DPIA mini-template, retention rules, and enforcement hooks.
L7) DPIA / Risk mini-template (copy-declare, fast)
- system: <service / component>
- purpose: bounded alignment for selection / routing / audit
- data_in: declared, dimensionless aggregates; no PII
- lawful_basis: legitimate interests (quality / safety measurement)
- pii_controls: none collected; structured tags if needed; retention limited
- risks: leakage via free-text logs; overfitting lenses to sensitive attributes
- mitigations: structured logs only; manifest review; drift checks
- residual_risk: low
- sign-off: <owners/date>
knobs_hash=<sha256>
This keeps privacy decisions explicit, auditable, and lightweight.
L8) Retention & rotation (default posture)
- Hot logs: 30-day rolling
Fields:iso_utc, svc, knobs_hash, m, a, U, W, RSI, g_t, RSI_env, band, stamp
→ no free text - Cold roll-ups: 12 months, hourly aggregates
(U,Wonly → reproduce alignment drift trends) - Deletion: stamp-chain cryptographic erasure when outside policy window
- Control: manifest declares the policy, CI enforces it
Retention remains just enough for reproducibility & monitoring.
L9) Enforcement hooks (reference pseudocode)
def enforce_governance(item, manifest, state):
# 1) privacy guard
assert item["agg_only"] and not item.get("free_text")
# 2) clamp + fuse
a = clamp_align(item["a"], eps_a=manifest["eps_a"])
state["U"] += item["w"] * atanh(a)
state["W"] += item["w"]
# 3) chooser + gate (alignment-only)
RSI = tanh((state["V_out"] - state["U_in"]) / max(state["W_in"], manifest["eps_w"]))
RSI_env = apply_gate(RSI, state["g"], mode=manifest["gate"]["mode"])
# 4) band policy
band = to_band(RSI_env)
if band in ["A--", "A-"]:
return fallback_classical(item["m"]) # m untouched
return {"m": item["m"], "RSI_env": RSI_env, "band": band}
→ Classical fallback is the safety net: if anything is off, use m directly and stamp the cause.
Navigation
Previous: SSM-AI – Appendix L — Governance Quick Reference (L4–L6)
Next: SSM-AI – Appendix L — Governance Quick Reference (L10–L12)
Directory of Pages
SSM-AI — Table of Contents
SHUNYAYA 