TETRA End to End, Part 14: Testing TETRA Without a Network — & What's Still Open

By Matt Cheramie · August 31, 2026

Part 14 — the finale — of TETRA End to End. Thirteen posts ago we started with a π/4-DQPSK constellation on a 25 kHz carrier and a promise: every claim pinned by a test or a capture. Since then we’ve built the slot grid, the channel coding, the scrambler, TCH/S, a clean-room ACELP vocoder proven bit-identical against the ETSI reference, soft decisions, two equalizers, a hardened control channel, and a DMO path that went from “reads the spec” to a production pipeline — pausing once to retract a wrong verdict in public. This close-out is about the machinery that made that pace survivable: the test lattice that lets a TETRA stack be developed hundreds of kilometers from the nearest TETRA network — and, in the same breath, the honest list of what that lattice cannot prove and what therefore remains open.

TL;DR: TETRA in GopherTrunk is tested in four layers, each catching what the previous one can’t: synthetic round-trips (fast, but blind to self-consistent bugs unless the encode side models the air — the colour-0 and slot-grid lessons); skip-guarded capture harnesses (TestTETRAMultiSlotReplay, TestTETRADMOReplay, TestTETRADMOColourScan — activated by GT_TETRA_DMO_IQ, GT_TETRA_DMO_RATE, GT_TETRA_DMO_COLOUR, GT_TETRA_DMO_CLEAR, GT_TETRA_DMO_SCAN, GT_TETRA_LMS, GT_TETRA_DMO_LMS); pipeline tests on modulated IQ (pipelines_dmo_test.go, the CC equalizer A/B); and full-daemon integration (TestDaemonDecodesTETRADMO, under make integration). Still open, stated plainly: the #1003 on-air A/B through the full daemon; DMO call-control (recordings file under group 0); and the 15aug lesson — a colour scan with no dominant winner means a marginal signal, not a broken guesser, and the descramble must never be bent to fit a 33%-yield colour.

Key takeaways

• Each test layer exists because the one below it lied at least once. Round-trips passed through the colour-0 bug; only an encode side that scrambles unconditionally — modeling the transmitter, not the code under test — made it fail first.
• Skip-guarded harnesses turn operators into the lab. A test that skips without GT_TETRA_DMO_IQ costs CI nothing, yet gives anyone with a capture a one-command, fully instrumented replay — every major TETRA finding in this series arrived through one.
• A refusal can be a correct answer. RecoverDMColourCode declining to pick a colour on the 15aug capture wasn’t a bug — the scan’s shape (several modest risers, no 3× dominance) is physically impossible for one radio with one keystream, so the gate correctly diagnosed a marginal signal.
• Green everything ≠ done. The DMO daemon path passes synthetic, capture, pipeline, and integration layers — and #1003 stays open anyway, because the one test that counts is a recording an operator can listen to.

Cheat sheet

Layer	What it proves	Where it lives
Vocoder conformance	bit-identical PCM vs the ETSI reference	internal/voice/acelp/etsi_reference_test.go
Synthetic round-trips	coding chains invert; encode models the air	internal/radio/tetra/dmo_decode_test.go, tch_test.go
TMO capture harness	grant-correlated multislot voice, LMS A/B	cmd/gophertrunk/tetra_multislot_replay_test.go (GT_TETRA_LMS)
DMO capture harness	lock, colour, TCH/S yield, verdict line	cmd/gophertrunk/tetra_dmo_replay_test.go (TestTETRADMOReplay)
Colour diagnostics	full 64-colour yield map	tetra_dmo_replay_test.go (TestTETRADMOColourScan, GT_TETRA_DMO_SCAN=1)
Pipeline on modulated IQ	lock + colour + grant from synthesized bursts	internal/scanner/ccdecoder/pipelines_dmo_test.go
CC equalizer proof	12%→100% BSCH on a real marginal fixture	pipelines_tetra_equalizer_test.go
Full daemon	config → DDC → pipeline → lock, no shortcuts	cmd/gophertrunk/integration_cc_tetra_dmo_test.go (make integration)

In this post

• The lattice, bottom to top — four layers and the failure each one owns.
• Making synthetics honest — the two fixture repairs that mattered.
• The capture harness as an instrument — env vars, verdict lines, A/B knobs.
• The 15aug capture: reading a refusal — when no colour dominates.
• The open list — what stays unverified, and the gate for each item.

