Find the defect. Prove why. Then act.

Traceon is a forensic agent for product-safety teams. It assembles scattered warning signs into one case, tests every suspect against factory data, exposes the ones that only look guilty, and quarantines the exact units at risk. Every claim it makes comes with a query you can re-run.

Open the case file →Watch a live investigation

About 30 seconds on 180,000 units with live Nemotron; no setup.

Fig. 1 — Seven suspects tested against 180,000 units. Two show no link; four look guilty only because they share units with the real cause. One survives every adjustment.

6 of 7suspects ruled out, each with its own test

92.8%of thermal incidents explained by one batch

$4.47Msaved against a blanket recall

3 claimsto statistical detection, Oct 10

250 / 250hidden scenarios solved blind

01 — Problem

The warning signs exist weeks before a recall. They sit in four places that never talk to each other.

Quality teams join spreadsheets by hand for weeks. When they finally act, they recall everything because they can’t prove which lot is bad, or they blame the wrong cause and miss it entirely.

SOURCE A

Warranty claimsPack hot near the discharge connector.125 claims. No pattern visible by eye.

SOURCE B

Reviews and forumsPin 4 has a black burn mark.Scattered across Reddit, Amazon and Facebook.

SOURCE C

Factory and supply chainWhich lots used cell batch 0917?Lots, BOMs, firmware, carriers, warehouses.

SOURCE D

Federal safety feedsSeven e-bike recalls since 2025.Live CPSC, openFDA and NHTSA data.

02 — Method

An investigation procedure, not a chatbot.

The model plans, challenges and explains. Every statistic is computed in code. Each step is recorded as an exhibit you can open.

Converge the signals

Meterless H-MEM

Stores every weak signal and pulls 131 of them into one thermal-connector cluster across six cities.

Plan the suspects

Nemotron 3.5

Ranks seven testable hypotheses from the evidence and logs the ones the data cannot test.

Test every suspect

Zetaris SQL

A federated query and a significance test for each, across ERP, supplier, telemetry and warehouse data.

Unmask confounders

Mantel–Haenszel

Stratified tests show which suspects stop mattering once the real cause is held fixed.

Inspect the photograph

Nemotron VL

Reads a real forensic photo of a lithium-ion pack after thermal runaway and locates the damage.

Attack the verdict

Nemotron red team

Raises objections. Each is answered by a deterministic test, or logged as an open question.

Pass the evidence gate

Seven criteria

The quarantine control does not appear until every criterion holds.

Act with a named approver

Human in the loop

Places an ERP hold on 1,380 serials, drafts the notices, and remembers the finding.

03 — Explain why

Four suspects looked guilty. Only one was.

On raw numbers, heat, the plant, the BMS board and the cell supplier all correlate with failures. A correlation tool would blame them. Traceon asks whether each effect survives once the leading cause is held fixed.

Raw data: five suspects look guilty.

**Table 1** — Raw chi-square tests, then Cochran–Mantel–Haenszel with the cell batch held fixed. The batch fails at 841 per 10k against 0.5 per 10k elsewhere (relative risk 1,668×) and holds in every firmware, carrier, climate, BMS and warehouse subgroup.
Suspect	Raw test	Verdict
Firmware v3.4	p = 0.44	Ruled out
Carrier shock	p = 0.33	Ruled out
Sun Belt heat	p = 1.6e-6	Looks guilty
Plant 2	p = 3.7e-42	Looks guilty
BMS board	p = 5.1e-25	Looks guilty
Supplier B	p = 3.7e-42	Looks guilty
Cell batch CB-B-0917	p < 1e-300	Looks guilty

Phoenix and Austin really do see more failures, because that is where the bad batch was shipped. Within the batch, and outside it, Sun Belt units fail at the same rate as everywhere else.Nemotron’s own first guess was heat. The evidence overruled it, and the agent trace records that.

04 — Blind tests

Same agent, different hidden causes. It finds each one, and refuses to act when there is none.

The obvious question is whether the answer was planted. So the agent runs blind on scenarios it has never seen: a firmware bug, a bad assembly plant, transit damage, faulty BMS boards, two causes at once, or no defect at all. The planted truth is revealed only after the verdict.

