Synthetic case · AIDCF-015 reference implementation v1.4 (draft)
One AI capability, traced end to end
A synthetic AI training and inference capability, assessed with the AIDCF method. Follow a failure from its origin to the service outcome it breaks, see which safeguards and controls should stop it, and how the evidence drives the gates and the final decision.
Synthetic data. This case describes no real facility. Every value is illustrative, and the approval record is synthetic: no human approved this decision.
01 Failure
Heatwave above design ambient. The cooling chain loses capacity.
- External supplyMake-up water supply
- Facility engineeringFacility water systemUNKNOWN U-003
- Facility engineeringChiller
- Facility engineeringCDU 2
- Facility engineeringLiquid-cooling secondary loop
Origin and path
HZ-02 Heatwave above design ambient (environmental hazard). Directly exposed: CI-001, Facility water system.
The failure starts in NORMAL operation and propagates through 7 recorded steps. Analysis confidence CONF-2.
02 Impact
Accelerators lose cooling, so inference latency breaches its tolerance.
- AI technology chainAccelerator compute cluster
- AI service functionProvide accelerator compute
- Service outcomeInference latency
Service outcome
SO-03 Inference latency: p99 request latency at the service edge, tolerance p99 below 300 ms at declared load. Applies in NORMAL and DEGRADED.
Illustrative: outcome breaches tolerance without the expected safeguards.
03 Evidence
The safeguards that should contain it are not shown to work.
- Leak detection and isolationTHM-019 Leak Detection Not assessedOPS-013 Alarm Rationalisation Partial
- Accelerator thermal throttlingCMP-019 Error and Degradation Detection Not assessed
- Exclusivity of the facility water plantUNKNOWN U-003, severity US5
Findings and unknowns
F-003 on MCF-OPS-013: Cooling and leak alarms are not rationalised. Critical alarms may not reach the operator with a defined response.
U-003 on Facility water system: It cannot be determined whether the facility water plant is shared with other tenants. Needed: Facility water service agreement and plant schematic.
04 Decision
Not assured
Alarm Rationalisation is a critical input to CG-07 Critical control-system integrity. A partial result fails the gate, and a failed gate caps the decision however good the rest is.
Why the gate fails
CG-07 has 6 critical controls. Under the AIDCF-010 roll-up rule a PARTIAL result never lets a gate pass.
- MCF-OPS-013 Alarm Rationalisation Partial
- MCF-POW-021 EPMS Coverage Not assessed
- MCF-SEC-005 BMS and EPMS Security Partial
- MCF-SEC-009 Security Zoning Not assessed
- MCF-SEC-023 Command Allowlisting Not assessed
- MCF-THM-021 Control Stability Not assessed
Walk through it step by stepDrawn from the AIDCF-015 synthetic dataset. It describes no real facility.
This is one of 12 failure scenarios instantiated in this case. The full graph below is the expert view: every element, dependency and scenario. The scenario library has the twenty cross-domain reference scenarios of AIDCF-004.
The decision question
Can the synthetic AI training and inference capability CI-001 be relied on to meet SO-01 to SO-03, as far as they apply in NORMAL and MAINTENANCE states, as at 2026-10-10? SO-02 (RECOVERY only) and the DEGRADED state of SO-03 are outside the assessed scope.
Scope: Synthetic capability CI-001 within snapshot SNAP-001; NORMAL and MAINTENANCE states.
Required service outcomes
| Outcome | Measure | Tolerance | Operating states |
|---|---|---|---|
| SO-01 Training availability | training cluster available for scheduled jobs | at most 4 hours unplanned unavailability per 30 days | NORMAL, MAINTENANCE |
| SO-02 Training checkpoint recovery | restart from last checkpoint after disruption | restart within 2 hours, losing at most 30 minutes of training progress | RECOVERY |
| SO-03 Inference latency | p99 request latency at the service edge | p99 below 300 ms at declared load | NORMAL, DEGRADED |
Dependencies and failure paths
30 elements and 17 recorded dependencies. 2 shared fault domains make apparently redundant equipment fail together.
Synthetic AI training and inference capability
Synthetic. Describes no real facility.
Synthetic data hall A, its power, cooling, facility water, compute, storage, fabric and supervisory control; the utility is external.
Illustrative: 2.0 MW IT load with accelerator transients; training jobs of up to 14 days.
How to use this
- Select any box to see what it relies on, what relies on it, and what can go wrong.
- Start the guided journey to walk one failure step by step, from the plant through the AI systems to the evidence, the gates and the decision.
- Choose a failure scenario to trace it from its origin to the service outcome it breaks.
- Boxes marked UNKNOWN carry unresolved evidence gaps that cap the decision.
Critical gate results
3 gates fail, 5 are UNKNOWN and 7 were not assessed, so none passes. Only 12 of 360 controls were assessed. Select a gate to see why.
CG-03 · In scope · evidence sufficiency S2
Critical dependency visibility FAIL
Material internal and external dependencies, common-mode exposure and concentration are identified and owned
Why: FAIL: CRITICAL control(s) MCF-DEP-001 PARTIAL; criteria not met: GC-03-1.
Effect on the decision: Gate FAIL caps the decision at LIMITED.
Assessed controls
- MCF-DEP-001 Asset-to-Function Mapping (critical) PARTIAL
No evidence gap: all required classes present at S4, floor S4.
Evidence detail: 2 objects and 1 test
Required classes: E1 Documentary and declarative; E2 Configuration and technical state.
Methods: document review, reconciliation, inspection, sampling.