The lattice, bottom to top

The bottom layer is conformance and round-trips: etsi_reference_test.go feeds the same 137-bit stream to GopherTrunk’s ACELP and the ETSI reference codec and demands bit-identical PCM (Part 7); encode/decode pairs pin every coding chain. Fast, deterministic, runs on every make vet test. Its blind spot is the one this series hit twice: a bug on both sides of a round-trip is invisible.

The second layer is skip-guarded capture harnesses — ordinary go test functions that skip unless an environment variable points at real IQ. The third is pipeline tests on modulated IQ: synthesize actual bursts, modulate them, and drive the production pipeline object — TestTETRADMOPipelineLocksAndGrants asserts lock, colour-3 recovery, and a grant from IQ in; TestTETRACCEqualizerLiftsMarginalBSCHYield holds the live TMO pipeline to the healthy regime on a real marginal fixture (Part 10). The top layer is full-daemon integration under make integration: TestDaemonDecodesTETRADMO boots the daemon from a protocol: tetra-dmo config and requires the whole path — DDC, pipeline, bus — to lock, with no hand-wired shortcuts. Above all four sits the layer no test framework provides: an operator, a radio, and the air.

The four automated layers of the TETRA test lattice, and the fifth — on-air verification — that no green run below it can substitute for.

Making synthetics honest

Two fixture repairs from the DMO arc are worth restating as method, because they generalize past TETRA. First: the encode side must model the transmitter, not the code under test. The colour-0 descramble skip (Part 12) survived a targeted sweep because the synthetic encoder shared the decoder’s conditional; dmo_decode_test.go now scrambles unconditionally — real-air behaviour — and the colour-0 rounds became the failing-first regression. Second: synthetic timing must be as structured as real timing. The noise-grant bug (Part 13) was uncatchable while buildDMODibitStream laid bursts at arbitrary filler offsets; a fixture transmitter that keeps the 255-dibit slot grid is what lets the DMSlotGrid regressions mean anything. The same honesty shows up smaller: blind CMA is degenerate on a noise-free constant-modulus input, so the daemon’s synthetic TETRA test adds 40 dB AWGN on purpose. A fixture flattered is a bug deferred — the full pattern catalog is in From the Issue Tracker #20.

The capture harness as an instrument

The harness layer deserves its manual, because it is how operators and the project collaborate. Everything keys off environment variables, so the tests cost CI nothing and cost an operator one shell line:

// cmd/gophertrunk/tetra_dmo_replay_test.go (shape) — the skip guard
path := os.Getenv("GT_TETRA_DMO_IQ")
if path == "" {
    t.Skip("set GT_TETRA_DMO_IQ (cs16 IQ) + GT_TETRA_DMO_RATE to run the DMO replay")
}

The knobs, as of this series’ close: GT_TETRA_DMO_IQ / GT_TETRA_DMO_RATE select the capture; GT_TETRA_DMO_COLOUR pins the DM colour (unset ⇒ auto-recovery); GT_TETRA_DMO_CLEAR=1 asserts the capture is known-clear, flipping the VERDICT line so a chance floor reads as a decode defect, never encryption; GT_TETRA_DMO_SCAN=1 runs TestTETRADMOColourScan, the full 64-colour yield map; GT_TETRA_DMO_LMS=1 A/Bs the per-burst trained equalizer on DMO, as GT_TETRA_LMS=1 does for TMO in TestTETRAMultiSlotReplay (compare traffic_marked_crc_soft across runs). The reproduction line for the Part 12 result is one command: GT_TETRA_DMO_IQ=<10aug .raw> GT_TETRA_DMO_RATE=144000 GT_TETRA_DMO_CLEAR=1 go test ./cmd/gophertrunk -run TestTETRADMOReplay -v. Every headline number in Parts 9–13 — 410→778, 12%→100%, 6→64, 1/269→35/269 — is re-derivable this way from the operator’s own files. That reproducibility is the series’ method.

