BDSM enforcement policy is undeployable: redacted thresholds [9.99] lack configuration guide

Problem

The BDSM (Behavioral Inauthentic-Account Detection) module's safety thresholds
are redacted in the public release to prevent adversarial evasion, but the
documentation and code lack clear guidance on how to create custom enforcement policies.

This undermines the repository's transparency mission, as users cannot understand:

• What threshold values are expected
• How to configure their own operating points
• The relationship between sentinel 9.99 and valid threshold ranges

Current State

Redacted threshold locations:

• bdsm/runtime/sink_policy.yaml (lines 8-31): All thresholds set to 9.99 sentinel
• bdsm/runtime/score_results_sink_focal.py (lines 175-207): Default thresholds all set to 9.99

Example of redaction:

thresholds:
  FollowBot:           [9.99, 9.99]
  EngagementAmplifier: [9.99, 9.99]
  RTBot:               [9.99, 9.99]

All probabilities in these fields should be in the range [0, 1], making 9.99
an obvious placeholder (from bdsm/README.md lines 104-112).

Documentation issue:
The README states users should "Supply your own via --policy-file / BDSM_SINK_POLICY"
but provides no template or example showing:

• The expected structure of each threshold (2-tuple: [tau, lambda])
• Valid value ranges ([0, 1] for probabilities)
• Which fields are safe to modify (vs. which control system behavior)

Impact

Users attempting to run BDSM cannot deploy it without either:

1. Reverse-engineering the policy structure from code parsing
2. Providing arbitrary 9.99 values that will never trigger enforcement
3. Guessing appropriate thresholds without domain knowledge

This is particularly problematic for the stated goal of the repo: public transparency
into how the For You feed works.

Proposed Solution

1. Create example template: bdsm/runtime/sink_policy.example.yaml

   • Show realistic threshold values in [0, 1]
   • Comment each field explaining its purpose
   • Label which values are likely to be too conservative/aggressive

2. Update bdsm/README.md:

   • Document the policy file format with concrete examples
   • Explain the 2-tuple structure: [tau, lambda] as threshold and ceiling
   • Add a "Configuration" section showing how to set policy via CLI or env var
   • Link to the example file

3. Add validation at startup:

   • Warn or fail if thresholds are exactly 9.99 (sentinel value)
   • Log policy version and source on load (already done, but clarify in README)

References

• README lines 104-112: Explanation of redaction
• README line 112: "Supply your own via..."
• sink_policy.yaml: Current redacted file
• score_results_sink_focal.py lines 175-207: Default policy dataclass
• score_results_sink_focal.py lines 220-260: Policy loading logic

Start with bdsm/README.md lines 104-112, then inspect bdsm/runtime/sink_policy.yaml, score_results_sink_focal.py lines 175-207, and its policy-loading logic at lines 220-260. Done means users have an example policy, documented tuple and value ranges, configuration instructions, and clear handling of the 9.99 sentinel.