AI and the Ecology of E-Waste: $14.33 pick unlocked a $13,900 AEC Scanner

Why the cleanest ecological case for AI isn't a Sustainability Report. It's taking a $14.33 lock pick to bypass proprietary software locks. Uncluttering the e-waste landfill.
Classification: Case Study · Published: 2026-10-29 · Status: production verified · Canon: axoworks.com/articles/ai-and-the-ecology-of-e-waste

This story was told by a human, cross-examined by AI models against each other, and edited by a human who signs it.

The Crime Scene

Nobody talks about the ecological crime scene at the back of the shop: six-to-eight-year-old commercial scanners, lab instruments, and industrial tools—packed with precision-milled aluminum, Fujifilm optical glass, stepper motors, and copper—hauled to the landfill in working mechanical condition because a recycler swapped a camera and the vendor locked the calibration math behind a $1,295.00 Windows-only license key.

In Q4 2026, that lock is broken. 4 days and $14.33 of AI tokens turned a $750 eBay door stop back into a $13,900 workhorse with better color than it ever had, on an OS the vendor never supported. Inside this open stack, we can vibe-code into whatever workflow we dream up next.

The ransom was never for the hardware. The hardware was fine. It was always just waiting for somebody to read its bytes.

The LockPickingLawyer Moment for Hardware

If you have ever fallen down a LockPickingLawyer rabbit hole on YouTube, you know this rhythm by heart.

A five-pound block of hardened steel lands on the bench. The box promises military-grade security. Then he slides a strip of spring steel from a used windscreen wiper into the keyway, applies light thumb tension, and narrates the takedown in seconds:

“Nothing on one… two is binding, nice click there… click out of three… four is in a false set… and we got this open.”

The heavy steel shackle was never the problem. The whole machine was held hostage by five tiny brass pins.

The Moat Breached

In The Moat Is Also a Prison, I wrote that proprietary software works the exact same way. A moat does 2 jobs: keeps the enemy out, keeps you locked in. I call it moat tax. Monthly rent on your own house. And I asked a simple question: If you stopped paying tomorrow, what would you still have?

What we reported: Q4 2026, we have watched the legacy pay-moat drained and breached. Vibe copied: Adobe, Office, AutoCAD, BIM with 90% of the features, 100% MCP for the future of work.

That pay-moat bypass? Weeks. Not years.

What changed. I received a shipment of artwork that needed to be scanned. Then I looked across the room at that 42-inch, two-person lift of a Contex scanner collecting dust in the corner. And I figured: if we can pick software locks, why aren't we picking hardware locks?

Here is the ledger on my studio floor:

• $750.02 for a 42-inch Contex HD Ultra i4250s scanner that carried a $13,900.00 list price new (receipts on file—$449.00 of it was actually the scanner; the rest was $562.83 in freight and $93.09 in tax, minus a −$354.90 partial refund).
• $109.20 for the wrong replacement camera when I blamed the hardware.
• Nearly four years as a dead door stop because the vendor wanted $650.00 to $1,295.00 just to unlock the scanning software—with zero guarantee the machine underneath even worked.
• Then 4 days and $14.33 of AI to make it sing—producing better color than it ever had, on an operating system the vendor never supported, inside a vibe-codeable open stack (Open.NEXT) that plugs straight into our AI workflows.

Unsetting the Pins, LockPickingLawyer Style

For those unfamiliar with lock picking, the goal is to get all the pins set. One by one.

Pin 1: Anatomy of a $13,900 Hostage

Why does a $13,900.00 commercial scanner with 1200 dpi optical resolution and 48-bit CCD imaging sell on eBay for $449.00? Because modern industrial equipment is built around a 3-pin software lock:

Pin 1A — Hardware Feature DRM: Contex ships 1 single giant scanner. Then sells you a la carte menu to 'activate' the features. Brilliant marketing strategy up to 2025. On their published price sheet:

• Unlocking a 36-inch bed to scan 42 inches? $475.00 (Mfg# 6799G525);
• Unlocking monochrome to 4-ips color? $2,375.00 (Mfg# 6799G537);
• Unlocking 4-ips to 8-ips color speed? $3,087.50 (Mfg# 6799G539).

Luckily, our eBay unit arrived with all of its hardware flags already activated. Full 42-inch width, HD, color, and Gigabit xDTR2.

