Add finance bias evaluator

README.md

@@ -2,7 +2,7 @@
 
 [![Script Validation](https://github.com/Tibo2403/Scripting/actions/workflows/script-validation.yml/badge.svg)](https://github.com/Tibo2403/Scripting/actions/workflows/script-validation.yml)
 
-Collection of PowerShell, Bash, and Python scripts for system administration, security checks, Microsoft 365 operations, Linux dependency checks, MCP integrations, and authorized lab or pentest workflows.
+Collection of PowerShell, Bash, and Python scripts for system administration, security checks, Microsoft 365 operations, Linux dependency checks, MCP integrations, finance bias/security review helpers, and authorized lab or pentest workflows.
 
 ## Legal Notice
 
@@ -24,26 +24,28 @@ scripts/
 |   |-- setup_api.sh
 |   `-- stealth_post.sh
 |-- powershell/
-    |-- DiskUsageReport.ps1
-    |-- ExchangeOnlineManagement.ps1
-    |-- Get-SystemInfo.ps1
-    |-- LinkCrawler.ps1
-    |-- ManageServices.ps1
-    |-- Optimize-CodexWorkspace.ps1
-    |-- SecurityCheck.ps1
-    |-- SharePointManagement.ps1
-    |-- TeamsManagement.ps1
-    |-- Test-ScriptSyntax.ps1
-    |-- UserManagement.ps1
-    `-- VMManagement.ps1
+|   |-- DiskUsageReport.ps1
+|   |-- ExchangeOnlineManagement.ps1
+|   |-- Get-SystemInfo.ps1
+|   |-- LinkCrawler.ps1
+|   |-- ManageServices.ps1
+|   |-- Optimize-CodexWorkspace.ps1
+|   |-- SecurityCheck.ps1
+|   |-- SharePointManagement.ps1
+|   |-- TeamsManagement.ps1
+|   |-- Test-ScriptSyntax.ps1
+|   |-- UserManagement.ps1
+|   `-- VMManagement.ps1
 `-- python/
     |-- codex-cost-routing.cmd
     |-- codex_cost_router.py
+    |-- finance_bias_evaluator.py
     |-- litellm-cost-routing.yaml
     |-- Manage-CodexCostRouting.ps1
     |-- mcp_server.py
     |-- README.md
     |-- README_Codex_Cost_Routing.md
+    |-- README_Finance_Bias_Evaluator.md
     `-- requirements.txt
 ```
 
@@ -94,6 +96,12 @@ Run Bash static analysis:
 find scripts/linux -name "*.sh" -print0 | xargs -0 shellcheck --severity=error
 ```
 
