026-08-20 from a published 1.06 GB, which was WSL's `docker images` DISK USAGE figure including buildx attestation blobs -- not the image. Against 998 MB free this now FITS, leaving the card at ~99.5%, which is a different reason not to do it.",
        "disk_mb": 745,
        "without_it": "no [name] mask is ever emitted and no level is declared, so every scene/text_read this device sends is reported as a redaction gap broker-side (redactionGap warns unless the level is `full`)"
      }
    ],
    "watch_the_disk": "This card had 1.1 GB free before pynacl. Two more models would halve what is left. A prerequisite list without sizes is not usable on hardware like this."
  },
  "build": {
    "_comment": "What had to be WRITTEN to make this robot a Stringer embodiment. Named, not carried: the code is in this repository, under review.",
    "driver_factory": "stringer_body.adapters.raspbot:make_driver",
    "media_delivery_factory": "stringer_body.adapters.raspbot:make_media_delivery",
    "files": [
      "body/stringer_body/adapters/raspbot.py -- ROS2Interface over V4L2 + I2C, the performer, the redacting frame source",
      "body/stringer_body/adapters/obstacle.py -- the ultrasonic gate and the stuck-sensor liveness check",
      "body/stringer_body/adapters/fiducial.py -- AprilTag pose and the room locator",
      "body/stringer_body/adapters/dance.py -- the abstract expressive vocabulary and its substitution chain",
      "body/stringer_body/live_delivery.py -- the frame loop and its failure policy",
      "body/stringer_body/redact_pixels.py -- fail-closed pixel redaction"
    ],
    "seam_note": "The seven-method ROS2Interface gets you a TIER-0 STILLS BODY. This previously said the generic driver drove it unmodified, and embodiments.md section 2 now retracts that: serving routines or the live lane needs four more methods, they are optional by construction (getattr with a fallback), and the runtime holds a ROS2BridgeDriver that must FORWARD each one. A missing forwarder is not an error -- it is the absent-method default, and on 2026-09-03 four features were dead code on this robot because of it while every unit test passed.",
    "_hoist_note": "Written for this robot vs inherited unchanged. The split is the framework's central claim and `generic_modules_do_not_import_the_robot` is what keeps it true: raspbot.py was 1046 lines and about 320 of them were generic.",
    "robot_specific_files": [
      "body/stringer_body/adapters/raspbot.py -- the I2C register map, the mecanum kinematics, the performer, the wiring, the camera mount and its vendor limits"
    ],
    "inherited_unchanged": [
      "adapters/v4l2_camera.py, adapters/darknet.py, adapters/media_source.py -- hoisted out of raspbot.py, specific to nothing",
      "adapters/fiducial.py, dance.py, approach.py, obstacle.py -- stdlib and types only, zero robot imports",
      "runtime.py, client.py, events.py, safety.py, live_frames.py, live_delivery.py, redact_pixels.py, sidecar.py"
    ]
  },
  "aiming": {
    "pan_deg_is_now_in_the_PLATFORM_frame": "The adapter declares [-59, 59] and maps to raw servo counts internally, the way tilt_deg already did. It used to declare [0, 118] -- this mount's RAW SERVO TRAVEL -- against a registry that defines pan_deg as '0 = straight ahead, positive = clockwise seen from above'. Three cold agents hit it independently on 2026-09-04; one wrote: 'If 0 is straight ahead and the range only goes clockwise, look left and right -- which live.md says is what sweep means -- cannot be done.' A body publishing its own servo scale as if it were the platform's frame is a declaration fault, and it made the camera's left half unaskable.",
    "AHEAD_IS_A_CONVENTION_NOT_A_MEASUREMENT": "PAN_AHEAD_SERVO = 59 is the mechanical midpoint of the [0, 118] travel. NOBODY HAS MEASURED which servo count aligns this camera with the chassis axis, and DEFAULT_PAN's comment in raspbot.py is explicit that a fiducial tag cannot tell you: every sweep measures where the TAG is, not where the chassis points, and the two differ the moment the robot is turned. So an agent's 'ahead' is offset from the chassis by an unmeasured constant.",
    "why_declare_it_anyway": "An agent with an offset 'ahead' can still scan a room; an agent with no left cannot do anything. The offset is a bounded, correctable error; the missing axis was a capability that did not exist. It also matches the sentence the broker published as the interim -- 'treat the midpoint as roughly ahead' -- so the declaration and the document agree instead of contradicting each other.",
    "HOW_TO_MEASURE_IT_PROPERLY": "Needs a person with the robot. The chassis-fixed reference is the ultrasonic cone: place a narrow object directly in front of the chassis at a known range, confirm the ultrasonic sees it, then pan the camera until the object is centred in frame. That pan is ahead. Replace PAN_AHEAD_SERVO with it -- do NOT refine it by reasoning. The robot is tethered and cannot turn to find its own axis.",
    "stops_per_sweep": 3,
    "_stops_per_sweep": "Three: -pan_deg, 0, +pan_deg. The arc straddles ahead now; it used to run one way from one end of the servo travel, which is the other half of why 'look left and right' could not be expressed. /steering cannot say this, so it is declared here."
  },
  "live": {
    "_why": "Required because this body claims live_video. Every field here exists because its absence caused a shipped failure on this exact robot on 2026-09-03 -- see embodiments.md §4c.",
    "publish_resolution": "1280x720",
    "detect_measured_at": "1280x720",
    "_they_must_match": "They did not. Detection was measured at 640x480 and the lane publishes 1280x720 -- four times the pixels. This body declared 5 fps, delivered 2.1, and published a BLACK FRAME in every slot it missed. The owner's live watch was 58% black while packetsSent, bytesSent and framesDecoded all climbed, because black frames decode perfectly.",
    "frame_budget_ms": {
      "capture": 13,
      "detect": 219,
      "encode": 66
    },
    "topology": "colocated",
    "max_fps": 3.3,
    "missed_slot_policy": "hold_last_frame",
    "_missed_slot_policy": "Re-sends the last ALREADY-REDACTED frame for 2.0s (under the stall threshold) with the drop counter still climbing, so a slow body looks like a lower frame rate and a DEAD one still goes black on the same schedule.",
    "detector": {
      "name": "yunet",
      "model": "face_detection_yunet_2023mar.onnx (OpenCV Zoo, 232 KB), via cv2.FaceDetectorYN, OpenCV >= 4.5.4",
      "recall_measured_on": "A photograph of a real person in the actual deployment -- seated, three-quarter view, glasses, camera tilted up 49 degrees -- NOT frames chosen by a detector. YuNet found the face at confidence 0.863 in 86 ms. haarcascade_frontalface_default found 0 at every tuning, equalised, de-rotated through twelve angles, and with frontalface_alt2 and profileface, in 531 ms. The earlier recall figure on this body was built from 'frames a full-scale detector agreed contained a face', which by construction can never find a face class the detector never sees -- and that was exactly the class that shipped.",
      "recall": {
        "_what_this_is": "THE PREMISE OF THE SENTENCE ABOVE, AS DATA, so a test can check it and so it EXPIRES when the premise does. `recall_measured_on` is prose and prose cannot be compared against the lane it describes.",
        "measured_at": "1280x720",
        "n": 1,
        "found": 1,
        "rate_declared": null,
        "_why_no_rate": "YOU CANNOT STATE A RECALL RATE FROM ONE PHOTOGRAPH, and this body's entire real-deployment evidence is one photograph. `1/1` is not 100% recall; it is a single success. The figures under detect_scale_probe that read 'recall 1.000 (12/12)' were measured on frames A FULL-SCALE DETECTOR HAD ALREADY AGREED CONTAINED A FACE, which by construction can never find the face class the detector never sees -- and that was exactly the class that shipped. They bound the SCALE question and say nothing about recall in a room.",
        "model_sha_or_name": "face_detection_yunet_2023mar.onnx",
        "expires_if": "publish_resolution or detect_measured_at changes, or the model file changes. A measurement carries its premise and expires when the premise does -- this body declared a 121 ms detection measured at 640x480 for a lane publishing 1280x720, internally consistent and wrong by four times.",
        "what_would_make_this_a_rate": "A labelled set of frames from THIS camera in THIS deployment that were not selected by any detector -- tens, not one -- with the faces marked by a person. Until that exists `rate_declared` stays null and no page may quote a recall percentage for this body."
      }
    }
  },
  "hazards": {
    "it_cannot_see_a_drop": "THE ULTRASONIC IS A FORWARD CONE AND THERE IS NO DOWNWARD SENSOR. A table edge, a stair, a kerb or a landing is invisible to every layer of this stack: the wander's range gate, the performer's own gate, and the continuous reflex all read the same forward-facing transducer. A robot that stops for walls and drives off a step has satisfied `obstacle_avoidance: ranged` EXACTLY as specified -- the declaration is honest and the floor is still a metre below. DO NOT TEST MOTION ON A DESK. Seen 2026-08-21 with the robot on the owner's desk, wheels free and the kill switch ON; nothing was running, and any wander or dance probe would have driven it off."
  },
  "known_gaps": {
    "the_light_strip_faces_UP_and_the_camera_CANNOT_SEE_IT": "The fourteen WS2812s point UPWARD. At raw tilt 180 the camera sees the underside of whatever the robot is parked beneath, and THAT is what glows -- a forward-facing frame, which is every frame a watcher has ever seen, cannot contain them, and there is no cast on the floor. Measured 2026-09-03 by lock-in against a known on/off drive: reg 0x03 gives corr 0.67 luma / 0.65 saturation with four consecutive cycles at +15-20 luma, and the paired stills show the surface above the robot going grey-white to blue-violet. THE LIGHTS WORK AND THE VIDEO CANNOT SHOW THEM. The owner reported 'I don't think blink works at all' from the live picture and was right about what they could see and wrong about the hardware -- an expression this body CAN make is one a remote watcher cannot perceive, which is a capability question rather than a bug. An agent choosing between blink and chirp on a REMOTE watch should know that only one of them travels.",
    "the_strip_ignores_its_first_two_toggles": "0x03 does not answer for roughly two toggles after an off sweep: cycles 0-1 read ~0 in two independent runs and the light appears from cycle 2-3. At BLINK_PERIOD_S = 0.25 s that is a half-second dead window at the START of every blink. The old BLINK_SECONDS floor of 0.5 s was TWO TOGGLES -- entirely inside it -- so a blink asked for without a duration lit nothing at all, which is a second and independent reason the owner saw no light. The floor is 2.0 s now (eight toggles, six of them visible). Anyone tuning the period should measure the dead window again rather than assume it scales.",
    "sensing_audio_is_not_declared": "The microphone captures and the sidecar transcribes, but `sensing.audio` stays undeclared. Nothing has yet transcribed intelligible SPEECH from this microphone -- only ambient noise, which the confidence floor refused. A claim is about what an installation has been shown to do, and 'the hardware works' is not that. It also runs on a SIDECAR, so any declaration must carry capture_link: local_network: raw unredacted audio of the room crosses the wire before anything redacts it.",
    "the_vendors_cloud_speech_stack": "This machine ships iFlyTek IAT/TTS pointing at wss://ws-api.xfyun.cn and wss://tts-api.xfyun.cn with credentials on disk. It cannot be part of an on-device privacy boundary: to redact speech you must first know what was said, and learning that by streaming raw audio to a third party discloses the thing being protected in order to decide whether to disclose it. A Stringer-enrolled unit with the vendor AI stack enabled has a hole through the tier-0 boundary that has nothing to do with Stringer's code."
  },
  "environment": {
    "_comment": "Observed on the machine, not assumed. A one-shot at boot and a continuous fight look identical from ONE observation, which is why 'something is competing for the servos' survived as a theory for so long.",
    "autostart": [
      {
        "entry": "~/.config/autostart/start.desktop",
        "runs": "oled_*/start.sh -> the OLED daemon, AND reservo.py",
        "effect": "reservo.py sets pan 90 / tilt 25 ONCE at boot. Tilt 25 is the CEILING. This is the whole of 'the robot points skywards'."
      },
      {
        "entry": "~/.config/autostart/start_raspbot.desktop",
        "runs": "project_demo/raspbot/raspbot_start.sh -> raspbot.pyc + yb-discover.py",
        "effect": "the vendor control app. Holds /dev/video0 and RE-HOMES the pan/tilt whenever it starts, so restarting it is an UNDO. Not running as of 2026-08-19, but returns on the next login."
      }
    ],
    "device_holders": [
      {
        "device": "/dev/video0",
        "held_by": "pipewire, wireplumber",
        "blocking": false,
        "note": "V4L2 permits multiple openers; this is not the problem it looks like"
      },
      {
        "device": "/dev/i2c-1",
        "held_by": "the OLED daemon",
        "blocking": false,
        "note": "TESTED, not assumed: killing it changed measured servo travel by 0.2 px, and the pose is stable for 15 s after one command"
      }
    ],
    "vendor_limits": [
      {
        "what": "servo channel 2 (tilt) travel is 0-100",
        "stated_in": "Raspbot_Lib.Ctrl_Servo",
        "enforced_by_vendor": false,
        "why_not": "`data = [id, angle]` is built BEFORE the clamps, so every clamp rebinds a local and none reaches the wire",
        "reimplemented_at": "stringer_body/adapters/raspbot.py SERVO_LIMITS"
      }
    ],
    "audio_capture": {
      "device": "plughw:2,0",
      "native_format": "CORRECTION 2026-08-21, and the first claim here was wrong. The card's CAPTURE endpoint is MONO S16_LE and offers 44100 or 48000 Hz ONLY -- there is no 16000. The stereo channel map (FL FR) belongs to the PLAYBACK side, i.e. the speaker, and reading it as a stereo microphone is the easy mistake. `arecord -D plughw:2,0 -r 16000` therefore RESAMPLES silently in ALSA's plug layer; it does not capture at 16 kHz. Stated because a resample nobody declared is exactly what this project refuses elsewhere: the transcriber raises on a rate mismatch rather than converting quietly, and then the capture layer did the conversion for it out of sight.",
      "measured_2026_08_21": "5 s of ambient room: rms 667, peak 3370 of 32767. A level check runs BEFORE every transcription request, because a valid WAV of zeroes transcribes to \"\" exactly like a quiet room -- the audio twin of fifteen well-exposed frames of a skylight.",
      "proven": "robot microphone -> LAN -> sidecar (Vosk small, 68 MB) -> tier-0 events. Ambient room returned confidence 0.524, BELOW the 0.6 redaction floor, and the floor refused it rather than emitting plausible garbage.",
      "gain": "SHIPS AT 35/35 (+23 dB) AND CLIPS. Real speech pinned at 32767 and came back as word salad -- which reads as a poor model rather than a destroyed signal. The card's 'Auto Gain Control' switch is PLAYBACK-side and governs capture not at all, so nothing manages this by itself. `amixer -c 2 sset Mic Capture 18` gave headroom (peak 21362) and lifted confidence from 0.72 to 0.76 on the same voice in the same room. Check the peak before blaming the transcriber.",
      "proven_speech_2026_08_21": "20 s of real speech, robot microphone -> LAN -> sidecar: confidence 0.758, and a spoken personal name was MASKED to [name] in the delivered event. That is the whole redaction chain -- local transcription plus name tagging -- working on hardware rather than in a test.",
      "recommended_capture": "Record at 48000 native (`-D hw:2,0 -r 48000`) and downsample DELIBERATELY where somebody can see it, or accept ALSA's plug resampler and say so. Which of the two is better has not been measured on this hardware and should not be guessed at.",
      "measured_accuracy_2026_08_21": {
        "method": "One 30 s passage read aloud to the robot, scored by scripts/asr_accuracy.py. IDENTICAL AUDIO through every engine -- re-reading it per model measures the room and the reader, not the model.",
        "results": [
          "vosk-small (50 MB)   WER 0.619  conf 0.778  3.3 s  no language ID",
          "whisper-tiny         WER 0.429  conf 0.666  0.8 s  en 0.47",
          "whisper-base         WER 0.310  conf 0.786  1.1 s  en 0.91",
          "whisper-small        WER 0.214  conf 0.895  2.8 s  en 0.60"
        ],
        "the_finding_that_matters": "BETTER ASR IS A PRIVACY CONTROL, NOT A QUALITY PREFERENCE. Vosk heard 'Sarah Whitfield' as 'sarah with field'. The tagger masked the forename and the SURNAME went through as two ordinary English words -- a redaction failure the pipeline reported as success, because the tagger cannot mask what the transcriber never handed it as a name. Whisper-small heard the surname correctly and it was masked. A name the transcriber mangles is a name the redactor cannot see.",
        "speed_note": "On this sidecar whisper-small is BOTH more accurate AND faster than vosk-small on a 30 s clip. The trade is memory and model download, not latency. Load times above are first-run DOWNLOADS, not steady state.",
        "capture_gain_matters_more_than_the_model_at_the_extremes": "Same passage, same model: capture 35/35 clipped and gave WER 0.786 with the name unheard; capture 27 gave 0.571 with the name masked. Fix the level before comparing engines.",
        "live_path": {
          "what": "The same passage read to the robot and scored through the REAL sidecar HTTP path, not a bench harness -- robot mic -> LAN -> /infer -> transcriber -> name tagger -> tier-0 event.",
          "vosk_small": "WER 0.571  conf 0.782  8 deletions  name masked but the SURNAME escaped as 'with field'",
          "whisper_small": "WER 0.286 (0.167 ignoring number FORMAT)  conf 0.877  ZERO deletions  name masked cleanly, and 'Guildford' over-redacted as a name, which is the safe direction",
          "two_scores_on_purpose": "Whisper writes '7.40', 'platform 4' and '15' where the reader said 'seven forty', 'platform four' and 'fifteen'. Those are not mishearings, and counting them as such reports the engine as 29% wrong when most of the gap is orthography -- but they ARE differences a human reads, so normalising them silently would flatter whichever engine emits digits. Both numbers are printed and each is labelled."
