CPU/AP Evidence Capture

This directory documents CPU/AP evidence expectations. Linux-capable AP evidence for the generated Chipyard ElizaRocketConfig target is archived under build/evidence/cpu_ap/, not committed here.

QEMU virt boot logs are software-reference evidence only. They do not satisfy the generated AP OpenSBI/Linux/trap/cache/benchmark gates.

Required generated-target evidence paths:

build/evidence/cpu_ap/eliza_e1_opensbi_boot.log
build/evidence/cpu_ap/eliza_e1_linux_boot.log
build/evidence/cpu_ap/eliza_e1_trap_timer_irq.log
build/evidence/cpu_ap/eliza_e1_isa_cache_mmu.log
build/evidence/cpu_ap/eliza_e1_ap_benchmarks.log

Use these commands to see the marker checklist, archive real generated-AP logs, and check the result:

python3 scripts/capture_cpu_ap_evidence.py template all
python3 scripts/capture_cpu_ap_evidence.py plan all --format shell
python3 scripts/wire_cpu_ap_capture_commands.py --format shell
scripts/capture_chipyard_linux_evidence.sh --help
scripts/capture_chipyard_linux_evidence.sh preflight
python3 scripts/check_cpu_ap_evidence.py --require-evidence
python3 scripts/check_chipyard_generated_linux_contract.py --require-boot-evidence

The capture wrapper requires one command environment variable per transcript. Each command must run the generated AP simulator or generated-target test and print the real transcript to stdout/stderr. The wrapper keeps raw output in build/evidence/cpu_ap/raw/ and only writes accepted evidence after the intake validator sees the required markers and binds the transcript to build/chipyard/eliza_rocket/ElizaRocketConfig.manifest.json.

Pass criteria

A generated AP evidence set must show all of the following across the accepted logs:

OpenSBI reset, CSR, timer, interrupt-controller, UART, DRAM, and next-stage handoff markers.
Linux early console, generated DTS hash, DT boot nodes, CONFIG_MMU, initramfs start, and e1 MMIO smoke result markers.
Trap, timer, software IRQ, PLIC external IRQ claim/complete, mret, and sret markers.
RV64GC, misa, riscv_hwprobe, Zicsr, Zifencei, Sv39, cache, TLB, and page-table markers.
Benchmark report hash, CoreMark/MHz, STREAM Triad, lat_mem_rd, fio, CPU frequency, run count, thermal state, and power method markers.

Linux Host Flow

On the Linux host, first generate/build the pinned Chipyard target and create build/chipyard/eliza_rocket/ElizaRocketConfig.manifest.json. Then derive the command variables that can be backed by checked-in generated-AP runners:

python3 scripts/wire_cpu_ap_capture_commands.py --format text
eval "$(python3 scripts/wire_cpu_ap_capture_commands.py --format shell)"
scripts/capture_chipyard_linux_evidence.sh preflight

The generated wiring exports ELIZA_OPENSBI_BOOT_CMD and ELIZA_LINUX_BOOT_CMD from the real scripts/run_chipyard_eliza_linux_smoke.sh path when the generated manifest and FireMarshal payload are present. It deliberately leaves ELIZA_TRAP_TIMER_IRQ_CMD, ELIZA_ISA_CACHE_MMU_CMD, and ELIZA_AP_BENCHMARKS_CMD unset until real generated-AP test or benchmark commands exist. Do not replace those with marker echo scripts, copied reference logs, or edited transcripts.

The two boot captures may use the same simulator command if one full boot log contains all OpenSBI and Linux markers. Trap/cache/benchmark captures should point at generated-target tests or scripts that emit those specific results. Do not edit raw transcripts to make the intake pass; fix the simulator payload or test command and rerun the capture.

Before starting a long simulator run, use the capture plan and preflight to confirm that all five command lanes are wired:

python3 scripts/capture_cpu_ap_evidence.py plan all --format shell
scripts/capture_chipyard_linux_evidence.sh preflight

preflight does not run the simulator or create evidence. It checks that build/chipyard/eliza_rocket/ElizaRocketConfig.manifest.json exists and that every ELIZA_*_CMD variable is set.

After all captures pass, copy the printed artifact_sha256.* and evidence_sha256.* values from python3 scripts/capture_cpu_ap_evidence.py hashes into the generated import manifest, then rerun the checks above.