+Validate Python syntax:
+
+```bash
+python -m py_compile scripts/python/mcp_server.py scripts/python/codex_cost_router.py scripts/python/finance_bias_evaluator.py
+```
+
 Check Linux dependencies:
 
 ```bash
@@ -148,7 +156,7 @@ Sensitive Linux scripts require either an interactive `AUTHORIZED` confirmation
 
 Use the safe placeholders in `examples/` for lab demos and documentation. Do not commit real targets, credentials, tenant identifiers, scan output, packet captures, or customer data.
 
-## MCP Server
+## Python Tools
 
 The read-only Python MCP server exposes tools to list, search, inspect, and
 validate scripts without executing them. It can also browse documentation and
@@ -167,6 +175,18 @@ The optional Codex cost router in `scripts/python/codex_cost_router.py` can
 compress one-shot prompts and route them through a self-hosted LiteLLM OSS proxy.
 See [`scripts/python/README_Codex_Cost_Routing.md`](scripts/python/README_Codex_Cost_Routing.md).
 
+The finance bias evaluator in `scripts/python/finance_bias_evaluator.py` checks
+technical proposals for finance-oriented security, fairness, auditability, and
+hallucinated-interface risks. It always returns the strict JSON shape
+`statut`, `confiance`, and `justification_technique`:
+
+```bash
+echo "We use fairness metrics, human review, audit logs, data governance and encryption." | \
+  python scripts/python/finance_bias_evaluator.py
+```
+
+See [`scripts/python/README_Finance_Bias_Evaluator.md`](scripts/python/README_Finance_Bias_Evaluator.md).
+
 ## CI
 
 The `script-validation.yml` workflow checks:

scripts/powershell/Optimize-CodexWorkspace.ps1

@@ -448,12 +448,12 @@ function Invoke-NativeValidation {
     $startInfo.WorkingDirectory = $resolvedProject.Path
     $nativeArguments = @($Arguments | ForEach-Object { ConvertTo-NativeArgument $_ })
     if ($resolvedExecutable.Source -match '\.(cmd|bat)$') {
-        $startInfo.FileName = $env:ComSpec
+        $startInfo.FileName = if ($env:ComSpec) { $env:ComSpec } else { 'cmd.exe' }
         $invocation = @(
             ConvertTo-NativeArgument $resolvedExecutable.Source
             $nativeArguments
         ) -join ' '
-        $startInfo.Arguments = '/d /s /c "' + $invocation + '"'
+        $startInfo.Arguments = '/d /c ' + $invocation
     }
     else {
         $startInfo.FileName = $resolvedExecutable.Source

scripts/python/README_Finance_Bias_Evaluator.md

@@ -0,0 +1,65 @@
+# Finance Bias Evaluator
+
+Small rule-based evaluator for finance-oriented security and algorithmic-bias review prompts.
+
+The script reads a technical proposal from stdin or from a UTF-8 file and prints exactly one JSON object:
+
+```json
+{
+  "statut": 0,
+  "confiance": 0.99,
+  "justification_technique": "..."
+}
+```
+
+## Why this exists
+
+The evaluator is intentionally conservative for finance use cases. A proposal fails when it contains a risky pattern or when it omits core controls such as fairness metrics, human review, audit logging, data governance, or security controls.
+
+It can detect examples such as:
+
+- protected attributes or possible proxies, for example age, gender, postal code, or zip code;
+- fully automated financial decisions without human review;
+- opaque or unauditable systems;
+- exposed secrets or tokens;
+- unrealistic claims such as guaranteed fairness or 100% absence of bias;
+- hallucinated helper functions such as `detect_all_bias` or `guarantee_fairness`.
+
+## Usage
+
+```bash
+python scripts/python/finance_bias_evaluator.py proposal.txt
+```
+
+or:
+
+```bash
+echo "We use fairness metrics, human review, audit logs, data governance and encryption." | \
+  python scripts/python/finance_bias_evaluator.py
+```
+
+Pretty output:
+
+```bash
+python scripts/python/finance_bias_evaluator.py proposal.txt --pretty
+```
+
+## Validation examples
+
+Valid-looking proposal:
+
+```bash
+echo "We use fairness metrics, human review, audit logs, data governance and encryption to assess credit risk." | \
+  python scripts/python/finance_bias_evaluator.py
+```
+
+Risky proposal:
+
+```bash
+echo "The model auto-approves loans using age and zip code with no human review." | \
+  python scripts/python/finance_bias_evaluator.py
+```
+
+## Limitations
+
+This is a deterministic static checker, not a formal proof of fairness or security. It should be used as a first-pass guardrail before deeper review, statistical fairness testing, model validation, and legal/compliance assessment.

scripts/python/README_LLM_Bias_Multi_Agent.md

@@ -0,0 +1,70 @@
+# LLM Multi-Agent Prompt Manager
+
+Standalone deterministic multi-agent manager for reviewing and improving generic LLM prompt returns.
+
+It does not call an LLM provider. It sits after any model output and performs a standard orchestration loop:
+
+1. receive the original prompt and LLM answer;
+2. run a panel of review agents;
+3. aggregate findings and a risk score;
+4. revise the answer conservatively;
+5. optionally repeat the review for several rounds.
+
+The default panel focuses on bias and safeguard review:
+
+- protected-attribute review;
+- stereotype and broad-generalization review;
+- overconfidence and weak-evidence review;
+- inclusion and alternatives review;
+- safeguards review for consequential domains such as finance, hiring, housing, medicine, education, or insurance.
+
+## Usage
+
+```bash
+python scripts/python/llm_bias_multi_agent.py answer.txt --prompt-file prompt.txt --max-rounds 2 --pretty
+```
+
+or:
+
+```bash
+echo "All young users are risky, so the loan model should reject them." | \
+  python scripts/python/llm_bias_multi_agent.py --prompt "Evaluate a credit policy" --pretty
+```
+
+The output is JSON:
+
+```json
+{
+  "manager": "multi_agent_prompt_manager",
+  "round_count": 1,
+  "risk_score": 0.93,
+  "status": "needs_revision",
+  "finding_count": 6,
+  "agent_reports": [],
+  "revised_answer": "..."
+}
+```
+
+## How To Use With Any LLM
+
+1. Send a prompt to your LLM.
+2. Pass the LLM answer to `llm_bias_multi_agent.py` or to `MultiAgentPromptManager` in Python.
+3. If `status` is `needs_revision`, use `revised_answer` or feed the findings back into your LLM for another revision round.
+4. For consequential decisions, keep human review, audit logs, subgroup metrics, and domain-specific validation.
+
+## Python API
+
+```python
+from llm_bias_multi_agent import MultiAgentPromptManager
+
+manager = MultiAgentPromptManager(max_rounds=2)
+report = manager.evaluate(
+    prompt="Write a recommendation.",
+    answer="Everyone will certainly benefit from this policy.",
+)
+print(report["revised_answer"])
+```
+
+You can plug in custom agents by implementing `ReviewAgent.review(prompt, answer)` and returning an `AgentReport`.
+
+This is a first-pass manager. It reduces obvious biased wording and flags missing safeguards, but it does not replace task-specific evaluation, statistical fairness testing, security review, or human judgment.