OH

Model, Prompt & Tool Registry

Versioned, reviewable, reproducible AI behavior: prompts, schemas, model policies and tools with owners, test results, effective dates, pinned provider snapshots and environment bindings. Promotion runs test → staging → production with rollback.

AI-02AI-07
Evaluation runs
Draft Peer review Evaluation — failed Staging soak Approval Production
All states Approved In review Blocked Deprecated
PromptVersionOwnerApproval stateRequired evaluationEnv bindingsEffective from
PRM-LEADSUM · Lead summary & next action
Use case: CRM recommendations (AI-05)
v14 RF Rashid Al Farsi Blocked — eval failed EV-GRD-01 groundedness ≥ 0.92 · 0.87 test
PRM-LEADSUM · Lead summary & next action
Current production version
v13 RF Rashid Al Farsi Approved EV-GRD-01 · 0.94 Pass prod staging 02 Jun 2026
PRM-LISTCOPY · Listing copy EN/AR
Grounded facts + disclaimer; AR variant reviewed
v9 EP Elena Petrova Approved EV-FCT-02 fact fidelity ≥ 0.95 · 0.97 Pass staging Pending promotion
PRM-CONTRACTQA · Contract QA & RAG
Citations + clause rules; legal review required
v6 MS Mariam Al Suwaidi In review — Legal EV-CIT-03 citation coverage ≥ 0.98 · 0.99 Pass test
PRM-KYCX · KYC document extraction
Feeds AI-04 pipeline; maker-checker downstream
v11 YK Yusuf Karim Approved EV-EXT-01 field accuracy ≥ 0.97 · 0.98 Pass prod 19 May 2026
PRM-CALLSUM · Call summary (pilot)
Recording policy gated; no employment action
v3 RF Rashid Al Farsi Pilot — staging only EV-GRD-01 · 0.93 Pass staging
6 of 23 prompt versions · deprecated hidden
Model policyPinned snapshotProvider profileApproval stateEnv bindingsEffective
MDL-GEN-STD · Standard generation
Reproducibility required → snapshot pinned
gpt-4o@2026-03-12PRV-AZO-01Approvedprod01 Apr 2026
MDL-TOOL-CP · Copilot tool use
Tool allowlist enforced by gateway
claude-sonnet-4-5@2026-01-12PRV-BED-02Approvedprod15 Apr 2026
MDL-EXT-OCR · Document extraction
On-prem; Restricted-PII permitted
idp-layoutlm-v3.4.1PRV-OCR-04Approvedprod19 May 2026
MDL-GEN-NXT · Next-gen generation
Candidate upgrade; A/B eval in progress
gpt-5-erp@2026-06-30PRV-AZO-01In reviewtest
ToolVersionPrivilegeServer validationApproval stateEnv bindings
TOOL-CRM-READ · Read lead/customer context
Caller RBAC re-checked per call
v5Read-onlyScope + row-level permissionApprovedprod
TOOL-PROP-MATCH · Query matching inventory
Delegates to matching engine
v4Read-onlyAvailability + masking rulesApprovedprod
TOOL-DRAFT-JE · Draft journal entry
Proposes only — enters finance approval flow, never posts (AI-07)
v2Draft-createBalanced lines + dimension rulesApprovedprod
TOOL-DOC-CLASSIFY · Classify uploaded document
v7Read-onlySchema + confidence floorApprovedprod
TOOL-DRAFT-REPLY · Draft customer reply
Send stays manual; consent & DND checked at send
v3Draft-createTemplate + consent policyIn reviewstaging
TOOL-PAY-EXEC · Execute payment
Requested by copilot team 03 Jul
Write — high riskRejected — violates AI-07 boundary
SchemaVersionUsed byApproval stateBreaking change policy
SCH-INV-EXTRACT · Invoice extraction output
Fields, confidence, source coordinates, validation errors
v8PRM-KYCX, RT-EXT-INVApprovedAdditive only within major
SCH-KYC-EXTRACT · KYC document output
EID/passport fields, masked render rules
v6PRM-KYCXApprovedAdditive only within major
SCH-RECO-CARD · Recommendation card
Score, confidence, main factors, feedback enum
v4PRM-LEADSUM, recommendationsApprovedAdditive only within major
Selected: PRM-LEADSUM — environment bindings
EnvironmentBound versionPinned model snapshotActivated byActivatedRollback target
testv14gpt-4o@2026-03-12Rashid Al Farsi15 Jul 2026v13
stagingv13gpt-4o@2026-03-12Vikram Nair28 May 2026v12
prodv13gpt-4o@2026-03-12Vikram Nair (approved: Ahmed Al Blooshi)02 Jun 2026v12 · one-click guided rollback
Content hash v14: sha256:7c1f…e2a9 · v13: sha256:44d0…91bc
Changelog — PRM-LEADSUMFull audit trail
  • 16 Jul 2026 · 17:05 · EVAL-2418
    v14 evaluation failed — promotion blocked
    EV-GRD-01 groundedness 0.87 < 0.92. Regression traced to shortened context window in retrieval step. Blocked
  • 15 Jul 2026 · 11:32
    v14 draft created by Rashid Al Farsi
    Change note: "Add next-best-action ranking for off-plan leads; tighten AR tone." Peer review: Layla Hassan.
  • 02 Jun 2026 · 09:00
    v13 activated in production
    Approved by Ahmed Al Blooshi. Effective date 02 Jun 2026, rollback target v12 retained. Active
  • 27 May 2026 · 14:18 · EVAL-2201
    v13 evaluation passed (0.94)
    Golden dataset GD-CRM-07 (412 cases), adversarial suite ADV-11 passed.
  • 03 Mar 2026 · 10:44
    v12 deprecated → rollback-only
    Retained for reproducibility of runs before 02 Jun 2026.
Promote to production — approval preview

A version that fails its required evaluation cannot reach this dialog — the promote action is disabled at source (AI-02 acceptance). Registry changes follow the same governance as configuration version lifecycle.

Static render mockup — no live data. Source: ERP Modern Development Specification v2.0 · Covers: AI-02 (versioned registry, pinned snapshots, gated promotion & rollback), AI-07 (tool allowlist states).