Measured Evidence & Test Records
We separate repeatable test-bench measurements from an operator's account of a NAS recovery. The incident account has a post-mortem, but no independent action log that identifies who performed each step.
Test boundary and constraints
Physical benchmarks were taken on one GL.iNet Comet KVM (firmware RM1 V1.10.1 release3) wired to one physical target machine (a mini PC called psycho). The benchmark records below cover that setup; use on other machines is listed on the usage page. The NAS recovery record is an observed first-person recovery event. Newer records add a Sipeed NanoKVM on a Raspberry Pi target, first runs of drivers for VNC consoles (against QEMU only) and for one Dell iDRAC9, PiKVM's software in a container, and the setup of one JetKVM with no driver yet. Each is one unit and a first run. We do not extrapolate to vendor-wide or fleet-wide claims.
Complete Evidence Registry
| Domain | Claim in one line | Measured figures | Evidence document |
|---|---|---|---|
| Real recovery | Diagnosed and fixed a real boot loop on a headless NAS | 180 s watchdog, under 400 ms boot window, 21 of 23 checks passed | NAS recovery account |
| Unattended install and rescue | A BIOS setup, a Debian install and a rescue card driven over the KVM, then the card booted on a second machine | 509 frames, 244 actions, 32 checks passing on the final run; mistakes listed | Install and rescue record |
| Host platforms | The unit suite passes on Linux and macOS in continuous integration; on Windows 11 it passed natively after a port, on one virtual machine; Windows use also appears in the operator account | Linux with Python 3.11 to 3.14 and macOS green; Windows 31 failures; 1052 passed locally | Platform evidence |
| Role of OCR | Local OCR read transient screens in the operator's NAS account | The field manual reports under 600 ms on its console frames. Separate BIOS tests measured 1081 to 1157 ms for Tesseract. | OCR role |
| Mouse accuracy | Target kernel sees absolute coordinates within 1 unit | Within 1/32767 in kernel; within 1 px in X (17 raw and 26 graphical checks) | Mouse accuracy |
| Typing reliability | Exact string delivery at full speed and slow mode | 25/25 fast, 25/25 slow (50/50 total trials) | Typing trials |
| BIOS OCR | RapidOCR and Tesseract performance on AMI Aptio setup | RapidOCR 0.98, Tesseract 0.94 (3 frames, 30 phrases) | BIOS OCR |
| Vision model vs OCR | Strict verbatim phrase reproduction comparison | VLM 31/32, Tesseract 22/32, RapidOCR 16/32 | Vision and OCR comparison |
| Boot-menu catch | Reached BIOS setup in one earlier benchmark run; no success rate measured | Delete tapped 113 times; SSH restored in 44 s | Boot menu catch |
| Comet quirks | Comet behaviors with live, source-only, and reported evidence | Eight catalogued behaviors; their evidence levels differ. | Comet behavior notes |
| Second KVM family | A Sipeed NanoKVM runs through the same driver layer: screen, typing, mouse and a mounted disk image the target saw. One unit, one target, a first run | Application 2.5.2; quirks found and handled are listed in the record | NanoKVM first runs |
| Keys and mouse against the target kernel | The target's own kernel sees the keys and mouse of two different KVMs as sent | 17 of 17 checks on each KVM, one run each, text console only | Input check, two KVMs |
| Healthy-machine baselines and reboot timelines | Reading a healthy machine through its KVM, ssh and the network together, then timing one reboot on each of two machines with two different KVMs | Ssh unreachable for 47.0 s and 35.9 s; 5 of 5 and 3 of 3 cross-channel facts agreed; one boot of each, first runs | Baselines and reboot timelines |
| Telling machines apart from the KVM side | What a KVM alone can say about the machine behind it, without reading the screen or logging in: a picture hash fails, the layout of the text does better, and the USB link gives a little | Twelve samples from each of 2 machines; the closest other machine was 1 to 2 bits away by picture hash and 0.0 apart by text layout; first measurements | KVM-side identification |
| VNC consoles, against QEMU | A driver for any VNC-speaking console reproduces a virtual machine's BIOS and UEFI screens exactly and sends keys into its firmware; no hardware controller has been tried | Identical to QEMU's own screen dump in three encodings on a text and a graphical screen; about 145 ms to fetch a screen; first runs | VNC first runs |
| Dell iDRAC9 screenshots | A read-only driver takes the server's screen through the management controller; it needs the console privilege on this firmware, which also allows typing | Refused to Read Only four times; a full-size picture in 1.48 s with the console privilege; one controller | iDRAC first runs |
| PiKVM software | The Comet client runs against PiKVM's own software in a container with fake hardware; no PiKVM unit, no video | Login, state reads and input calls accepted; four repeatable tests | PiKVM software in a container |
| JetKVM setup | One unit set up and its interface recorded; a scratch client got a frame; no Bootscry driver yet | First frame 1.2 s after the start of a run that included login | JetKVM day one |
The Comet BIOS video, which led the home page until early October, is a separate capture with 119 timed Delete taps. Its timing is not combined with the earlier benchmark. Its capture record has the frame timeline and file hashes.
ARM-board evidence: two ODROID-M1S boards extends the older-PC installation records. The watchdog failure required manual power cycling. Libre Computer Le Potato setup is operator-confirmed; its card was prepared on another host, not delivered through Bootscry. See the research agenda.
How each check was verified
Mouse positioning (Mouse accuracy)
The test script scripts/kvm/live_mouse_test.py generates raw mouse motion packets and drives an Xorg graphical session on Ubuntu 26.04 at 1920x1080. Coordinates are independently verified over SSH by querying the kernel input layer with evtest and the X server pointer position with xdotool. In 17 raw event checks and 26 graphical checks, motion delivered within 1/32767 in the kernel and within 1 px in X.
Keystroke delivery (Typing trials)
scripts/kvm/live_typing_test.py types strings containing mixed-case alphanumeric text, doubled letters, and punctuation marks. Strings are captured on the target via SSH and compared against the sent payload. All 25 strings of 64 characters arrived verbatim at full speed, and all 25 arrived verbatim in slow mode.
BIOS setup OCR (BIOS OCR)
Screen frames from 3 tabs of an AMI Aptio UEFI setup utility were captured at 1920x1080 using arrow keys only. 30 hand-verified phrases served as ground truth. RapidOCR scored a mean fuzzy phrase match of 0.98, while Tesseract scored 0.94.