Pin 1B — The Seller “Refurb” Calibration Trap: While our scanner was activated, the 5 cameras were completely uncalibrated. My seller was an equipment recycler, pulled the machine from a working shop, likely “refurbished” it by swapping a camera module or controller board without calibrating it. 3 months of head scratching later, I assumed it was faulty hardware. Spent $109.20 on a salvaged HP DesignJet HD Pro (G6H51A) camera, different scanner family but same OEM part. Result: same discoloration and streaks. Uncalibrated, in an analogue 5-camera system, each sensor has its own hue and character. Every scan was raw chaos, poisoning the B&W and grayscale math. It wasn't a broken part; it was an opaque 'calibration' lock.

Pin 1C — The NextImage Ransom & Version Treadmill: Calibration doesn't live entirely in the box; it also lived in the software you no longer have. Contex splits its Windows stack in two: WIDEsystem and NextImage. The former a free low-level driver/utility, the latter a massive, Windows-only app that locks into demo mode after a 30-day trial. You have a giant and expensive door stop once that trial expires. Here's the ransom:

• $650.00 for Nextimage SCAN+ARCHIVE (Mfg# 9691A601)—which strips out closed-loop ICC color profiling;
• $1,295.00 for Nextimage REPRO (Mfg# 9691A602);
• $445.00 for their Calibration Sheet, 42″ HD Ultra (Mfg# 6798G016)—and $880.00 for the Customer 'Care' Kit (Mfg# 2898A171), which is a lamp cartridge, a white background plate, and a micro fiber cloth.

Adding salt to the wound. Contex warns: “Nextimage 6 needs a new license to run and will not work with previous version licenses.” Every time Windows moves, your license dies. Cough up or head to the landfill.

Pin 2: Putting Tension on the Core (Wall #1)

Oct 24, 2026, we said 2 'no's. No e-waste. No ransom tolls. Started draining the Contex moat with DeepSeek Harness.

Our goal? Replace NextImage and WIDEsystem altogether with Open.NEXT. Our DIY native Linux driver, host-side calibration engine, and web-based application on Ubuntu.

We put 3 AI models on the bench, DeepSeek, GLM 5.3 Flash, and Gemini, on a zero-hardware solution. 'Focus on what you can control,' I reminded myself.

Immediately, we hit Wall #1: Contex did not publicly publish the communications protocol.

The scanner speaks SCSI-over-Ethernet—TCP 58170 for commands, UDP for the image stream, and UDP 58169 for discovery. Standard SCSI commands (INQUIRY, READ, WRITE BUFFER) are wrapped in undocumented vendor datagrams over a network cable. Yes, as gibberish as it sounds.

The NextImage trial was on the 30-day stopwatch. We needed every reference packet captured before the trial window slammed shut.

DeepSeek harness (DSH) did all the hard work. I, just the errand-boy loading the scanner and waiting for further directions. DSH decoded the captures byte by byte, mapped the capability pages (5 cameras, 42.00 in, firmware 3-5-1-015), and proved that raw RGB decoded cleanly while the vendor's proprietary color transform did not. Color would be our job, not the scanner's.

When our first raw decode correlated at 0.76 against the vendor's TIFF from the exact same raw scan, we felt the first pin on the lock set. Click out of Pin #2.

Pin 3: Two GitHub Shims and a 12:26 A.M. False Set

By late night on October 25, we were stuck in what lock pickers call a false set: the core rotated slightly for a brief moment that afternoon, and then never again. We could query the scanner all night, but no mechanical movements.

9pm. DeepSeek recognized the amount of time spent and advised me to literally 'stop' (session-678b645c-95dd-41f7-b2c4-2af4a3b0f4e0).

Needed fresh eyes. So I called in Gemini to recon the web. It found blueprints on GitHub:

• IvanRolero/WideScan gave us the Contex SDK C header and the full SET WINDOW struct.
• MerlijnWajer/contex-scanning gave us the Contex SDK reference: docs, DLLs, the FSS/ICD/005 specification, and the 28-function vendor API.

Yet the rollers still wouldn't budge. Why? Because our own repo's HANDOFF.md had recorded a fatal false assumption: “the vendor also sends no 1-byte window parameter — stop chasing this.”

12:26 a.m. October 26, Gemini cross-examined our repo's analysis against the raw C++ SDK source. It caught the lie the other AIs missed:

ctx_scan_2000.h:367, scanScan(hScanner, BYTE*, BYTE) is called in simplescan.cpp with a 1-byte buffer. The vendor does send the 1-byte window parameter.

