How a 2019 MacBook Pro that failed every macOS install was diagnosed down to a single memory chip, then brought back to life with a memory cap and a page-pinning daemon.
If your Intel Mac (2018 to 2020, the T2 generation) fails every macOS install with errors like these, this post is for you:
- OSISUnarchiveMobileSoftwareUpdateOperation failed, error 1004, "An error occurred loading the update"
- "Installation cannot proceed because the installer is damaged"
- pkgdmg checksum does not match, possible corruption
- ZIP bad CRC errors inside installer payloads
My 16-inch 2019 MacBook Pro produced all of them. Apple Diagnostics found nothing. Disk Utility said the SSD was healthy, SMART status: Verified. A DFU revive and a full restore changed nothing. Internet Recovery, USB installers, different networks, fresh downloads: every attempt died during verification, and here is the detail that turned out to matter most: a different file failed every time.
The tell: corruption that moves
A bad download fails in the same place on every retry. A bad disk sector fails in the same place too. Corruption that lands on a different file each run is not sitting anywhere. It is happening in transit, while data moves. During an install, the busiest transit point in the machine is RAM: gigabytes are staged in memory, unarchived, and verified byte for byte. Installers are one of the few workloads that both fill memory and check everything that passed through it. That is why a machine with bad RAM can browse, boot utilities, and pass diagnostics, yet fail every single install at the verification stage.
It is also why the usual tests lie. The faulty region of memory only gets used once RAM is nearly full. Small tests never touch it.
The method: an AI agent inside Recovery, running controlled experiments
Guessing was getting nowhere, so I changed the setup. From a second, older Mac, I put an AI coding agent (Claude Code) in the driver's seat and gave it a root shell inside the sick Mac's own Recovery environment, using the open-source recovery-ssh toolkit from aurora-silicon (https://github.com/aurora-silicon/recovery-ssh), which builds a reverse SSH relay into macOS Recovery. The agent could then run experiments on the patient directly, for hours, without anyone retyping commands at a Recovery terminal.
Instead of swapping parts, it ran a controlled matrix against one known-good 18 GB installer image, hashing it through every path the machine offers:
- Streamed over Ethernet from the healthy Mac: identical, correct hash, twice.
- Read raw from the internal SSD: identical, twice.
- Read raw from an external SSD: identical, twice.
- Small files and small RAM tests: perfect.
- Anything that filled RAM: failed. A pure RAM-disk test, 8 GB written once and hashed three times with no disk, no USB, and no network involved, returned three different hashes.
Storage, network, CPU, and the T2 were all exonerated by elimination. Only memory remained.
The smoking gun: one byte lane
The agent then read the same 256 MB chunk twice and diffed the two copies at byte level. In every differing line, only the bytes at offset 7 (mod 8) had changed: always the same byte of every 64-bit word, bits 56 to 63, the whole byte replaced. On a DDR4 channel, one byte lane is served by one specific DRAM chip. This was not scattered cell decay. It was a single faulty chip (or its solder joints) on one lane.
A purpose-built memory test confirmed the scale: 8 bad 4 KB pages, every single error in byte lane 7. Capping memory with maxmem boot arguments bounded the bad rows to somewhere between 4 GB and 12 GB of physical RAM. Total damage: a few dozen kilobytes of a 16 GB machine. Apple Diagnostics, which never stresses memory this way, reported a clean bill of health throughout.
The fix: cap it to install, pin the bad pages to run
With the fault mapped, the recovery had three steps:
1. Install under a memory cap. Setting nvram boot-args="maxmem=4096" (4 GB) kept the installer entirely below the bad rows. For the first time ever, every installer hash verified correctly, and macOS installed cleanly, start to finish. The cap was removed afterward (nvram boot-args=""), restoring the full 16 GB.
2. Pin the bad pages at boot. The agent wrote a small daemon that runs at startup, scans memory, finds the pages that fail, and locks just those pages away so the system never allocates them, freeing everything else. Cost on this machine: 12 pages pinned, 48 KB, with a safety margin. A daily watchdog re-scans and logs any change, so if the fault ever grows, it is caught.