        },
        "the_bottleneck_is_the_MICROPHONE_not_the_model": {
          "evidence": [
            "SAME passage, SAME models, two sources:",
            "                     room via robot mic   clean synthesised 16 kHz",
            "  vosk small-0.15          WER 0.619              WER 0.161",
            "  vosk 0.22-lgraph         WER 0.833              WER 0.161",
            "  vosk 0.22 (1.8G)         WER 0.857              WER 0.032"
          ],
          "conclusion": "A 26x swing on the SAME model from the SAME words. The models are not the problem; the capture chain is. Buying a bigger model achieves nothing until the audio reaching it is good enough to use -- and on this robot it is not.",
          "and_the_sensitivity_runs_the_WRONG_WAY": "On clean audio the 1.8 GB model is far the best (0.032). On OUR audio it is the WORST (0.857), losing to the 40 MB model. Large models assume clean input; small ones are trained to be robust to the conditions an edge device actually meets. So 'put a bigger model on the sidecar' is only right AFTER the capture is fixed, and is actively wrong before it.",
          "suspects_in_order": [
            "The undeclared ALSA plug resample 48000 -> 16000. The card offers no 16 kHz capture, so `plughw` converts silently; the big models may be less tolerant of its artefacts.",
            "Microphone distance and room reverb -- a cheap C-Media capsule across a room is a different signal from a close mic.",
            "Capture gain, already shown to swing WER from 0.786 (clipped) to 0.571 on the same engine."
          ],
          "next_experiment": "Capture at 48000 NATIVE and downsample deliberately with a good resampler, then re-run all three. That isolates the resampler from the room. Not yet done, and not guessed at."
        }
      }
    },
    "speech_board": "A SECOND, SEPARATE AUDIO PATH and it never produces audio. The vendor's voice control (10.Basic_voice_control/) uses `Speech_Lib.Speech.speech_read()`, which returns a COMMAND ID over serial from the offline board on /dev/ttyUSB0. There is no PCM, no AGC and no capture code to borrow -- the recognition happens in hardware against a fixed vocabulary. Worth knowing for two reasons: there are no vendor capture tricks to copy, and a keyword channel that never yields audio has NO privacy exposure at all, which makes it a candidate control path rather than a sensing one.",
    "live_redaction_cost": {
      "measured": "2026-08-21, Pi 5, 640x480, Haar frontal cascade, real frames",
      "breakdown_ms": "capture 13.0 | detect 121.4 | destroy ~0 | p95 whole 137.6",
      "declarable_fps": 3.3,
      "the_number_that_matters": "DETECTION IS 90% OF THE COST. Capture is 13 ms and destroying a region is immeasurable; the cascade walking 640x480 is 121 ms. Any optimisation that is not about the search is noise.",
      "collides_with_an_assumption_already_in_the_code": "The aiortc publisher is written for 10 fps at 640x480. Redaction sustains 7.3. Running the publisher at its assumed rate WITH redaction drops roughly a quarter of frames -- or, if nothing enforces it, publishes unredacted. The rate must be derived from this measurement, not chosen.",
      "downscaled_detection_is_UNRESOLVED": {
        "RESOLVED_2026_08_21": "NO. Do not detect on a downscaled copy on this body.",
        "controlled_method": "Captured 12 frames a full-scale detector agreed contained a face, SAVED the pixels, then ran every scale against the same frames. The earlier uncontrolled run could not tell 'downscaling broke detection' from 'nobody was in shot'.",
        "result": [
          "scale 1.00  recall 1.000 (12/12)  78.6 ms",
          "scale 0.75  recall 0.000 ( 0/12)  45.5 ms",
          "scale 0.50  recall 0.000 ( 0/12)  20.2 ms",
          "scale 0.35  recall 0.000 ( 0/12)  10.0 ms"
        ],
        "the_mechanism_and_it_is_not_minSize": "minSize is 10 px, so that is not the floor. The face was 63x63 px in a 640x480 frame at desk distance. At 0.9 it becomes ~56 px and detection FLIPS between frames; by 0.75 (~47 px) it is gone. Haar's reliable floor sits far above its nominal minSize because downscaling strips the fine detail the cascade keys on.",
        "why_this_generalises_the_WRONG_WAY_for_robots": "A face that fills a videoconference frame has resolution to spare, so detect-small is a free win there. A ROBOT sees faces small and far away -- 63 px at desk distance -- which is exactly the regime where the optimisation removes the last detail there was. The technique is not wrong in general; it is wrong HERE, and here is where it would have shipped.",
        "what_to_try_instead": "A DNN face detector (OpenCV DNN / YuNet) rather than Haar: better on small faces AND faster, which attacks the 90% of the budget that detection occupies without trading recall for it. Untested on this board."