- EV-002 E1 Documentary and declarative · S4 · quality Q3 · 2026-07-01 to 2026-09-30 · NORMALSHA-256 88d4a879ebf6951d… fixture EV-002
- EV-015 E2 Configuration and technical state · S4 · quality Q3 · 2026-07-01 to 2026-09-30 · NORMALSHA-256 c915f2708431fecc… fixture EV-015
- SMP-001 Sample: Sample of required functions for MCF-DEP-001 (Whole population)
- EV-002 E1 Documentary and declarative · S4 · quality Q3 · 2026-07-01 to 2026-09-30 · NORMAL
- MCF-POW-001 Utility Capacity (supporting) CONFORMING
No evidence gap: all required classes present at S5, floor S5.
Evidence detail: 4 objects
Required classes: E1 Documentary and declarative; E2 Configuration and technical state; E3 Operational and observational; E7 Derived and analytical.
Methods: document review, analysis, reconciliation.
- EV-003 E1 Documentary and declarative · S5 · quality Q3 · 2026-07-01 to 2026-09-30 · NORMALSHA-256 b8f4dae1acbaf7f2… fixture EV-003
- EV-016 E2 Configuration and technical state · S5 · quality Q3 · 2026-07-01 to 2026-09-30 · NORMALSHA-256 78768a5984a2dda6… fixture EV-016
- EV-017 E3 Operational and observational · S5 · quality Q3 · 2026-07-01 to 2026-09-30 · NORMALSHA-256 5c9f749bea5d4fba… fixture EV-017
- EV-018 E7 Derived and analytical · S5 · quality Q3 · 2026-07-01 to 2026-09-30 · NORMALSHA-256 008c40d6c9f50bb7… fixture EV-018
- EV-003 E1 Documentary and declarative · S5 · quality Q3 · 2026-07-01 to 2026-09-30 · NORMAL
8 of 10 mapped controls were not assessed. 1 criterion not met, 5 criteria not evaluated, 1 criterion unknown.
F-001 (Weakness, MCF-DEP-001): Asset-to-function mapping omits the facility water plant and recovery dependencies. Shared-source dependencies for RF-02 cannot be analysed.
U-003 (US5): Exclusivity of the facility water plant. Systemic: decision capped at INCONCLUSIVE by this UNKNOWN alone; CG-03 and CG-06 cannot pass.
The assurance decision
Decision state
NOT_ASSURED
Controlling reason: CG-07 FAIL (failCeiling NOT_ASSURED), Critical control-system integrity. The most restrictive ceiling wins; it is not an average.
Why the score band is withheld: coverage 3.0556% below 80%; US4/US5 UNKNOWN open in scope. AIDCF withholds the band below 80% coverage, or while a US4 or US5 UNKNOWN is open, so a score cannot imply more than the evidence supports.
Valid until 2027-04-10 unless a reassessment trigger fires first. Synthetic record; no human approval was given.
Other ceilings that did not control
Each would cap the decision on its own; NOT_ASSURED is more restrictive than all of them.
- CG-01 UNKNOWN (INCONCLUSIVE)
- CG-02 NOT_ASSESSED (LIMITED)
- CG-03 FAIL (LIMITED)
- CG-04 NOT_ASSESSED (LIMITED)
- CG-05 UNKNOWN (INCONCLUSIVE)
- CG-06 UNKNOWN (LIMITED)
- CG-08 UNKNOWN (LIMITED)
- CG-09 NOT_ASSESSED (LIMITED)
- CG-10 UNKNOWN (LIMITED)
- CG-11 NOT_ASSESSED (LIMITED)
- CG-12 NOT_ASSESSED (LIMITED)
- CG-13 NOT_ASSESSED (LIMITED)
- CG-14 NOT_ASSESSED (LIMITED)
- CG-15 FAIL (LIMITED)
- U-001 US3 (CONDITIONALLY_ASSURED)
- U-003 US5 (INCONCLUSIVE)
- U-004 US4 (LIMITED)
- U-005 US4 (LIMITED)
Open UNKNOWN records
- U-001 US3 E4 test evidence for MCF-RES-017. Control result UNKNOWN; decision capped at CONDITIONALLY_ASSURED by this UNKNOWN alone. Evidence needed: Component test reports (E4).
- U-002 US2 Current fabric topology export. Limitation recorded; no ceiling. Evidence needed: Fresh topology export.
- U-003 US5 Exclusivity of the facility water plant. Systemic: decision capped at INCONCLUSIVE by this UNKNOWN alone; CG-03 and CG-06 cannot pass. Evidence needed: Facility water service agreement and plant schematic.
- U-004 US4 Integrated cross-domain recovery exercise (GC-10-5). Critical: CG-10 UNKNOWN; decision capped at LIMITED by this UNKNOWN alone. Evidence needed: Integrated recovery exercise report.
- U-005 US4 Achieved security level (SL-A) of the BMS zone. Critical: CG-07 cannot pass; decision capped at LIMITED by this UNKNOWN alone. Evidence needed: Zone SL-A assessment.
Check it yourself
Everything on this page is generated from the reference implementation at commit 8b0a941.
Data
Re-run the checks
git clone https://github.com/Sivas1187/AIDCF cd AIDCF && git checkout 8b0a941 cd implementation/aidcf-015-v1.4 pip install -r requirements.txt python3 -B tools/run_all.py
Score calculation hash: 5783596f4e5330457c0fb74dd333f2be893a49b2cccf6bcc2787fd617f980060
Terms such as not assessed and the US1 to US5 UNKNOWN severities are defined in AIDCF-010 and the vocabularies in the repository.