**Table 2** — npm run eval. Correct means exactly the planted causes found, and action allowed only when a cause exists.
Benchmark	Result
Held-out random scenarios (seeds 151–400)	250 / 250
All scenarios (seeds 1–300 and four named cases)	304 / 304
Two simultaneous causes	both found
No defect present	action blocked
Runs that acted on a wrong cause	0

Cell batch46/46

Assembly plant47/47

Firmware release36/36

BMS board supplier44/44

Carrier41/41

Two causes at once44/44

No defect46/46

Try it yourself. Each run draws a new hidden cause from a random seed.

Run a hidden scenario →Watch it refuse to recall

05 — Audit

An agent you can audit, not just believe.

Safety decisions are challenged by regulators, suppliers and lawyers. Every answer Traceon gives is built to survive that.

The model never computes a statistic: Every p-value, rate and dollar figure is computed in code and checked against scipy and statsmodels. Nemotron plans, challenges and explains.
Every figure in the prose is checked: Numbers in the written explanation are matched against the evidence. Anything unverified is flagged.
The agent can block itself: No quarantine unless significance, confounders, subgroups, residuals, failure signature, red team and scope all pass.
Every query is replayable: Evidence is hashed into a SHA-256 chain. One control re-runs all 45 queries and verifies each hash and the root.
Detection is dated: A day-by-day replay shows the defect became detectable on Oct 10 after three claims; 113 later incidents were preventable.
Findings are remembered: An approved finding becomes long-term H-MEM memory with lineage to its signals. The next investigation recalls it.

— Disposition

Quarantine 1,380 units, not 180,000.

The cause is pinned to one cell batch and mapped to exact serials, SN-88400 to SN-89779. Traceon places the hold across six warehouses and drafts a stop-sale notice and a CPSC Section 15(b) report for counsel.

Blanket recall$4.50M180,000 units

Surgical quarantine$34,5001,380 units

Saved$4,465,500

Beats the multiple-testing barBonferroni
Survives adjustment for every other suspectCMH ×4
Consistent in every subgroup5 / 5
Nothing unexplained remains92.8% explained
Consistent failure signaturephoto + 85/116
Red-team objections answeredNemotron
Scope mapped to exact serialsL-409

EVIDENCE GATE · ACTION ALLOWED

— Beyond one case

Before the case, after the verdict, and on real recalls.

The same tests watch the warranty stream every day, run on your own data, are checked against public recall history, and keep checking after the fix.

Would the signal have fired first?: Public NHTSA complaint histories replayed day by day through a standard safety-signal test. The Chevrolet Bolt EV fire signal fires 14 days before recall 20V701000; on 3 of 5 vehicles it comes first.
Screen every day, open a case on alert: The warranty stream is screened daily across every factor. The day it fires, a case opens on the data as it stood that day.
Bring a CSV: One row per unit, a failure column and your suspects. Same agent, same tests, same evidence gate. A 12,000-unit sample is included.
Ask, test and challenge the verdict: Ask questions answered only from the exhibits, move what-if sliders that re-run the statistics, see every interval, and read what would prove the verdict wrong.
Verify, or reopen: Thirty simulated days after the fix are checked against the pre-registered falsifiers. Act on a confounder instead and the case reopens.
Reports, review and a timestamp: An 8D report and a supplier claim that cite their exhibits, a read-only case link for reviewers to comment on, and a signed root anchored with OpenTimestamps.

06 — Stack

Each component does real work, and says when it is not live.

The app labels every integration as live, simulated or offline, so nobody has to guess what they are looking at.

Component	Role in the investigation	Status
NVIDIA Nemotron 3.5	Super VL inspects the photo and writes the explanation with thinking on; Lightning plans and red-teams.	live
Zetaris	Federated SQL across ERP, supplier, telemetry, warehouse and warranty silos, with a hash-parity check.	SQL API · mirror fallback
Meterless H-MEM, Markovian	Durable signal memory with a trust ledger; a bounded investigation chain.	live, local-first
Bright Data	Reviews and forum posts, where early complaints surface first.	live with token
CPSC, openFDA, NHTSA	Base rates for e-bike recalls and lithium-battery incidents.	live
Factory and warranty data	A fictional company, VoltRide, with planted defects. The agent is never told the answer.	simulated

Open the case. Try to break the verdict.

Select any node to see the SQL behind it. Read the red team’s objections. Then approve the quarantine yourself.

Open the case file →Watch a live investigation