One AI caught the other two models' blind spot in writing. In the very next session, we fired SCAN with that 1-byte window parameter. Status flipped to 0x24 SCANNING.

Click out of Pin #3.

Pin 4: The Bent Tension Wrench — Wall #2

2026-10-26 09:10am. The paper moved and bytes poured in.

241 records (707 KB), and then, once we matched the exact command shape and ACKed every single datagram, 8,464 records (24,850,304 bytes) in one burst.

The exhilaration soon turned to frustration. Getting Frankenstein to move doesn't mean its head is screwed on right. And how screwing on another head may be the solution. Here's what happened:

The scanner moved, but the raw image was shredded: ~26% of records were missing, every gap shifting the rows below it and dragging adjacent-row correlation down to 0.35.

For a day and a half of live captures (~22 hours), we blamed software. The flow control. Every counter read clean. All red herrings.

We were trapped inside our own framing.

So around noon on October 26, we did the thing that works: we screwed a different head onto the DeepSeek harness. A head with fresh eyes. An AI session with explicit permission to call us wrong. Its brief (FRESH-EYES-PROMPT) literally read:

“Do not treat any conclusion in the repo as established.”

We've been down that road before in our 3 Ways an AI Lies to You article. Ask if they made a mistake and they will defend their mistakes and lie when it pleases you.

GLM was the new head and fresh eyes. And its verdict exposed Wall #2, the one that had been invisible from Day 1. Our workstation's network patch cable was damaged. It had quietly negotiated 100 Mbit/s while both ends advertised Gigabit. Whenever the scan saturated the wire, the network switch dropped bursts of 14 to 18 datagrams (~41–53 KB) on egress-buffer overflow. Host error counters read zero because the dropped packets happened in the switch.

Swap to a good Cat5, Gigabit up, rescan: 98,505 / 98,505 records, zero holes, 288.9 MB, 16-bit RGB, full 5-camera width—geometrically verified against a calibration target at 300 dpi.

Lesson learned: Don't trust the AI, but don't trust your rig either. Cost of that entire wrong-diagnosis detour: 1 new patch cable. The humility was free.

The hardest part wasn't teaching our software to speak SCSI. It was finding out the phone line was cut.

Un-locked.

Beating the Maintenance Middle: Why Open.NEXT Outclasses NextImage

When the LockPickingLawyer opens a lock, he sets it down on the mat. When you pick a manufacturer's lock with an AI harness, you build a better machine than the one the factory sold you. More importantly, you are now an Owner, not a Renter. You get to remodel the house and keep the upgrades.

We disassembled the pieces and built Open.NEXT (a pun), and made it live calibrate with frontend and backend AI hooks to automate:

5-Camera Host-Side Calibration: We decoded the circular sensor readout, 13,640 pixels/row from the physical scanner bed. Applied per-camera optical black pedestals, flat-field curves, per-channel gains, and pre-calibration CIS red/green sensor line registration, and stitched a seamless 12,600 px / 42.000-inch @ 300 dpi composite with zero dead columns across all four joins.

Studio-Grade IT8 Color: Calibrated against the 264-patch Contex IT8 target (default-it8 preset), hitting a median ΔE*ab = 2.30. Delivered the closed-loop color accuracy Contex locks behind its $1,295.00 NextImage REPRO tier.

Thermal & Color-Drift Auto-Recalibration: Added per-scan multi-preset calibration profiles (architectural bond vs. artwork/photo) with backup, one-click rollback, and temperature-drift detection with live A/B preview.

Daily-Driver Polish & Physical Front Panel: On Day 4, we fed real sheets through the rollers—fixing narrow-sheet mis-registration and DPI vertical scaling, adding held scan settings, orientation control, cancel, and format-aware autosave, and wiring the physical front-panel buttons plus the ready LED. The NextImage REPRO tier was nothing more than a dumb watcher for the lo-tech button signals (SCAN 0x20, COPY 0x10, EMAIL 0x08).

Fully Customizable via Vibe Coding: NextImage is a frozen Windows .NET binary you cannot modify. Open.NEXT is clean, documented open-source code with a live WebSocket WebUI and an MCP server (ctx/mcp_server.py). Want a new feature? You vibe-code it in minutes. It already feeds straight into Photocraft for image processing. Now in the works: piping scans into PDFCraft for automatic OCR so our entire wide-format PDF collection is text-searchable, OCR-driven auto-file naming from drawing title blocks, and Vision-LLM auto-preset configuration that inspects the preview scan and selects the right profile automatically.