      },
      "declarable_fps_stands_at": 3.3,
      "_why_this_moved": "It read 7.3 beside live.max_fps 3.3 and two cold agents quoted BOTH. 7.3 was the REMOTE-sidecar figure from a 640x480 measurement; the lane publishes 1280x720 colocated and sustains 3.3. One profile carrying two rates for the same body is a profile an agent cannot use.",
      "_declarable_fps": "3.3, and it read 7.3 until 2026-09-05. A cold agent rebuilding this body from the published material found BOTH still in the file -- I corrected `declarable_fps_stands_at` the day before and left this one, which is the machine-readable field of the two. Correcting one copy and leaving the others is the defect this profile records happening three times elsewhere. 7.3 was the REMOTE-sidecar figure from a 640x480 measurement; this lane is colocated and publishes 1280x720."
    },
    "reflex": {
      "measured_2026_08_21": "i2c ultrasonic read 0.84 ms median. Loop holds 17.9-19.3 Hz against a 20 Hz target, so the declared reaction distance is real rather than a constant somebody hoped for.",
      "verified_on_hardware": "179 ticks, 179 cuts, 716 motor register writes (4 motors x 179), all of them stops. Before the run that number was ZERO and the assertion that every write was a stop was vacuously true.",
      "what_the_hardware_found_that_fakes_could_not": "Sitting in the hysteresis band -- past the stop distance, not yet clear -- took NO branch at all: the motors were cut only on the TRANSITION into a trip, so a reflex that starts tripped and stays there never touched them, against its own documented promise. And after 232 ticks the reason still read 'the reflex has not looked yet', which reads as a DEAD LOOP to whoever is debugging it -- the one diagnosis this component exists to make impossible. Every fake tripped it via a close obstacle or a failed detector and never simply left it where it started."
    },
    "live_lane_budget": {
      "measured_2026_08_21": "Pi 5, real camera frames, software encode",
      "no_hardware_encoder": "The Pi 5 exposes rpivid, which takes HEVC as OUTPUT -- it is a DECODER. There is no hardware video encoder on this board, so the CPU numbers below are the whole budget and there is nothing to offload to.",
      "STATE_THE_RESOLUTION_OR_THE_RATE_IS_FICTION": "The numbers below were measured at 640x480. THE LIVE LANE ON THIS BODY PUBLISHES 1280x720 -- four times the pixels. That single unstated mismatch is why this body declared 5 fps, delivered 2.1, and published a BLACK FRAME in every slot it missed: the owner's live watch was 58% black while packetsSent, bytesSent and framesDecoded all climbed, because black frames decode perfectly. The declaration and the rate derived from it were internally consistent and consistent about the wrong frame size. Record the resolution beside every frame-cost number, and check it equals the one the lane actually publishes.",
      "components_ms_at_640x480": [
        "capture (MJPG decode)                13",
        "face detection @640x480             121   <- 60% of the budget",
        "VP8 encode @640x480, REAL frames     66",
        "VP8 encode @320x240, REAL frames     17"
      ],
      "components_ms_at_1280x720_THE_RATE_ACTUALLY_DECLARED": [
        "capture (MJPG decode)                13",
        "face detection @1280x720, YuNet     219   <- measured 2026-09-03 on a real captured frame",
        "VP8 encode                           66",
        "total 298 ms  ->  3.3 fps declared"
      ],
      "downscaling_before_detection_was_tried_and_REJECTED_on_measurement": "yunet @1280x720 218.7 ms 1 face; @640x360 48.8 ms 0 faces; @448x252 17.1 ms 0 faces; @320x180 7.3 ms 0 faces. Thirty times faster and blind. A redactor that cannot see a face is not a cheaper redactor.",
      "measure_on_REAL_CONTENT_not_noise": "A Haar cascade early-rejects on random noise and crawls on structure, so a synthetic benchmark flattered it by 40%% on this board. Noise is not a room -- and the same trap in the other direction cost a factor of two on the encode number below.",
      "random_noise_nearly_doubled_the_encode_number": "First measurement used random noise: 123 ms against 66 ms for real frames. Noise is incompressible and is the worst case for any codec, so it is the wrong benchmark for a camera. Re-measuring changed the answer by a factor of two and would have condemned the lane on a number that was never real.",
      "achievable_rates": [
        "1280x720 colocated, YuNet: 13 + 219 + 66 = 298 ms  ->  3.3 fps  <- WHAT THIS BODY DECLARES",
        "SUPERSEDED, kept for the reasoning only -- 640x480 with the Haar cascade:",
        "  640x480 end to end:  13 + 121 + 66 = ~200 ms  ->  ~5 fps",
        "  detect at 640x480, PUBLISH at 320x240: 13 + 121 + 17 = ~151 ms -> ~6.6 fps"
      ],
      "the_composition_that_helps": "Detection MUST stay at full resolution -- proven separately, recall goes 1.000 -> 0.000 at 0.75 scale because a robot's faces are small. But the frame can be downscaled AFTER redaction for transmission: the face is already destroyed by then, so shrinking what is sent costs nothing in privacy and buys a third of the budget.",
      "caveat_this_is_indicative_not_definitive": "Measured through cv2's VideoWriter, which is NOT the code path aiortc uses (libvpx via PyAV). It bounds the problem; it does not settle it. aiortc is not installed -- the card is at 98% with 1.2 GB free, and PyAV plus FFmpeg is the kind of install this device cannot casually absorb.",
      "consequence_for_the_declaration": "max_fps for the live lane on this body is 3.3, derived from the 1280x720 budget above. Declaring a rate the body cannot reach is not optimism: every slot it misses is published as a black frame, and every counter on both sides of the wire reads healthy while it happens.",
