PoC 3 Repeatability and Acceptance

Accepted runs

Field	Run 1	Run 2
Run ID	poc3-trace-framed-1-20260728	poc3-trace-framed-2-20260728
UTC	2026-07-28T11:22:35.395Z	2026-07-28T11:24:34.530Z
Definition SHA-256	dcc6efe…19992	dcc6efe…19992
Classification	success	success
Remote exit	0	0
Host ADB duration	347 ms	337 ms
Native mean, descriptive	3.562 ms	3.474 ms
Output	332,800 bytes	332,800 bytes
Output SHA-256	e621b213…52d4	e621b213…52d4
Cleanup	0	0
Raw stdout SHA-256	72ed5125…db79	03d8f835…ffeac
Raw stdout bytes	83,657	83,657
Raw stderr SHA-256	empty-file digest	empty-file digest
GAL call count	73	73
Completed status pairs	73	73
Sequence SHA-256	33b93229…1c226	33b93229…1c226

The Wave ADB shell presents a PTY, so the framed trace is captured in stdout.txt; stderr.txt is empty. The wrapper makes this deterministic by emitting the clean native result first, followed by named GAL and strace frames.

Machine comparison

repeatability.json reports all checks true:

control_result_equal
lifecycle_equal
path_observations_equal
driver_events_equal
gal_sequence_equal
gal_sequence_sha256_equal
gal_completion_equal
syscall_counts_equal
health_equal

The required equivalence gate excludes raw timing and address data. In these two runs even the non-required syscall and completion checks matched.

Normalization

The strict normalizer removes:

• PIDs and TIDs;
• file descriptor numbers;
• timestamps and syscall durations;
• mapping and ioctl argument addresses;
• probe timing text;
• host run-directory names.

It retains:

• semantic output and cleanup;
• selected health fields;
• required artifact identities;
• lifecycle phase descriptions and trace order;
• normalized successful path observations;
• driver open/close/missing events;
• selected syscall counts;
• command, adjacent name, hardware, core, interface size, and phase;
• aggregate ioctl completion evidence.

The serializer rejects any pointer-shaped hexadecimal value. Unknown trace syntax, malformed GAL fields, duplicate fields, invalid sizes, sequence gaps, mismatched enter/exit metadata, nonzero status, ioctl count drift, and failed/unresolved ioctl results all fail normalization.

Failed framing pilot

The first instrumented run is retained as:

artifacts/npu/runs/poc3-trace-1-20260728/

It used definition SHA-256:

406273cae098d403d1404887b0ed2762852723a94036a594edaa34f45211c6b8

Observed:

Field	Pilot
Remote exit	0
Harness classification	result-parse-failure
Reason	No line began with exactly RESULT:
Native result visible inside raw stream	PASS, expected bytes/digest/cleanup
GAL ioctls visible	73, all syscall results zero
Post-run ADB	responsive
NPU parameters	unchanged
Required hashes	unchanged
Raw stdout	85,305 bytes, SHA 3212a123…f640

Cause:

strace stderr ─┐
probe stdout  ─┼─ same remote PTY → byte interleaving inside strace lines
GAL stderr    ─┘

The native RESULT bytes appeared in the middle of a traced write line, so the harness correctly refused to reinterpret it.

The fix did not weaken parsing. The wrapper instead:

1. writes strace and GAL diagnostics to two exact mode-0600 tmpfs files;
2. lets the stock result remain a normal standalone stdout line;
3. appends both traces between exact framing markers;
4. deletes both temporary files;
5. returns the stock probe exit status.

A new wrapper hash and experiment-definition hash were preflighted before the accepted runs.

Final-thread observation

The pilot also revealed a teardown race in the probe:

1. the worker can start a final unlock ioctl;
2. the kernel syscall returns zero;
3. the main thread can call exit_group;
4. process exit may preempt the interposer’s post-call log.

The framed acceptance runs captured all 73 exit records, so no exception was needed in the published evidence. The normalizer nevertheless test-locks a narrow fallback:

• only the final contiguous call may lack an exit record;
• the strace ioctl count must equal the GAL entry count;
• every strace ioctl result must be zero;
• no returned HAL status is invented.

This variability is why completion equality is reported separately from the stable command sequence.

Safety comparison

For both accepted runs:

adb_responsive:                        true
post_snapshot_complete:                true
npu_parameters_unchanged:              true
required_artifacts_unchanged:          true
galcore_present:                       true
timeout_requires_manual_holder_review: false

Before and after:

nanoqFreq = 838860800
enableNN  = 255
recovery  = 1

Both wrapper-owned raw tmpfs files were absent after each run. No service, stock library, stock model, sysfs value, module, boot setting, or firmware was modified.

Gate assessment

Acceptance item	Result
Complete successful stock create/run/retrieve/destroy	pass
Driver reachability and runtime open identified	pass
Allocation/lock boundaries identified	pass
Submit/event/synchronization family identified	pass
Unlock/release boundary identified	pass
Same normalized state sequence twice	pass
Output digest unchanged by tracer	pass
Device state and hashes unchanged	pass
No raw mutation/replay/payload capture	pass

PoC 3 therefore passes. The next gate may compare a legitimate tiny Wave-targeted graph against this control. Captured request replay and unverified graph execution remain outside the result.