Logs in, root cause out.

An 802.11a-style OFDM test harness, and a model that triages its failures.

Built by Akshit Gupta

Read the code The results

Pilot phase per symbol, two failing 16QAM packets

tc_00017 · EVM −10.41 dB, limit −16.0

Injected: cfo_not_corrected
Slope, off-line wander: 45.6°/sym, 3.8°
Rules: cfo_not_corrected
Trees: cfo_not_corrected 1.00

tc_00351 · EVM −12.42 dB, limit −16.0

Injected: phase_noise
Slope, off-line wander: −27.3°/sym, 32.4°
Rules: ~~cfo_not_corrected (wrong)~~
Trees: phase_noise 1.00

The rules call any slope past 20.7°/sym a CFO first, and a drifting random walk has one. Dashed: straight-line fit.

Measured on a separate campaign

Root cause named correctly by the trees: 99.2%
By hand-ordered rules on the same 17 features: 84.9%
Tests to a diagnosis, of a 48-test suite: 2.2
Unseen fault types flagged as new, ~5% false alarms: 89.5%

1,554 single-cause failures from 4,000 cases, seed 1. Trained on 8,000 cases, seed 0.

Eight root causes, each with a signature in the log

Root cause	Signature in the log	Rules	Trees
`low_snr`	Error matches the LTF noise estimate	75.1	99.8
`cfo_not_corrected`	Pilot phase ramps linearly	90.1	100
`stale_channel_estimate`	Error follows an old channel’s fades	90.1	99.6
`timing_late`	Impulse response wraps to negative delay	97.1	98.6
`pa_clipping`	Outer points compressed, gain below 1	74.0	96.6
`iq_imbalance`	Error tracks the mirror subcarrier	100	100
`phase_noise`	Pilot phase random-walks	51.4	98.1
`delay_spread_exceeds_cp`	Impulse-response energy past the CP	98.7	100

Recall, %. Ground truth is counterfactual: each failure is re-run on the same random streams with the fault removed.

A diagnosis in two tests, not 48

Run the test with the largest expected information gain, update the posterior over which bug, stop at 99%.

Policy	Tests, mean / p90	Diagnosis	Missed bug
Full suite	48 / 48	97.1%	3.3%
Random order	3.9 / 7	96.9%	3.7%
Information gain	2.2 / 3	98.3%	2.0%

350 fresh builds. The full suite scores lower: 48 correlated results make the posterior overconfident.

Where it breaks

40% of failures still fail with the fault removed: the SNR was too low as well. The model never says both.

Of those, given confidence of 0.9 or more: 91%
Caught by deferring the least-confident 10%: 23%
Clipping recall on 64QAM, held out of training: 58%

Run it

33 tests, including simulated BER against the closed-form Gray QAM curves.

Uncoded, single antenna, one packet; EVM is the only pass criterion.
Textbook impairment models, not measured hardware.

git clone https://github.com/akshit9162/ofdm-phy-triage
cd ofdm-phy-triage
python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
.venv/bin/python -m pytest -q
.venv/bin/python -m scripts.campaign 8000 runs/main 0
.venv/bin/python -m scripts.campaign 4000 runs/test 1
.venv/bin/python -m scripts.evaluate runs/main runs/test results