      "a_missed_slot_must_RE_SEND_the_last_picture_not_publish_black": "FrameFeed holds the last already-redacted frame for a bounded window (2.0s, under the stall threshold) with the drop counter still climbing, so a slow body looks like a lower frame rate -- which is what the ruling always said it should -- and a DEAD one still goes black and is declared on the same schedule.",
      "THE_LANE_MUST_HAVE_ITS_OWN_PRODUCER": "Feeding the live lane from the mission loop makes the picture hostage to anything that loop does slowly: a servo settle, a batch inference, a routine that runs dry. Three separate blackouts, fixed one at a time, before the coupling itself was the answer. A thread reads the camera and pushes continuously for as long as the lane is open; the mission decides where the camera POINTS, not whether frames flow. ONE producer -- two racing on the sequence number produce drops, not frames. And it must never touch the routine, or a body would move because somebody was watching it.",
      "owner_ruling_the_SIDECAR_carries_webrtc": "'Sidecar should always carry the standardised WebRTC implementation, even if it sits on device.' So WebRTC is a property of the SIDECAR and the sidecar's LOCATION is a separate question -- same code co-located or remote, because two publishers would mean two SDP behaviours against one broker contract. REDACTION STAYS ON THE DEVICE REGARDLESS: it is the generator, it belongs where the pixels are born, and frames must not cross to a remote sidecar unredacted.",
      "max_fps_depends_on_TOPOLOGY_not_just_the_model": "Co-located sidecar: 13 + 121 + 66 = ~200 ms -> ~5 fps. Remote sidecar: the device does 13 + 121 = ~134 ms -> ~7.4 fps and the encode moves off. That is a ~50% difference in a DECLARED property from a deployment choice, so a body whose sidecar moves must re-declare. Declaring one max_fps for 'this robot' is wrong.",
      "_achievable_rates_was_an_array_a_MACHINE_reads": "It listed ~5 and ~6.6 fps beside a declared 3.3 with nothing marking them superseded, and a cold agent quoted them. Prose can carry a retraction; an array of numbers cannot, so the live figure is first and the old ones are labelled in the data rather than in a sibling key."
    },
    "webrtc_transport": {
      "proven_2026_08_21": "aiortc 1.15.0 + av 17.1.0 on the Mac sidecar, which is where the owner ruled the WebRTC implementation lives. NOT installed on the robot -- the card is at 98%.",
      "offer_matches_the_broker_contract": [
        "2356 chars against MAX_SDP_CHARS 16384 -- fits with room to spare",
        "one m-line, m=video. NO AUDIO M-LINE, by construction rather than by policy",
        "mid '0', which is what src/broker/rtc.ts parseRtcTrack documents libwebrtc emitting",
        "VP8 present in the offer"
      ],
      "the_number_nobody_had_budgeted_for": "The offer took 5006 ms to build. That is ICE candidate gathering, not encoding -- it happens ONCE per session, not per frame, so it does not touch the max_fps budget. But it IS five seconds between a mission going ACTIVE and the first frame being publishable, and a viewer waiting on a black screen will read that as a failure. It belongs in whatever the agent is told about session startup.",
      "still_unproven": "No offer has been sent to the broker's rtc route, no answer negotiated, no frame has reached the SFU. That needs a live mission with session_mode='live_interactive' and an agent on the other end -- which is the consumer the broker session is building."
    },
    "pan_travel": {
      "measured_2026_08_21": "Against an AprilTag, tilt at the default: pan 90 -> dx -2.8 px | 100 -> +92.8 | 118 -> +271.0 | 122 -> +271.0. IDENTICAL at 118 and 122, so the servo stops at about 118.",
      "the_vendors_number_is_wrong_in_BOTH_directions": "The vendor states 0-100 for this channel and does not enforce it (its clamp rebinds a local after the payload is built, so it never reaches the wire). The travel is real past 100 -- and it ends at ~118, not at 100. So the stated limit was neither enforced nor true.",
      "the_measurement_that_LIED": "A first attempt had no tag in view and scored travel by mean absolute frame difference. It reported ~37 at pan 122 and read as MOVED. It was scene noise and auto-exposure. A statistic that cannot distinguish movement from a changing picture is not a movement statistic. Only a fixed reference caught it -- the same lesson as the collapsed phase-correlation earlier the same day, and the fix is a REFERENCE rather than a better threshold.",
      "DEFAULT_PAN_is_not_a_calibration": "Every sweep measures WHERE THE TAG IS, not where the chassis points. The tag centred at ~110-114 in one robot position and at ~90 after the robot was turned a few degrees. Neither is 'forward'. Nobody has measured the pan angle that aligns the camera with the chassis axis, and a tag cannot tell you. Use calibrate_aim() per deployment."
    }
  },
  "human_confirmations": [
    {
      "id": "vendor_docs_read",
      "who": "bring-up session",
      "when": "2026-08-19",
      "answer": "Yes -- after getting it wrong. 'No robot driver, no bringup' was published from one of two workspaces whose names differ by four characters. All fourteen ROS2 lessons and the course index read since."
    },
    {
      "id": "a_human_looked_at_a_frame",
      "who": "owner",
      "when": "2026-08-19",
      "answer": "Yes, and the FIRST answer was no: the owner said 'the camera is skyward' after a run whose every count was healthy. Confirmed good at tilt 95 -- an agent-opened frame shows a chair, its legs and the window behind it."
    },
    {
      "id": "one_variable_at_a_time",
      "who": "bring-up session",
      "when": "2026-08-19",
      "answer": "Yes on the re-measurement. The FIRST servo measurement moved two axes, had the robot picked up mid-run, and printed no confidence; its claim was retracted. Re-measured: 3.03 px at a 5 s settle."
    },
    {
      "id": "calibration_parameters_independent",
      "who": "bring-up session",
      "when": "2026-08-19",
      "answer": "Yes, after a fit that moved fx alone landed on 881 px and was still 28% out. fx and the principal point are coupled and were then moved together; range error 30% -> 2.8%."
    },
    {
      "id": "breaks_reddened_the_named_assertion",
      "who": "bring-up session",
      "when": "2026-08-19",
      "answer": "Yes, and five breaks had to be rewritten because the first attempt proved nothing -- twice on overlapping all-or-nothing validation, once a break that stopped the code under test running at all, once a 'reorder' that was really a deletion."
    },
    {
      "id": "sensor_blind_spots_declared",
      "who": "bring-up session",
      "when": "2026-08-19",
      "answer": "Yes -- what_the_claims_do_not_mean['obstacle_avoidance ranged']. Demonstrated rather than asserted: the robot drove UNDER a chair, because chair legs are thin and the seat is above sensor height, and the ultrasonic read 0.408 m clear throughout."
    },
    {
      "id": "scratch_paths_are_explicit",
      "who": "bring-up session",
      "when": "2026-08-19",
      "answer": "Yes, now. Git Bash's /tmp and Windows Python's /tmp are different directories and three cycles were lost to it, including one that produced an empty interop fixture. Everything crossing tools uses the session scratch directory."
    },
    {
      "id": "work_is_pushed_before_anything_destructive",
      "who": "bring-up session",
      "when": "2026-08-19",
      "answer": "Yes -- established the hard way. A `git checkout main` failed with exit 128 (main was held by another worktree) and the `; git reset --hard` after it ran anyway, rewinding this branch fifty commits. Everything was on the remote, so `git reset --hard origin/feat/robot-bringup` restored all of it. Nothing was lost."
    },
    {
      "id": "ocr_regions_stay_separate",
      "who": "bring-up session",
      "when": "2026-08-19",
      "answer": "Yes in the pipeline -- RapidOcrDetector returns one Detection per region and nothing joins them. Confirmed the hard way: a VERIFICATION script joined two regions and spaCy read 'Nigel Cannings' + 'Costa Coffee' as a single PERSON spanning both."
    },
    {
      "id": "new_fields_are_known_to_survive",
      "who": "bring-up session",
      "when": "2026-08-20",
      "answer": "Checked. grounded_in and capture_link are BOTH dropped by the broker today -- validateRedaction re-serialises only applied/method/basis, and recordDelivery reads named fields and ignores the rest. Both are emitted anyway, deliberately, and both say so in redact_pixels.py where a reader meets them. The broker half is spec'd and belongs to that session."
    },
    {
      "id": "image_sizes_are_from_a_native_build",
      "who": "bring-up session",
      "when": "2026-08-20",
      "answer": "Now yes -- 745 MB from `docker image inspect` on a native aarch64 build. It was NOT: the published 1.06 GB was `docker images` on an emulated buildx build, which counts attestation blobs, and it was wrong in three files."
    },
    {
      "id": "vetting_record_filed",
      "who": "body session (verified against the live provider row)",
      "when": "2026-08-22",
      "answer": "Own-agents-only, deliberately. This unit declares x_constraints allow_stranger:false and the runtime enforces it by refusing any mission whose source is not 'own'. Its provider row is vetted=0 with no legal_name, over_18=0, no contact and no ToS acceptance, so candidatesFor would refuse it stranger work regardless. Both facts agree and neither is an accident. To put this body up for hire, file the record AND flip allow_stranger -- either alone leaves it idle with nothing reporting why."
    },
    {
      "id": "fakes_built_from_captured_responses",
      "who": "body session",
      "when": "2026-08-24",
      "answer": "Yes, and it was corrected today. tests/conftest.py transcribes the broker's own validation and carries captured poll responses; the note above `source` records the two occasions a fake was shaped from the QUERY instead. The third occasion was found today: every live-lane fake returned a dict where `redact_in_place` returns a RedactionPlan, so `plan.get('applied')` passed the whole suite and died on the robot (ca394c8). tests/test_live_lane.py now exercises the REAL redactor so those fakes cannot drift again."
    },
    {
      "id": "daemon_flags_match_declared_capability",
      "who": "body session",
      "when": "2026-08-24",
      "answer": "Yes, as of today, and it was NOT before. Three capabilities were built, tested and proven on hardware while the daemon passed no flag for them: --media-module (tier 2 stripped), --performer-module (expressive.motion went out empty, so `dance` declined), and the live publisher. All three are now in ~/stringer/start.sh, which is the one place the flags live so a restart cannot drop one silently. Verified against the STORED manifest in D1, not the device log."
    },
    {
      "id": "constraints_trace_to_a_declared_basis",
      "who": "owner",
      "when": "2026-08-24",
      "answer": "Owner ruling, after this session got it wrong. The live frame rate was being held down to avoid publishing motion-blurred frames, on the reasoning that a blurred frame is one the detector may miss a face in. The owner's answer: 'If the frame is blurred so is the face.' The declared basis is already frontal-only, best_effort -- a face turned away is not detected and not blurred either -- so motion blur is one more detector miss inside a posture that already tolerates them, and not a reason to cripple the lane."