Frequently Asked Questions

Machine-readable FAQPage JSON-LD schema lives in the page <head> on axoworks.com and mirrors this section verbatim — see canonical schema definitions.

Can a Contex large-format scanner work without NextImage?

Yes. The scanner speaks SCSI-over-Ethernet (TCP 58170 commands, UDP image stream). We replaced the entire NextImage/WIDEsystem stack with our open-source Open.NEXT driver and app on Ubuntu Linux in 4 days, for $14.33 of AI tokens.

Why do scans from a used Contex scanner come out streaky or discolored?

Almost always the 5 cameras are uncalibrated, not broken. Each sensor has its own hue, dark pedestal, and gain; in an analogue 5-camera system the imbalance poisons even B&W mode. A seller who swaps a camera or board without recalibrating leaves you with streaks. Per-camera flat-field, black pedestal, and gain correction fixed ours.

How much does NextImage software cost?

$650.00 for Nextimage SCAN+ARCHIVE (Mfg# 9691A601) and $1,295.00 for Nextimage REPRO (Mfg# 9691A602), plus $445.00 for the 42″ HD Ultra calibration sheet (Mfg# 6798G016). Contex's post-sale upgrade sheet sells width and speed unlocks at $475.00–$3,087.50, and NextImage 6 will not run on previous-version licenses. (Source: wholesalescanners.com price sheet, verified 2026-10-29.)

Was the scanner broken?

No. The hardware was fine — feature flags activated, 48 kg of working optics and mechanics. It was a software and calibration lock, not a broken machine.

What is Open.NEXT?

Our open-source replacement for NextImage/WIDEsystem: a native Linux SCSI-over-Ethernet driver, 5-camera host-side calibration (13,640-px circular readout, per-camera pedestals, flat fields, gains), a WebSocket WebUI, front-panel button support, and an MCP server so AI agents can drive the scanner as a tool. github.com/Axotopia/contex_scanner_open

How good is the color without the vendor's proprietary pipeline?

Calibrated against the official 264-patch Contex IT8 target: median ΔE*ab = 2.30. That is studio-grade color accuracy, delivered free of the $1,295 REPRO tier.

What did the whole build cost?

$14.33 of AI tokens across 4 days: DeepSeek $9.35 (3,533 requests, 1.77B tokens), GLM 5.3 Flash $4.98, Gemini $0.00 on the free tier. The wrong-diagnosis detour cost one Ethernet patch cable.

What is “software bricking”?

Landfilling a mechanically working machine because its software license expired. The hardware survives; the owner's permission to use it doesn't.

The Final Word

The ransom was never for the hardware. The hardware was fine. It was always just waiting for somebody to read its bytes.

Three AIs spent $14.33 and four days to prove a $13,900 machine was never junk — it was hostage. The vendor's price sheet sells you your own scanner back, feature by feature. Our price sheet says: Owner, not Renter.

Next time someone counts the watts of an AI prompt, show them the 106-pound scanner still humming at the back of the shop. Tiny is Mighty. Believe it!

Appendix A: The Receipts & Forensic Timeline

All [SESSION_TS] markers below were populated 2026-10-29 from the local git log (/home/song/Documents/DSH/Scanner) and the DSH session transcripts (~/.dsh/sessions/--home-song-Documents-DSH-Scanner--/, zstd JSONL v4, per-record epoch-ms timestamps in America/Los_Angeles).

Pre-History — The $750 Door Stop

2023-02-07 [eBay Invoice]: Purchased Contex HD Ultra i4250s (original list price $13,900.00). Paid $449.00 item + $562.83 freight + $93.09 tax − $354.90 partial seller refund = $750.02 net. Arrived with hardware flags activated (42″, HD, color, Gigabit), uncalibrated cameras (likely seller refurb with swapped camera/board), free WIDEsystem utility, and 30-day NextImage trial (full license $650.00–$1,295.00).

2023-05-08 [eBay Invoice]: Purchased HP DesignJet HD Pro (G6H51A) camera for $109.20 to fix streaking. Same streaks remained—confirming the issue was 5 uncalibrated sensors, not a single broken camera.