The 15aug capture: reading a refusal

The most recent capture is the best closing exhibit, because its result is a correct refusal. The operator recorded a purpose-built vector — 25 s of clear, colour-0, silent PTT (dmo_test_15aug/25sec_ptt_then_off_30sec_cs16_144khz.raw). Signalling: DSB SCH/S at ~90% (105/117). Voice: near the chance floor at every colour. The colour scan shows several colours rising modestly at once — 28→140, 57→74, 30→46 of 831 DNBs — rather than one dominating. One radio scrambling with one keystream cannot produce that shape, so RecoverDMColourCode’s dominance gate refuses to pick (140 < 3×74) — which the operator experienced as “the colour guesser is broken” and which is actually the gate working: a marginal signal with partial-keystream artifacts, not a recoverable colour. The receiver’s side of the ledger is already maxed — the CMA equalizer lifts DSB 77→105 and TCH 80→140; LMS doesn’t move it — and the SYNC PDU’s extended-colour parse matches osmo-tetra-dmo’s scrambler-init offsets, so that path isn’t the gap. The conclusions are procedural: a silent PTT is a poor test vector (DTX/comfort-noise frames), the next capture must be known-colour and actually talking — and above all, do not bend the descramble to fit a 33%-yield non-dominant colour. That would be the self-consistent trap, volume three.

The open list

What this series leaves open, each with its closing gate:

• The #1003 on-air A/B. A clear DMO capture through the full daemon (protocol: tetra-dmo) must land an intelligible recording. Known sharp edges for whoever runs it: recordings file under group 0 (no talkgroup on the air), colour recovery needs ~20 qualified DNBs (a very short first PTT grants before the colour is known — the voice chain re-recovers), and the grant lands ~0.5 s into a transmission by design. dnb_qualified is the traffic number; dnb_total is a noise meter.
• DMO call control. EN 300 396-3’s source/destination/group PDUs are not yet decoded — attribution and group filing wait on it.
• The trained LMS as a production default. Staged on both TMO and DMO soft paths, pinned synthetically (13%→0% and 12%→0% payload error), gated on capture A/Bs that so far say “no harm, no clear win.”
• A decisive DMO voice capture. The 15aug refusal stands until a known-colour, talking capture either recovers cleanly or fails informatively.

That is the honest ledger. The arc it closes is real, though: one carrier, followed from a rotating constellation to a conformant vocoder to recorded clear voice — TMO hardened end to end, DMO taken from spec-reading to a production pipeline — with every claim along the way carrying either a test, a capture, or an explicit “not yet verified.” The villain of the series never changed: the green test that wasn’t evidence. The method never changed either: make it fail first, then make it pass, then make the air agree.

FAQ

Can I develop against this stack with no TETRA network in range? Yes — that’s the lattice’s whole point. The bottom three layers need no RF at all, and the capture harnesses run on files. What you cannot do without air is close anything: on-air verification is a distinct layer, not a formality.

Why environment variables instead of test flags or fixtures in the repo? Captures are large, operator-owned, and often sensitive, so they can’t live in testdata. Env-var skip guards keep the tests permanently wired into go test (no bit-rot) while making CI cost zero and operator activation trivial.

How do I contribute a capture that actually helps? Record cs16 IQ at a known rate, note everything the radio’s config declares (colour, security class, frequency), and prefer a talking transmission — the 15aug lesson is that silent PTT vectors mostly exercise DTX frames. Then run the relevant harness and file the full log output; the verdict lines are written to be pasted into issues.

What would make the colour-scan gate pick wrongly? Very little, by construction — it needs 6 CRC-valid bursts and 3× dominance, and wrong colours live at the ~1/256 floor. The realistic failure is the one observed: a signal too marginal for any colour to dominate, where the gate refuses. Trusting a non-dominant winner is exactly what the gate exists to prevent.

Is TMO TETRA “done”? The decode path is verified end to end — conformant vocoder, capture-proven soft/equalized TCH/S, a control channel that rides through the marginal regime. Open engineering remains around the edges (LMS default, symbol-scope pooling), but nothing in the TMO voice path currently waits on evidence.

Series navigation

Part 14 of 14 · ← Part 13: DMO III — A Production Pipeline: Grid Votes, Grants & Noise · This is the finale — back to the series index.

Where to next? The equalizers, soft decisions, and metric discipline this series applied to one protocol are a general toolkit — see how they generalize in Weak-Signal Engineering, which owns the theory these TETRA posts kept pointing at.