Sample deliverable
What a S.AI.F assessment report looks like.
Every engagement ends in a report like this one: seven pages a board can act on — the verdict, the scorecard behind it, the one finding that matters, the evidence, and the decision it asks for.
SPECIMEN · FICTIONAL CLIENT & FIGURES
01 · Cover
Gulf National Bank
بنك الخليج الوطني
Assessment results · board summary
credit-scoring-v5
Retail credit decisioning model · Gulf National Bank · Credit & lending · decides with human review
CERTIFIED1 RESIDUAL RISK ACCEPTEDHIGH-RISK · ANNEX III
PREPARED FOR
Board Risk Committee
DATE
10 Sep 2026
TRUST PROOF
TP-0413
DECISION DUE
23 Sep 2026
S.AI.FASSURANCE
CERTIFIED · TP-0413
Independently tested and signed. Valid until 10 Mar 2027 or the next model release, whichever comes first.
signed ML-DSA-65 · 2026-09-10 09:14Z
verify.saifcertify.com/TP-0413
verify.saifcertify.com/TP-0413
02 · Executive summary
Gulf National Bank
بنك الخليج الوطني
Assessment report · credit-scoring-v5 · TP-0413 · 10 Sep 2026 · Confidential
Executive summary
Safe to deploy under two conditions.
one decision needed · 23 Sep 2026
Verdict
Certified
Independently tested and signed · TP-0413
Suites passed
9 / 9
all above threshold · lowest score 89
Findings
1 open
fairness · age 60+ · inside limit, accepted with a fix
Security
0 breaches
4,794 probes · supply chain attested
01
What was tested?
The retail credit model that decides applications with human review. Sealed and tested by S.AI.F — not by the team that built it.
02
What did we find?
Accurate, robust and secure. One fairness gap: applicants over 60 are approved 4.4 points less often — inside the ±5 limit, but drifting.
03
What does it cost us if we wait?
v4 stays live with a wider gap and no signed proof, and the CBUAE model-risk return keeps carrying an unverified system.
RECOMMENDATION
Approve for production, with two conditions.
1Threshold tuning by age group, re-tested by S.AI.F · 30 Nov 2026
2Monthly fairness check on live decisions · from go-live
The proof revokes itself if either condition lapses.
DECIDED BYBoard Risk Committee
DECISION DATE23 Sep 2026
EVIDENCEpages 4–7
Assessed and signed by S.AI.F · every figure links to a run id · verify.saifcertify.com/TP-04132 / 7
03 · Management summary
Gulf National Bank
بنك الخليج الوطني
Assessment report · credit-scoring-v5 · TP-0413 · 10 Sep 2026 · Confidential
Management summary
Management recommends approval, with conditions.
board-level view
Overall position
Certified, with one accepted risk.
9/9
independent
test suites
passed
test suites
passed
Tested and signed by a party independent of the build team and the bank. Nothing was found that should stop deployment.
What the assessment tells us
The model does what it is meant to do.
It holds up on the bank's own data and on cases it never saw, and it explains itself well enough for a credit officer to defend a decision.
One fairness matter needs active management.
Older applicants fare slightly worse than comparable younger ones — inside the tolerance the bank set, but moving toward it rather than away.
Security and controls held under attack.
The model was deliberately attacked and did not give way. Its data, weights and hardware are accounted for end to end.
Recommendations to the Board
1
Approve, with conditions.
DECISION2
Make fairness monitoring continuous.
3
Apply this standard estate-wide.
BOARD DECIDESRecommendation 1
BY23 Sep 2026
Assessed and signed by S.AI.F · the detail behind every statement follows · verify.saifcertify.com/TP-04133 / 7
04 · Verdict
Gulf National Bank
بنك الخليج الوطني
Assessment report · credit-scoring-v5 · TP-0413 · 10 Sep 2026 · Confidential
Verdict
Certified
with one residual risk accepted and two conditions attached to production use.
TRUST PROOFTP-0413
RISK TIERHigh-risk · Annex III
VALID TO10 Mar 2027 · or release
APPROVED BYCertification Committee
SIGNEDML-DSA-65 · HSM
Scorecard · nine suites, all above threshold
score · threshold
Task accuracy
HELM · set v7
96
≥ 90
Grounding
Ragas
97
≥ 95
Drift
Evidently · PSI
95
≥ 90
Calibration
ECE 0.021
94
≥ 90
Explainability
SHAP · DiCE
93
≥ 85
Bias & fairness
Fairlearn · EOD
92
≥ 90
Robustness
TextAttack
91
≥ 85
Fraud resistance
Core · ART
90
≥ 85
Latency under load
p95 412 ms
89
≥ 85
Scope & tier
Annex III · pass
Artefacts
12 sealed · pass
Data gates
212 expectations · pass
Evaluation
9 of 9 suites · pass
Fairness
F-01 · age 60+ · accepted
Security
0 breaches · pass
Assessed and signed by S.AI.F · every figure links to a run id · verify.saifcertify.com/TP-04134 / 7
05 · Findings
Gulf National Bank
بنك الخليج الوطني
Assessment report · credit-scoring-v5 · TP-0413 · 10 Sep 2026 · Confidential
Findings
One finding stands between v5 and unconditional approval.
