Datasetcredit-nov

credit-nov5 findings traces-nov5 findings

Ava L.

rootplant-credit-nov-20261104.zip is imported. The file is a bundle, so this dataset has Bundle Analysis beside Traces & Spans.

credit-nov

This page shows what the dataset contains and what Auga found in it.

Source: rootplant-credit-nov-20261104.zip 1.9 MB
sha256: c25aafc5fb48…
Received: 4 Nov 2026 14:44 UTC
Window: 06:00 – 17:59 UTC (12.0 h)
Period: 900 s
Analysis run: otlp-c25aafc5fb48
State: Active

Upgrade

Bundle Analysis is where you practise with Auga and adjust it. Each bundle is a simulated dataset with a known correct answer and a root cause hidden inside it. Run a bundle and see whether Auga finds the root cause. Read the documentation when you want to learn more.

Findings

3 grade A · 1 grade B · 1 grade C

Traces

463

399 flow · 64 sync

Signals after screening

8of 14

6 set aside

The answer

Sealed

until you open it

Root causeThe answerSealedOpenRead what Auga found below, and decide which cause you would choose. Then open the answer and compare.This shows what was hidden in the data, and what Auga found.Show the answerHide the answer

Root cause

Reached

at the first step that read the data, not at the sync step

Planted links found

3of 4

Auga missed the sync step

False causes accepted

1of 2

sanctions_list_sync, at grade C

Affected runs

68of 463

1 extra finding, harmless

Root cause

This comes from ground-truth.json. Only admins and operators can see it. It is never stored in a run, a finding or the ledger.

bureau_feed_sync wrote an empty batch (0 records). Every bureau_fetch after it returned a stale score.

4 Nov 2026 11:15 UTC · signal: attributes · target outcome: credit.decision=deny

Hop	Span	Mechanism	Why
1	bureau_fetch execute_tool	state	bureau_fetch read the empty feed (0 records) and returned a stale score
2	risk_agent invoke_agent	data	the stale score raised the risk score
3	policy_agent invoke_agent	data	the risk score crossed the deny threshold

Did Auga find it? Yes, at the first step that read the data. The cause chain starts at credit.feed_batch_records, because the sync step itself is not a signal.

False causes planted One is sanctions_list_sync, a new list version two hours before the fault. The other is policy_config_deploy, a deployment one hour after the fault, when the denials had already started.

Found and planted

What Auga found, beside the correct answer. Auga found 3 of the 4 planted links.

Cause	Effect	Grade	Lag	Strength	In words	Planted	Result
credit.feed_batch_records	credit.score_source	A	1	0.93	An empty feed batch makes the bureau score stale one period later. We are very sure: 3 of 3 challenge tests held.Prediction If the feed batch is empty again, stale scores follow within one period. · recheck 11 Nov 2026	yes	match
credit.score_source	credit.risk_score	A	0	0.91	A stale bureau score raises the risk score in the same period. We are very sure: the effect held in 6 other environments.Prediction A stale score will raise the risk score by about 12 points. · recheck 11 Nov 2026	yes	match
credit.risk_score	credit.decision	A	0	0.96	A higher risk score makes a denial more likely in the same period. We are very sure: 4 of 4 registered predictions held.Prediction Above a risk score of 62, denials increase. The next 30 decisions will show it. · recheck 8 Nov 2026	yes	match
credit.income_verified_ratio	credit.affordability_ratio	B	0	0.78	We are fairly sure that a better-verified income is followed by a higher affordability ratio. One of three challenge tests did not hold.no registered prediction · recheck 18 Nov 2026	no	extra · harmless
credit.sanctions_list_version	credit.decision	C	2	0.41	Auga has no plain wording for this link yet (coincident). Ask for a wording review. · 422 ·	no	false cause accepted
bureau_feed_sync (sync span)	credit.feed_batch_records	—	—	—	—	yes	missed A sync step is not a signal. The link starts at the first step that reads it.

Results are calculated against ground-truth.json by the benchmark scorer. A false cause that Auga accepted is red. An extra link that is true but not planted is harmless, and has the normal colour.

Traffic over time

Runs per 15 minutes. The planted fault and the false causes are marked.

runs per 15 minplanted faultdecoy

Who calls whom

The orchestrator, the agents in call order, and the tools each one uses. The planted cause chain is red.

Traces

463 of 463. Open shows each step of a run in time order. A click on a dot, bar or link above brings you here.

affected only

Time	Root span	Spans	Duration	Status	In the scenario
06:00Z	Bureau agent invoke agent	2	2,107 ms	ok	healthy	Open
06:02Z	Loan orchestrator invoke agent	12	5,832 ms	error	healthy	Open
06:04Z	Loan orchestrator invoke agent	12	3,248 ms	error	healthy	Open
06:05Z	Loan orchestrator invoke agent	10	3,945 ms	ok	healthy	Open
06:06Z	Loan orchestrator invoke agent	11	4,726 ms	ok	healthy	Open
06:07Z	Loan orchestrator invoke agent	10	4,219 ms	ok	healthy	Open
06:09Z	Loan orchestrator invoke agent	10	4,846 ms	ok	healthy	Open
06:12Z	Loan orchestrator invoke agent	12	3,693 ms	ok	healthy	Open
06:13Z	Loan orchestrator invoke agent	10	4,425 ms	ok	healthy	Open
06:15Z	Bureau agent invoke agent	2	2,063 ms	ok	healthy	Open
06:15Z	KYC agent invoke agent	2	1,289 ms	ok	healthy	Open
06:15Z	Loan orchestrator invoke agent	12	4,493 ms	ok	healthy	Open

At most 50 of 463 are listed. Use the search to see the rest.