3. The permanent repair is one chip. Because the fault is a single DRAM chip on a known byte lane, a board-level repair shop can replace that one chip instead of the entire logic board Apple would quote. That is the difference between a repair and a write-off, and a written diagnosis (fault, lane, page map) is what makes the shop's job possible.
Rules I now run under, and would recommend to anyone using this mitigation: never run a macOS update without re-applying the memory cap first (install under the cap, verify, then clear it), keep real backups, and treat the daemon as a bridge to the chip replacement, not a cure.
What to take from this
- "An error occurred loading the update" plus a different failing file each attempt is a data-integrity signature, and RAM belongs at the top of the suspect list, above the SSD and the network.
- A passed Apple Diagnostics does not clear memory. It misses load-dependent faults by design.
- SMART Verified does not clear the disk either, but controlled hashing can: if the same large file reads back identically via disk, network, and external media, your storage is not the problem.
- Soldered RAM failing does not automatically mean a dead logic board. One chip can be mapped, worked around in software, and ultimately replaced on its own.
- The practical enabler was leverage: an AI agent with a real shell inside Recovery, running patient, repeatable experiments through a relay like aurora-silicon/recovery-ssh (https://github.com/aurora-silicon/recovery-ssh), turned a "haunted" machine into a measured one. The diagnosis took hours of machine time and almost no human time at the keyboard.
"Unrepairable" very often just means "not yet diagnosed." If your old Mac fails installs the way mine did, do not let the first verdict be the last one.
Related accounts of the same failure
- A board-level repair tech's 2019 16-inch MacBook Pro (A2141) with the identical installer log signature, still unresolved in the thread: https://forums.macrumors.com/thread...os-install-issues-nand-wifi-firmware.2367576/
- An iMac whose Code 1004 payload CRC mismatches were eventually traced to faulty RAM (removable in that machine, soldered in ours): https://www.rezteksystems.com/2021/10/an-evening-at-the-genius-bar/
- The same error family, documented in the OpenCore Legacy Patcher issue tracker: https://github.com/dortania/opencore-legacy-patcher/issues/786
If your Intel Mac (2018 to 2020, the T2 generation) fails every macOS install with errors like these, this post is for you:
- OSISUnarchiveMobileSoftwareUpdateOperation failed, error 1004, "An error occurred loading the update"
- "Installation cannot proceed because the installer is damaged"
- pkgdmg checksum does not match, possible corruption
- ZIP bad CRC errors inside installer payloads
My 16-inch 2019 MacBook Pro produced all of them. Apple Diagnostics found nothing. Disk Utility said the SSD was healthy, SMART status: Verified. A DFU revive and a full restore changed nothing. Internet Recovery, USB installers, different networks, fresh downloads: every attempt died during verification, and here is the detail that turned out to matter most: a different file failed every time.
The tell: corruption that moves
A bad download fails in the same place on every retry. A bad disk sector fails in the same place too. Corruption that lands on a different file each run is not sitting anywhere. It is happening in transit, while data moves. During an install, the busiest transit point in the machine is RAM: gigabytes are staged in memory, unarchived, and verified byte for byte. Installers are one of the few workloads that both fill memory and check everything that passed through it. That is why a machine with bad RAM can browse, boot utilities, and pass diagnostics, yet fail every single install at the verification stage.
It is also why the usual tests lie. The faulty region of memory only gets used once RAM is nearly full. Small tests never touch it.
The method: an AI agent inside Recovery, running controlled experiments
Guessing was getting nowhere, so I changed the setup. From a second, older Mac, I put an AI coding agent (Claude Code) in the driver's seat and gave it a root shell inside the sick Mac's own Recovery environment, using the open-source recovery-ssh toolkit from aurora-silicon (https://github.com/aurora-silicon/recovery-ssh), which builds a reverse SSH relay into macOS Recovery. The agent could then run experiments on the patient directly, for hours, without anyone retyping commands at a Recovery terminal.
Instead of swapping parts, it ran a controlled matrix against one known-good 18 GB installer image, hashing it through every path the machine offers:
- Streamed over Ethernet from the healthy Mac: identical, correct hash, twice.