Day 1 — Protocol Capture, Ground-Truth Race & First Prior Art (Oct 24, 2026)

2026-10-24 08:11 [DSH Sprint Kickoff — session-f6e578ce-cf58-45ba-858b-8195e469e160]: Launched clean-room Linux driver/app sprint inside DSH using DeepSeek, GLM 5.3 Flash, and Gemini (Antigravity + free tier). Enforced non-destructive rule: SCSI WRITE BUFFER 0xA0 (RAM→flash commit) blocked unconditionally in ctx/proto/cdb.py. (First user message logged 08:11:20 PDT; the 0xA0/0xA2 block was designed 08:20–08:30 in this session and written into ctx/proto/cdb.py 2026-10-25 15:42–15:50 in session-678b645c-95dd-41f7-b2c4-2af4a3b0f4e0, first committed in 224db35.)

2026-10-24 14:06 [Trial-Expiry Capture Pass — session-20908fa5-ef0c-489f-8cf7-c390bc77dd28]: Captured ground-truth pcap traces from vendor software inside the expiring 30-day NextImage trial window (HANDOFF-COMMS.md). Mapped SCSI-over-Ethernet transport: TCP 58170 (commands), UDP (image stream), UDP 58169 (discovery). (Trace analysis 14:06–14:20; HANDOFF-COMMS.md written 2026-10-25 15:40 → commit 4e40015.)

2026-10-24 15:40 [Capability & Decode Check — session-20908fa5]: Extracted scanner identity (5 cameras, 42.00 in, DPI ladder, firmware 3-5-1-015). Verified raw RGB decode at 0.76 correlation against same-scan vendor TIFF; documented cross-scan mismatch trap (0.06 correlation). Initialized SPECS.md, DESIGN.md, and STATUS.md. (Capability pages 15:40–15:42; raw-RGB 0.76 correlation result 19:38–19:42 same session. SPECS/STATUS/DESIGN landed as commits 224db35 15:12, d3ab540 15:14, ad2bb05 15:17, e926474 15:35 on 10-25.)

2026-10-24 15:38–19:13 [HANDOFF.md Rev 2 / Rev 3 — session-20908fa5]: Surfaced first GitHub prior-art repo: IvanRolero/WideScan containing Contex SDK C header and SET WINDOW struct. (GitHub hit logged 15:38; evening HANDOFF.md revisions written 19:13; captured in initial commit 224db35, later handoff update 39d063c 10-26 08:40.)

2026-10-24 13:58 → 14:21 [First Hand-Off & Session Pause — session-f6e578ce + session-20908fa5]: DeepSeek closed out Day 1 at the user's request. At 13:58 the user said “Give me a detailed hand-off on everything we have learned to use on another session. The context on this session is getting too big”, and DeepSeek wrote analysis/HANDOFF.md (1,185 lines, 43 sections as committed) at 13:59 — including “the one unresolved problem: the image cannot be reassembled, with every stride and interpretation I tried listed so the next session doesn't repeat them” — then ran the partial REA pass on CtxScan64.dll (14:01–14:03, host-side stitching + 14-byte stitch values) before pausing. In the follow-up session, at 14:19 the user said “just add and revise the handoff for now, we will continue another time”, and DeepSeek recorded the day's bottom line at 14:21: “extracting a 4848 B/line raster is 1–2 h of work; assembling a correct image is ~1 week, gated on a capture that actually contains a drawing.” (DeepSeek issued no stop-working advice on Oct 24 itself — searched all sessions for “call it a night / long day / rest” phrasing. The day's last message came at 19:50:18 — the “honest odds” wrap (“framing solved tonight/tomorrow: ~75 %... Want me to start step A now?”), which went unanswered overnight; no further Oct 24 activity follows in either session.)

Day 2 — First Live Captures, The Invisible Cable Wall & Second GitHub Find (Oct 25, 2026)

2026-10-25 14:06 [logrun-1025-1406.pcapng — session-20908fa5]: First live capture run initiated. Unknown to the team, a damaged workstation patch cable negotiated 100 Mbit/s instead of Gigabit, silently dropping 14–18 datagram bursts (~41–53 KB / ~26% of records) inside the network switch egress buffer while host error counters read zero.

2026-10-25 14:49 [IP Fragmentation Fix — session-20908fa5, commit d4138f3]: Diagnosed and fixed 49.5% record loss caused by standard IP fragmentation (initially misdiagnosed in docs as an “MTU artifact”). (Loss diagnosed 14:49–14:57, reassembly fix documented 15:09–15:35; the diagnosis correction was committed as d4138f3 on 10-26 13:06.)