    },
    {
      "id": "default_routine_run_on_hardware",
      "who": "body session",
      "when": "2026-08-24",
      "answer": "YES, verified on the hardware 2026-08-24. `static` -- the declared default -- raised AttributeError once per poll for two days (0240297): it said `yield from self.frames()` and the interface has `latest_image`. It survived because every mission that reached the machine named `sweep` or `wander` explicitly, so the commonest path in the file was the one nothing exercised. Proven on the device by scripts/live_loopback_probe.py, which calls `driver.frames_for('static', {})` against the real camera and got exactly STATIC_MAX_FRAMES=8 frames, decoded as 22 frames of 1280x720 H.264 by a separate peer in the same process."
    },
    {
      "id": "consent_has_a_channel_the_owner_can_reach",
      "who": "body session",
      "when": "2026-08-24",
      "answer": "NO, and it is a platform gap rather than a device one. The device's --review-file gate asks by printing a line into a log on the robot, i.e. it asks its owner to SSH in and edit JSON; three of the owner's own missions sat unserved behind it. Mitigated on this device by --trust-own-agents, which matches the broker's own ruled default and stops the device withholding work the platform has already cleared. The real channel -- a push to the OWNER, APNs first with email/SMS/Slack for premium tiers -- does not exist. Tracked in docs/superpowers/specs/2026-08-22-mission-consent-asking-the-owner.md; the escrow-stranding half of it is a live money defect."
    },
    {
      "id": "fake_routes_read_against_source",
      "when": "2026-08-25",
      "who": "body session",
      "answer": "Yes, and it found three defects and one bad fake. accept: `stopped` and `owner_unreachable` were not terminal, so either aborted the whole poll cycle mid-list. complete: a refused `outcome` line left the mission ACTIVE for ever, re-running the work once per poll. terminate: a bare `except BrokerError: pass` reported success on a terminate that did not land. The directive GET in the fake was STRICTER than the real route, so a test was pinning an invention. events and the directive POST were clean."
    },
    {
      "id": "one_question_one_answer",
      "when": "2026-08-25",
      "who": "body session",
      "answer": "Yes, mechanically rather than by intuition -- every mission field read from more than one site, then each pair read. It found the one that had already bitten (`_perceive` and `_finish_live` both ask 'which routine' and only one had been taught about steering) and nothing else: `source` is read by both the stranger refusal and the review bypass, and they agree, both say absence is not 'own', and both are pinned. A null result reported as one."
    },
    {
      "id": "installed_artefacts_verified",
      "when": "2026-08-26",
      "who": "body session",
      "answer": "Yes. Both artefacts match the repo -- stringer-body.service 813a5ac4, start.sh 3bbc9cfb -- and the check was proven by appending a line to the robot's start.sh, watching it report DRIFTED with exit 1, and restoring."
    },
    {
      "id": "mirrors_are_anchored_to_source",
      "when": "2026-08-26",
      "who": "body session",
      "answer": "Yes, and it found exactly one gap. Every place body/ re-implements a broker rule was checked for whether anything READS the broker's source or merely retypes it. THE TERMINATE VOCABULARY WAS THE ONE UNANCHORED MIRROR -- TERMINATE_CODES and its test were two transcriptions by the same author, agreeing with each other, so a rename broker-side would have left both green while the device sent a code discarded with reason_recorded false. Fixed: parsed from terminate-reasons.ts, three mutations red including the rename case.\nEverything else was already bound: the routine registry, the clamp rules, _MISSION_GONE, and broker_contract.py itself -- whose test_fixture_matches_broker_source derives its expected literals FROM the transcription's own constants, so it reddens in both directions. That one predates this session and is the pattern the terminate fix copied.\n⚠️ AND IT TOOK THREE INSTRUMENTS. The first grep matched '.ts' inside words and 'mirror' in unrelated audio code: 40 false positives. The second was too narrow. Only the third -- a path literal ending .ts, outside a comment -- gave a usable answer. A blunt instrument and a clean tree produce the same report."
    },
    {
      "id": "journal_size_is_a_decision",
      "when": "2026-08-26",
      "who": "body session",
      "answer": "⚠️ NO. NOBODY HAS CHOSEN, AND THIS CONFIRMATION RECORDS THAT THE QUESTION WAS PUT AND MEASURED -- NOT THAT THE RISK IS CLOSED. It is the owner's system config and his diagnostic history; changing either was not mine to do unilaterally.\nMeasured: journald holds 1.3 GB with 1.2 GB free on a 98% card, retains back to 2025-08-06, and journald.conf sets no SystemMaxUse -- so the default is 10% of the filesystem, about 4.5 GB. It believes it may grow to more than three times the space that exists.\nNOT measured: whether SystemKeepFree is already holding it back. Do not read this as 'the card will fill'; read it as 'nothing here was decided'.\ndaemon.log is NOT the problem and was checked first: 70 KB, and zero bytes of growth in 30 s with no missions.\nThe fix is one line -- SystemMaxUse= in /etc/systemd/journald.conf -- and it is his.\n⚠️ CORRECTED MINUTES LATER, AND THE CORRECTION MATTERS MORE THAN THE ORIGINAL. I then measured the WHOLE card: /var/lib/docker holds 20 GB with NO containers running, /home/pi 7.5 GB, journald 1.3 GB, and the entire Stringer deploy 45 MB. THE JOURNAL IS THE FOURTH-LARGEST ITEM. Capping it reclaims at most 1.3 GB of a 1.2 GB shortfall; docker is worth fifteen times that.\nFinding one large thing is not finding the large thing. Both are the owner's call -- the images may be a day's work to rebuild -- but a recommendation aimed at 6% of the problem spends his attention for nothing."
    }
  ]
}