Scenario

From scenario.yaml: the settings used for this bundle.

template	credit (loan application, 8 agents)
seed	20261104
fault.type	poisoned_data
fault.origin	tool:bureau_feed_sync
fault.at	+05:15
fault.blast_radius	30%
causal_distance	4 hops
noise.decoys	2
spans_per_trace	10–12 (auto)
annotation_mode	blind
auga_profile	true

Links found

What Auga found in run otlp-c25aafc5fb48. Grade A means identified, and grade C means a guess.

Cause	Effect	Grade	Lag	Strength	In words
credit.feed_batch_records	credit.score_source	A	1	0.93	An empty feed batch makes the bureau score stale one period later. We are very sure: 3 of 3 challenge tests held.Prediction If the feed batch is empty again, stale scores follow within one period. · recheck 11 Nov 2026
credit.score_source	credit.risk_score	A	0	0.91	A stale bureau score raises the risk score in the same period. We are very sure: the effect held in 6 other environments.Prediction A stale score will raise the risk score by about 12 points. · recheck 11 Nov 2026
credit.risk_score	credit.decision	A	0	0.96	A higher risk score makes a denial more likely in the same period. We are very sure: 4 of 4 registered predictions held.Prediction Above a risk score of 62, denials increase. The next 30 decisions will show it. · recheck 8 Nov 2026
credit.income_verified_ratio	credit.affordability_ratio	B	0	0.78	We are fairly sure that a better-verified income is followed by a higher affordability ratio. One of three challenge tests did not hold.no registered prediction · recheck 18 Nov 2026
credit.sanctions_list_version	credit.decision	C	2	0.41	Auga has no plain wording for this link yet (coincident). Ask for a wording review. · 422 ·

Traffic over time

Runs per 15 minutes. The marks show where Auga found a link to one period later.

runs per 15 minAuga's finding

Who calls whom

The orchestrator, the agents in call order, and the tools each one uses. Dashed lines are the grade A links that Auga found.

Traces

463 of 463. Open shows each step of a run in time order. A click on a dot, bar or link above brings you here.

Time	Root span	Spans	Duration	Status
06:00Z	Bureau agent invoke agent	2	2,107 ms	ok	Open
06:02Z	Loan orchestrator invoke agent	12	5,832 ms	error	Open
06:04Z	Loan orchestrator invoke agent	12	3,248 ms	error	Open
06:05Z	Loan orchestrator invoke agent	10	3,945 ms	ok	Open
06:06Z	Loan orchestrator invoke agent	11	4,726 ms	ok	Open
06:07Z	Loan orchestrator invoke agent	10	4,219 ms	ok	Open
06:09Z	Loan orchestrator invoke agent	10	4,846 ms	ok	Open
06:12Z	Loan orchestrator invoke agent	12	3,693 ms	ok	Open
06:13Z	Loan orchestrator invoke agent	10	4,425 ms	ok	Open
06:15Z	Bureau agent invoke agent	2	2,063 ms	ok	Open
06:15Z	KYC agent invoke agent	2	1,289 ms	ok	Open
06:15Z	Loan orchestrator invoke agent	12	4,493 ms	ok	Open

At most 50 of 463 are listed. Use the search to see the rest.

The bundle

7 files · manifest.json

File	Lines	Format
traces.otlp.json	465	otlp-trace-jsonl-v1
ground-truth.json	2,307	rootplant-ground-truth-v0.1
scenario.yaml	41	rootplant-scenario-v1
edge.yaml	95	auga-edge-v1
actions.yaml	42	rootplant-actions-v1
logs.otlp.json	465	otlp-logs-jsonl-v1
manifest.json	61	rootplant-manifest-v1

Validation

Rootplant's validation results, read from the bundle, and Auga's blind check.

L1 · structure	ok	Every step has its parent, no links are broken, and the timestamps are in order.
L2 · content	ok	Blind: no signal or event names the fault. 465 flow and sync runs, minimum 60.
L2-006 · size	ok	2.0 MB uncompressed, under 50 MB
Blind (Auga)	ok	No value contains root_cause, planted_fault or decoy.

Example data: this page generated a Rootplant-style credit dataset from seed 20261104 (463 runs, 4,619 spans). Auga's figures for screening and findings are example values until that part of the product is built. This is a static example. The search fields do not work, and every Open link leads to the same example run.