2026-10-25 15:12 [Git Commit #1 — 224db35]: Initial repository commit landed. (“Contex HD Ultra i4250s reverse engineering: protocol notes, tools, decode findings”)

2026-10-25 20:46 → 21:00 [DeepSeek Recognizes the Hours Spent and Advises Stopping — session-678b645c]: Unprompted, three times in fifteen minutes, DeepSeek flagged the length of the working session and recommended ending the day: at 20:46 its reasoning read “It's late; I've been at this a long time; the user has been patient” and “The user has been at this for hours. Let me be brief and give a clean handoff,” telling the user “I'd stop here for tonight. The next session doesn't need the scanner at all”; at 20:55 “Recorded. Now the final reply for tonight… I'd stop here for tonight”; and at 21:00:40, after the ninth straight hypothesis elimination (control-from-TCP-port → 120 frames, nz=0), its reasoning stated “It's ~21:15, end of a very long session” and “So my recommendation: 1. Tonight: stop.” — delivered to the user as “Short answer: stop for tonight, then work in this order.” followed by a four-point plan (the §12.5 warm-state test next session; accept the Nextimage-scans/we-capture split; spend the trial window — expiring 2026-11-07 — on the missing vendor captures; finish the driver offline from the matched pair). Each stop recommendation was paired with a concrete next-session plan rather than a bare “quit,” and the user pushed back once more before accepting. (The stop advice was workload-recognition — hours spent, diminishing returns — not a wellbeing statement; DeepSeek never used “overworking” or rest/health language. Four minutes later it closed the day with the “Night.” sign-off at 21:04:34.)

2026-10-25 21:04 [The “Night.” Sign-Off — DeepSeek Advises Stopping for the Day — session-678b645c]: At 21:04 the user asked, “before I call a night, anything that we failed to monitor between the Windows station and scanner?” DeepSeek answered with the four monitoring blind spots (pre-Start connection setup, local IPC needing procmon, the driver's own SCSI log, process attribution), shipped the upgraded 04-triple-collect.ps1, and closed the working day with: “Night. The state is clean: no writes to the scanner, no flash commands, and everything we established is written down in COMMS-LAYER.md §10–12 and STATUS.md §17.” — the sprint's only explicit end-of-day sign-off from DeepSeek, confirming no persistent state was touched. (Timestamp 21:04:34 PDT; the same transcript is mirrored in session-91c73dd8-07f7-4e09-8242-c4aa7d34c618.)

2026-10-25 23:36 [Repo Documentation — session-678b645c, commit 088c055]: Surfaced second GitHub prior-art repo: MerlijnWajer/contex-scanning with FSS/ICD/005 spec, DLL exports, and 28-function vendor API—explaining the 5-camera line structure (8,892 B single-camera 8-bit stride @ 300 dpi vs. 81,840 B full 16-bit 5-camera row) and 32-command OEM arming sequence. (MerlijnWajer prompt spliced into the session at 23:28; documentation commit 088c055 landed 23:36, digest-pinning follow-up 78c7203 23:38.)

Day 3 — The 12:26 A.M. Cross-Examination, First Motion, Fresh-Eyes Verdict & 5-Camera Stitch (Oct 26, 2026)

2026-10-26 00:26 [CONTEX-SDK-AND-LINUX-DRIVER_GEMINI.md — commit 0166a2d]: Gemini cross-examined repo HANDOFF.md claim (“the vendor also sends no 1-byte window parameter — stop chasing this”) against SDK source (ctx_scan_2000.h:367, simplescan.cpp) and proved the vendor does send the 1-byte parameter in scanScan(hScanner, BYTE*, BYTE).

2026-10-26 09:10 [First Paper Motion — session-375359e3-a31c-4bfe-aca1-f7970136d23f, commit cee2a0b 08:58]: Sent SCAN with 1-byte window parameter; status flipped to 0x24 SCANNING, rollers engaged, and first nonzero bytes arrived (241 records / 707 KB), though missing ~26% of records (0.35 adjacent-row correlation).

2026-10-26 09:31 [START_PROPS & ACK Gate — session-375359e3]: Matched exact START_PROPS record shape and switched to ACKing every datagram instead of every 32nd; burst throughput jumped to 8,464 records (24,850,304 bytes), but 26% packet holes persisted. (First hit on 8,464 logged 09:30–09:32; the ACK-policy change itself was coded 2026-10-25 16:35 in session-678b645c.)

