Each of these is a real Truveil report on a real agent run. Nothing was selected for its grade, and nothing was edited afterwards. Two of the eight audit live workflows rather than test data, and both of the two lowest-scoring reports audit agents we built ourselves.
These reports are real Truveil output: real agents, real runs, real hash-chained logs, real engine determinations. Where a scenario involves people or organisations, they are fictional; no real individual's data appears in any sample. Seven were scored by engine v1.4.0 against methodology bible-1.9 and the eighth by engine v2.0.0 against bible-2.0, the same path that produces a customer's report. Every finding below is the engine's determination, not a model's opinion.
India · Email triage
Email Triage Agent
Highest stake in the set: a real, irreversible action.
A81.2 / 100
Transparency83.3
Accountability89.5
Data trust70
Reversibility76.9
6 logged events · Truveil methodology, DPDP Rules 2025 · Live mail workflow, Truveil's own mailbox
What the agent did
Fetched a live mail thread and calendar context, classified the message as high-stakes, drafted a reply, and dispatched it only after a human approved through a messaging checkpoint.
What the report shows
One of two samples audited on a live workflow rather than test data. Graded A, and it still carries two HIGH findings. Grade A is not a clean bill.
Captured a hiring requisition, parsed a resume and reviewed its scoring rubric for protected attributes before scoring, ran three panel interviews and a debrief, and routed the offer through HR and finance approval before emailing it to the candidate.
What the report shows
The hire decision itself put the agent in Category 1: the category came from the record, not a declaration. Accountability scores 87.5 while Data Trust and Reversibility sit at 25 and 23.1, and 10 of 21 applicable signals were evidenced. The report withholds any statement it cannot verify against the run's record.
Multi-agent environment: one accountability trail, whichever model runs.
C62.8 / 100
Transparency91.7
Accountability64.3
Data trust55.6
Reversibility46.7
16 logged events · Truveil methodology, DPDP Rules 2025 · Live production run
What the agent did
Ran a research and suggestion cycle across a pipeline using separate models for research, analysis and critique, persisted briefs for human review, and recorded user interaction with the queue.
What the report shows
The lowest grade in the set. Reversibility scores 46.7 with no override control, no appeal channel and no correction workflow. Step 05 states plainly that no formal approval event was logged at any stage.
Profiled six candidates for a senior engineering role, ranked them on seniority, stack fit and location, raised one approval request per candidate, and issued advance, conditional-advance and do-not-advance verdicts.
What the report shows
Article 50 renders as binding while Articles 9, 10(3), 14(4)(e) and 15 render with their December 2027 effective dates. Two kinds of legal claim, distinguishable on the page. The finding: candidates were never told an AI was assessing them.
9 logged events · Singapore MGF-GenAI · Synthetic data
What the agent did
Ingested three vendor quotations, logged a conditional validation pass with four recorded exceptions, logged an individual assessment for each of the three vendors with its uncertainties and open items recorded, and recorded an award against explicit human approval.
What the report shows
Every Singapore citation reads “advisory, not binding”, because Singapore's Model AI Governance Framework is guidance rather than statute. The agent declined its own automatic approval gate and waited for a person.
10 logged events · NIST AI RMF, NIST AI 600-1 · Synthetic data
What the agent did
Ranked ten accounts for churn risk, proposed three retention discounts, surfaced its contested ranking, and recorded the operator substituting one account before approving the final set.
What the report shows
The override is the point. The change itself, not just the outcome, is in the record. NIST citations read “voluntary standard” throughout.
Applied an opt-in rule to twelve customer records, logged each inclusion and exclusion as its own decision with the consent basis stated, and finalised a send list only after operator approval. Nothing was sent.
What the report shows
The second-densest decision trail in the set, and the highest-severity finding in an otherwise strong run: no risk assessment on file.
Screened five candidate profiles against three stated criteria, flagged ambiguity on the profiles it could not resolve rather than rejecting them, ran a cohort fairness check, and requested human sign-off before any shortlist.
What the report shows
The category was not declared, it was derived. Profiling of natural persons reclassified this agent to Category 1, raising obligations that bind on 13 May 2027 and are not yet met.