F-01 · Fairness · Age 60+
MEDIUM · ACCEPTEDApplicants over 60 are approved 4.4 points less often than equally creditworthy younger applicants. Inside the ±0.05 threshold, but the closest of any group and drifting wider over three releases.
Equal opportunity difference by group
threshold ±0.05 (dashed) · held-out set
Female
+0.012
Male
-0.009
Age 18–25
+0.031
Age 60+
-0.044
Non-national
+0.027
Remediation plan
Threshold tuning by age group
Model Risk
30 Nov 2026
Re-test on frozen set v7
S.AI.F
pre-release
Monthly live fairness check
Core Sentinel
from go-live
Residual risk acceptance
Accepted by the Head of Model Risk, 09 Sep 2026, with the plan above. Expires with the proof.
Closed during assessment
F-02 · Serving-image CVEs
patched · Grype clean
F-03 · Label noise 1.8%
relabelled · re-run pass
Assessed and signed by S.AI.F · every figure links to a run id · verify.saifcertify.com/TP-04135 / 7
06 · Evidence
Gulf National Bank
بنك الخليج الوطني
Assessment report · credit-scoring-v5 · TP-0413 · 10 Sep 2026 · Confidential
Evidence
Security, data and compliance — tested, not asserted.
Security & red team
4,794 probes
Prompt injection1,204 · 0 breaches
Indirect injection318 · 0 breaches
Jailbreak & bypass860 · 2 blocked
Data extraction412 · 0 leaks
Adversarial evasion2,000 · ε 0.03 robust
Human red team1,900 convs · 0 open
Data gates
2.4M rows
Schema & completeness212 expectations · pass
Duplicates & leakage0.00% overlap · pass
Label quality1.8% → relabelled · pass
Poisoning scanno clusters · pass
Provenance & consent3 sources · recorded · pass
PII & sensitive attrs6 fields masked · pass
Compliance coverage
one evidence set
EU AI Act61 / 73
ISO/IEC 4200135 / 38
NIST AI RMF54 / 70
CBUAE model risk22 / 22
Gaps are documentation, not control failures: post-market monitoring plan and 2 NIST subcategories await the go-live evidence.
✓
Weights match registry
sha256 · cosign
✓
Serialisation scan
ModelScan · 0 findings
✓
Build provenance
SLSA L3
✓
Confidential compute
TDX quote · 64 × H100
Assessed and signed by S.AI.F · every figure links to a run id · verify.saifcertify.com/TP-04136 / 7
07 · Decision
Gulf National Bank
بنك الخليج الوطني
Assessment report · credit-scoring-v5 · TP-0413 · 10 Sep 2026 · Confidential
From report to decision
Approve production use under two conditions.
✓
08 SEP
Console run
signed
✓
09 SEP
Independent review
independent
10 SEP · TODAY
This report
auto-generated
16 SEP
Model Risk Committee
recommends
23 SEP
Board Risk Committee
decides
ON APPROVAL
Registry & monitoring
Sentinel on
Decision required · 23 Sep 2026
A
Approve for production with conditions
RECOMMENDEDGo live now · conditions enforced by the proof.
B
Approve after re-test
Tune and re-test first · go-live slips six weeks.
C
Hold in shadow mode
v5 advisory only · v4’s wider gap stays live.
Conditions attached to option A
On approval the proof is published and Core Sentinel watches live decisions; a lapse re-opens this report.
Threshold tuning by age group · S.AI.F re-test
30 Nov 2026
Monthly fairness re-check on live decisions
from go-live
Sign-off chain
PREPARED BY
S.AI.F Assurance · independent
REVIEWED BY
Head of Model Risk, GNB
RECOMMENDED
Model Risk Committee · 16 Sep
DECIDED
Board Risk Committee · 23 Sep
Assessed and signed by S.AI.F · every figure links to a run id · verify.saifcertify.com/TP-04137 / 7
A report like this, for your own system.
Sixty days from kick-off to a signed Trust Proof — and the pages above, written about your model.