2026-10-26 12:18 [FRESH-EYES-PROMPT / FRESH-EYES-VERDICT — session-a5f0dcac-2e5a-4c49-bdeb-aa4044ae97ef, commits 11a2e8b / 96569ff 12:18]: After ~22 hours of chasing false flow-control theories since the first live capture, launched an adversarial fresh-eyes AI session instructed: “Do not treat any conclusion in the repo as established” (Authority order: raw bytes > measured facts > prior session notes). Verdict isolated the 100 Mbit/s damaged patch cable and switch egress buffer overflow. (Prompt written 10:15 in session-375359e3; fresh-eyes session opened 10:17; the 100-Mbit finding recorded 11:46–11:47.)

2026-10-26 11:47 [Gigabit Clean Scan — session-a5f0dcac, commit b1834d6 12:24]: Swapped damaged cable for good Cat5 (Gigabit link up). Rescan captured 98,505 / 98,505 records, zero holes, 288.9 MB, 16-bit RGB across all 5 cameras at 300 dpi. (98,505 first observed 11:46–11:47; format-stability experiments committed 5195b72 12:31 and 6a20a3c 12:42.)

2026-10-26 17:34 [5-Camera Stitch & IT8 Calibration — session-95c422ad-27bc-4009-8bad-927a13168a3c, commit c3434ff 17:46]: Unpacked circular row readout ([cam5R][gap][cam1..4][cam5L], 13,640 px/row), applied per-camera black pedestal, flat-field, and gain corrections, and stitched seamless 12,600 px / 42.000″ @ 300 dpi composite with zero dead columns. Calibrated against 264-patch Contex IT8 target to median ΔE*ab = 2.30 (default-it8). Refuted vendor pitch-offset hypothesis. (Session opened 15:52; 16-bit stitcher committed 15436a0 16:46; ΔE 2.30 computed 17:34; calibration committed c3434ff 17:46.)

2026-10-26 22:57 → 2026-10-27 10:32 [WebUI, Front Panel & MCP — session-678b645c (ends 22:56), commits 5f2604c 22:57, 85b7ea0 10:27, 858f2c8 10:32]: Shipped WebSocket WebUI, mapped front-panel hardware buttons (SCAN 0x20, COPY 0x10, EMAIL 0x08 + ready LED; commits 85b7ea0, 858f2c8), and shipped MCP server (ctx/mcp_server.py, commit 5f2604c 10-26 22:57; panel key-map commit 85b7ea0 landed 10-27 10:27, key monitor 858f2c8 10:32).

Day 4 — Hardening the Daily Workhorse & Final Audit (Oct 27, 2026)

2026-10-27 08:36 [Real-Sheet Tuning — session-29197d7c-8661-4d53-bbbb-3c2f4ca6720d, commits eb6eb14 09:14, e0582c8 09:42, 12a82c3 12:15]: Fixed narrow-original auto-width registration, DPI-dependent vertical scaling, and pre-calibration CIS red/green sensor line registration. Added held scan settings, orientation control, cancel, format-aware autosave, multi-preset profiles (bond vs. artwork) with rollback, and thermal-drift A/B preview. (CIS registration eb6eb14 09:14; narrow/dpi fixes e0582c8 09:42; held settings/orientation/cancel/autosave 12a82c3 12:15, built in session-432beebf-39a3-4357-9c12-27a9ff96e6d4 (opened 10:59); artwork seams/calibration workflow 9f6d021 14:55.)

2026-10-27 10:45 [Security Audit & Commit 0fb2e5d — physical verification in session-9d6082cf-f44f-4912-94fe-a697bd0adf91 / session-debc1db3-0a6b-47b1-a4d4-e383b4b25112]: Published internal security audit and recorded end-to-end physical front-panel button verification (0fb2e5d 10:45). (docs/SECURITY-AUDIT-2026-10-27.md was drafted 11:14–11:16 in session-9d6082cf and left untracked; 0fb2e5d “Record front-panel end-to-end verification” is the audit-day anchor commit.)

2026-10-27 14:55 [Git Commit #67 — 9f6d021]: Final sprint commit landed (67 total commits). Total AI compute cost: $14.33 ($9.35 DeepSeek across 3,533 requests / 1.77B tokens + $4.98 GLM 5.3 Flash + $0.00 Gemini free tier).