Guardrails & evals
Guardrails & Evals Review Packet
A guardrail/eval coverage map, failure-mode taxonomy, and release decision gates.
The first deliverable is a decision package: what is in scope, what is required, what is blocked, and what can proceed.
No production testing, adversarial activity, access to secrets, or customer-data processing happens without explicit authorization and the right agreement path. Do not enter secrets or credentials here.
What to gather for this packet
Needed: Current guardrails · Eval suite & coverage · Release criteria · Known failure modes
Helpful: Logs/metrics · Abuse cases · Model details
Bring names and high-level descriptions only — exact targets, accounts, and credentials are shared later through a secure channel.
Engagement
Readiness inputs
Deliverables you want
Packet modules
- Scope Brief — What is in scope, what decision is being made, and what success looks like.
- Authorization Statement — Confirmation the org owns, controls, or is authorized to assess the targets.
- Evidence Handling Plan — Where evidence is stored, redaction, retention, and deletion.
- Contract Requirements — The required and conditional agreements for this engagement.
- Draft SOW Inputs — Scope, window, deliverables, and acceptance inputs for the SOW.
- Open Questions — What is still required before a scoping call or private offer.
- Follow-On Recommendations — Natural next services once this engagement completes.
- Deliverables Plan — The artifacts the buyer will receive and in what format.
Guardrails & Evals Review Packet — preview
Guardrails & evals
Scope Brief
- Organization: To be specified during scoping
- Decision: A guardrail/eval coverage map, failure-mode taxonomy, and release decision gates.
- Driver: Our guardrails, evals, refusal behavior, and release criteria have unknown coverage and failure modes.
Authorization Statement
- Not yet confirmed — required before any access or testing.
Evidence Handling Plan
- Evidence stored in an access-controlled encrypted store; redaction; agreed retention + deletion.
Contract Requirements
- Mutual NDA — Required for this engagement.
- Evidence Handling Policy — Required for this engagement.
- Statement of Work Template — Required for this engagement.
- Assessment Terms Addendum — Required for this engagement.
- No-Cost Scoping Retainer (conditional) — Scope before any paid work or active testing.
Draft SOW Inputs
- Engagement: Guardrails & Evals Review Packet
- Deliverables: 7 selected
- Budget category: AI product / product security / trust & safety
Open Questions
- Authorization (own / control / explicitly authorized to assess the targets)
Follow-On Recommendations
- ai red team adversarial testing
- secure ai sdlc
- model provider risk
Deliverables Plan
- Guardrail Inventory
- Eval Coverage Map
- Failure Mode Taxonomy
- Abuse Test Plan
- Metrics and Decision Gates
- Remediation Recommendations
- Monitoring Requirements
Onboarding
Move four tracks in parallel
We put legal, finance, procurement, and technical scoping on parallel rails so the work can start without waiting on every internal process sequentially.
Technical Scoping
Output: Draft Launch Review Plan
- architecture
- demo/staging
- prompts
- RAG
- agents/tools
- authz
- logs/evals
- test boundaries
Legal
Output: NDA + Scoping Authorization
- mutual NDA
- data handling
- authorized testing boundaries
- confidentiality
- work-product terms
Finance / Procurement
Output: Procurement Packet
- vendor profile
- tax/payment details
- budget category
- fixed-fee quote path
- invoice terms
- onboarding answers
Internal Approval
Output: Approval Memo
- why now
- business pressure
- risk if delayed
- expected deliverables
- timeline
- decision needed
Output
Your output: a clear, measurable prescription.
SCOPE delivers a one-page engagement plan you can share and act on immediately.
- Situation & core problem
- Desired outcome & success criteria
- Key risks & assumptions
- Recommended path(s)
- Effort, timing & impact
- Open questions & next step
SCOPE Prescription
Situation
Criteria
Recommended path
Next step
Export as markdown ↗