- Read raw from the internal SSD: identical, twice.
- Read raw from an external SSD: identical, twice.
- Small files and small RAM tests: perfect.
- Anything that filled RAM: failed. A pure RAM-disk test, 8 GB written once and hashed three times with no disk, no USB, and no network involved, returned three different hashes.
Storage, network, CPU, and the T2 were all exonerated by elimination. Only memory remained.
The smoking gun: one byte lane
The agent then read the same 256 MB chunk twice and diffed the two copies at byte level. In every differing line, only the bytes at offset 7 (mod 8) had changed: always the same byte of every 64-bit word, bits 56 to 63, the whole byte replaced. On a DDR4 channel, one byte lane is served by one specific DRAM chip. This was not scattered cell decay. It was a single faulty chip (or its solder joints) on one lane.
A purpose-built memory test confirmed the scale: 8 bad 4 KB pages, every single error in byte lane 7. Capping memory with maxmem boot arguments bounded the bad rows to somewhere between 4 GB and 12 GB of physical RAM. Total damage: a few dozen kilobytes of a 16 GB machine. Apple Diagnostics, which never stresses memory this way, reported a clean bill of health throughout.
The fix: cap it to install, pin the bad pages to run
With the fault mapped, the recovery had three steps:
1. Install under a memory cap. Setting nvram boot-args="maxmem=4096" (4 GB) kept the installer entirely below the bad rows. For the first time ever, every installer hash verified correctly, and macOS installed cleanly, start to finish. The cap was removed afterward (nvram boot-args=""), restoring the full 16 GB.
2. Pin the bad pages at boot. The agent wrote a small daemon that runs at startup, scans memory, finds the pages that fail, and locks just those pages away so the system never allocates them, freeing everything else. Cost on this machine: 12 pages pinned, 48 KB, with a safety margin. A daily watchdog re-scans and logs any change, so if the fault ever grows, it is caught.
3. The permanent repair is one chip. Because the fault is a single DRAM chip on a known byte lane, a board-level repair shop can replace that one chip instead of the entire logic board Apple would quote. That is the difference between a repair and a write-off, and a written diagnosis (fault, lane, page map) is what makes the shop's job possible.
Rules I now run under, and would recommend to anyone using this mitigation: never run a macOS update without re-applying the memory cap first (install under the cap, verify, then clear it), keep real backups, and treat the daemon as a bridge to the chip replacement, not a cure.
What to take from this
- "An error occurred loading the update" plus a different failing file each attempt is a data-integrity signature, and RAM belongs at the top of the suspect list, above the SSD and the network.
- A passed Apple Diagnostics does not clear memory. It misses load-dependent faults by design.
- SMART Verified does not clear the disk either, but controlled hashing can: if the same large file reads back identically via disk, network, and external media, your storage is not the problem.
- Soldered RAM failing does not automatically mean a dead logic board. One chip can be mapped, worked around in software, and ultimately replaced on its own.
- The practical enabler was leverage: an AI agent with a real shell inside Recovery, running patient, repeatable experiments through a relay like aurora-silicon/recovery-ssh (https://github.com/aurora-silicon/recovery-ssh), turned a "haunted" machine into a measured one. The diagnosis took hours of machine time and almost no human time at the keyboard.
"Unrepairable" very often just means "not yet diagnosed." If your old Mac fails installs the way mine did, do not let the first verdict be the last one.
Related accounts of the same failure
- A board-level repair tech's 2019 16-inch MacBook Pro (A2141) with the identical installer log signature, still unresolved in the thread: https://forums.macrumors.com/thread...os-install-issues-nand-wifi-firmware.2367576/
- An iMac whose Code 1004 payload CRC mismatches were eventually traced to faulty RAM (removable in that machine, soldered in ours): https://www.rezteksystems.com/2021/10/an-evening-at-the-genius-bar/
- The same error family, documented in the OpenCore Legacy Patcher issue tracker: https://github.com/dortania/opencore-legacy-patcher/issues/786