[devin-ai-integration]

Summary

software/conformance/ is the conformance harness: it proves the Argent flight host, the generated tables and the Meridian ground suite agree with the Tethys model as the pinned OpenSysML (5cd37f288) executes and evaluates it, and turns the software VCRM's planned flight-software-in-the-loop cases into as-run evidence. make software-conformance-check (root Makefile include after the flight include, CI job software-conformance-check after the flight job) builds the host, starts sysml-grpc from the pin, runs three tiers and compares every output byte for byte with software/conformance/expected/ and software/conformance/evidence/; make software-conformance-update rewrites them. Whole gate ≈ 50 s wall-clock, sysml-grpc ≈ 0.83 GB peak, so one tier — no fast/full split.

Counts

• Trace tier: 16 runs (the 14 pinned sw-* runs + vc-final-targeting, vc-thrust-arc) × machines = 149 machine-runs compared, 146 agree, 3 adjudicated divergences (F-SW1304 ×2, F-SW1305 ×1), 348 configuration changes, 434 model transitions, 282 model accepts compared.
• Calc tier: 47 calcs — agree 10 (C and Go, 1e-9 relative / exact), disagree 0, does-not-compile-at-pin 15 (quantity-typed inputs or results, package constants: F-SW1308), not-driveable 22 (record/sequence inputs; what covers each is in the file). No Python target at the pin (F-SW1307). The power margin 11.9 % vs 15 % and the bus slots 36 of 64 vs the 32 reserve (TF-011) come out identical on both sides and failing.
• Cases: 5 run / 5 pass / 0 fail / 26 still planned (12 unit, 4 integration, 4 hardwareInTheLoop, 6 flatsat — benches this repository does not hold). No planned case is at level modelInTheLoop.

F-SW1201 adjudication. The pinned run is right and the host was wrong: the model keeps a transfer routed to a port until an accepter takes it (SysML v2 §7.17.8 — an accept action accepts a payload received by its receiver; §7.18.3 — a transition fires when its accepter can accept an incoming transfer), so attitudeControl later takes the earlier ProximityOperationsGo/DockingGo from wheels. sm_runner now holds unaccepted items and offers them again; attitudeControl agrees in every run and the wildcard in compare_runs.py is gone. The model's own design point (rows that consume stale items) stays visible in the pinned run.

Flight-side edits the comparison forced (nothing under software/ground/, software/gen/, software/generated/):

• software/flight/sm/sm_runner.[ch] — held items (F-SW1301), unit test test_held_items.
• software/flight/host/argent_host.c — fast-forward stopped one frame short of an exact deadline (F-SW1302) and skipped the sequence engine's next step (F-SW1306).
• software/flight/tests/compare_runs.py + regenerated software/flight/scenarios/*.stim — exact campaign times instead of the render's six significant digits, and only this host's uplink (F-SW1303).

Model edit (model/88-software-assurance.sysml, runs re-recorded with make runs, make check green): for the five cases that ran, verdict = VerificationVerdict::planned; evidence = "not yet run" → verdict = VerificationVerdict::pass; evidence = "software/conformance/evidence/<case>.txt" — finalTargetingFlightSoftwareInTheLoop, thrustArcFlightSoftwareInTheLoop, captureSequenceFlightSoftwareInTheLoop, faultDetectionFlightSoftwareInTheLoop, groundTwoStepIntegration. Run diff: sw-assurance-passed 10 → 15 rows (PassedCases), sw-assurance-vcrm/sw-assurance-vcrm-document and sw-doc-test-plan carry the five verdicts and evidence paths, the VCRM and test-plan documents (runs/documents*/…SoftwareVcrmDocument, …SoftwareTestPlan) and render-documents-*.txt re-rendered — 10 files, nothing else moved. IV&V status, SA-001/SA-002 untouched.

Findings (docs/findings/WP-SW-13.md): F-SW1301 host dropped unaccepted items (fixed); F-SW1302 fast-forward off-by-one (fixed); F-SW1303 scenario derivation rounding/uplink (fixed); F-SW1304 host sequence-3 step items have no model source (open, adjudicated divergence); F-SW1305 ExecuteState over gRPC traces two FdirSupervisor instances in earthImpactScenario (pin; probe software/conformance/probes/F-SW1305-duplicate-fdir-instance.py); F-SW1306 fast-forward skipped sequence steps (fixed); F-SW1307 no Python -compile target; F-SW1308 -compile refuses quantity/record inputs and package constants; F-SW1309 client events are untimed and payload-less; F-SW1310 -compile cost (parallelised).

Decisions / alternatives rejected. Retention semantics fixed in the host rather than wildcarded or tolerated per machine; exact stimulus times derived by snapping to the driver script's delay gcd rather than hand-editing the shipped files or loosening the time comparison; the model side of every run is a scenario part (the client cannot inject timed payloads, F-SW1309) rather than CLI fallbacks; the duplicate FDIR instance is worked around by rebuilding the run's own usage and recorded, not hidden; compile steps parallelised rather than moved to a slower tier; calcs the pin cannot compile are bucketed with the reason rather than re-implemented by hand.

Coverage: docs/coverage/WP-SW-13.md. How to read a divergence and add a case: software/conformance/README.md.

Link to Devin session: https://nasa-jpl-demo.devinenterprise.com/sessions/a18b635f8fa84ae5848db4d43bbfb422
Open in Devin Desktop: https://nasa-jpl-demo.devinenterprise.com/desktop/session/a18b635f8fa84ae5848db4d43bbfb422?variant=devin
Requested by: @HuiJun

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring