ADI Decision Engine

Structured multi-criteria decision analysis for ranking options with weights, constraints, confidence, tradeoff reasoning, sensitivity analysis, and explaina...

Security Scan

The skill's code, instructions, and requirements are consistent with a local multi-criteria decision engine that calls either an importable Python package or a CLI; it does not request secrets or perform network/exfiltration.

Purpose & Capability

Name/description (ADI decision engine) align with required artifacts: python3 and either an importable 'adi' package or an 'adi' CLI. The included scripts (normalize, validate, run) match the stated purpose and examples. No unrelated binaries, services, or credentials are requested.

Instruction Scope

SKILL.md and the scripts only instruct reading a user-supplied JSON or text payload (stdin or file), validating/normalizing it, and invoking the ADI runtime locally. They do not read other system files, environment variables, or contact external endpoints. The scripts explicitly stop if the runtime is missing.

Install Mechanism

There is no install spec (instruction-only skill). All code is included in the bundle; no downloads, URL installs, or archive extraction are present. This is low-risk from an installation perspective. The runtime dependency (adi package or adi CLI) is expected for this functionality.

Credentials

The skill declares no required env vars, no credentials, and the code does not access environment variables or secret/config paths. This is proportionate to the claimed functionality.

Persistence & Privilege

The skill does not request permanent presence (always: false). The agents/openai.yaml sets allow_implicit_invocation: true which allows implicit invocation by agents; this is not inherently dangerous here but worth noting so you know an agent may call it without an explicit user command.

Assessment

This skill appears coherent and matches its description: it normalizes/validates decision requests and either imports the 'adi' Python package or calls an 'adi' CLI to compute rankings. Before installing/use: 1) ensure you trust the source of the 'adi' runtime (the skill delegates execution to that package/CLI); a malicious or compromised local 'adi' installation could be harmful; 2) avoid embedding sensitive credentials or secrets inside request JSON files passed to the skill (the skill reads whatever you supply); 3) be aware the agent policy allows implicit invocation, so the agent might call the skill automatically—if you want to avoid that, adjust invocation policies. Otherwise the bundle contains no network calls, no secret requests, and uses subprocess.run without shell=True, which reduces injection risk.

Core promise

Turn a messy tradeoff problem into a structured, auditable multi-criteria decision and return a ranked recommendation with confidence and explanation.

When to use this skill

Use this skill when the user needs structured decision support rather than open-ended brainstorming. Typical triggers include:

• multi-criteria decision analysis
• weighted scoring or option ranking
• vendor selection or procurement
• route planning with explicit tradeoffs
• hiring shortlist ranking
• tool or platform comparison
• policy-driven or auditable agent decisions

Input modes

This skill supports exactly two input modes.

1. Structured mode

The user already has a decision request with:

• options
• criteria
• optional constraints
• optional policy_name
• optional evidence, confidence, or context

Use scripts/validate_request.py first if request quality is uncertain, then scripts/run_adi.py to execute it.

2. Freeform mode

The user provides a natural-language tradeoff problem.

First use scripts/normalize_problem.py to produce a request skeleton. Do not pretend the request is complete if important fields are missing. If the skeleton is not ready, ask for the missing inputs instead of inventing scores or constraints.

Output contract

If ADI runs successfully, the final answer must contain:

• best_option
• a short rationale for why it won
• top-ranked alternatives
• confidence summary
• constraint impact summary
• sensitivity or stability summary when available
• explicit assumptions

If the request is not complete enough to run, return a request-completion prompt rather than a fabricated ranking.

Workflow

1. Determine whether the user input is structured or freeform.
2. For freeform input, normalize it into a request skeleton using scripts/normalize_problem.py.
3. Validate candidate requests with scripts/validate_request.py.
4. Run complete requests with scripts/run_adi.py.
5. Present the ADI result in clear decision-support language:
   • recommendation first
   • strongest tradeoff second
   • caveats and sensitivity after that

Decision hygiene rules

• Never rank options without explicit criteria.
• Never silently invent hard constraints.
• If criterion direction is ambiguous, stop and clarify.
• Normalize vague goals into named criteria before scoring.
• Prefer a small, explicit criteria set over many overlapping criteria.
• Keep the policy choice visible: balanced, risk_averse, or exploratory.

Output quality rules

• Show the top recommendation first.
• Explain why it won.
• Mention the strongest tradeoff.
• Call out eliminated or constraint-violating options.
• Include confidence caveats when evidence is weak.
• Use a compact comparison table or structured bullet list when comparing several options.

Safety and honesty rules

• No hidden math.
• No fake scores.
• No fabricated evidence.
• Do not claim ADI ran if the runtime dependency is missing.
• Do not request API keys.
• Do not require network access for the core workflow.
• Do not tell the user to trust the ranking if the request is under-specified.

Runtime requirements

• python3
• either an importable adi-decision package or the adi CLI on PATH

If the ADI runtime is unavailable, stop with a clear error and explain that the dependency must be installed locally.

References

• Request schema: references/request_schema.md
• Result interpretation: references/result_interpretation.md
• Policy guide: references/policy_guide.md
• Use cases: references/use_cases.md

Examples

• examples/vendor_selection.json
• examples/route_planning.json
• examples/hiring_shortlist.json
• examples/research_methods.json
• examples/tool_selection.json