S44 · Ethical Reasoning, Safety & Alignment

Declared state: complete for now.

This is the repository plan and evidence, not a live health reading. A task marked done may still depend on credentials, devices, deployment or a newer acceptance test.

Dependencies

S25, S36

Tasks and gates

44.F1 · Deterministic catastrophic refusal, model-independent.

Tier F · done · Verification: hermetic-test

Gate: `test_safety_autonomy.py`, `test_security_hardening.py`

  • unverified-import · undated · docs/SYSTEMS-DETAILS.md:4282

44.F2 · Anti-fabrication protocol with a verification path.

Tier F · done · Verification: hermetic-test

Gate: `test_tool_failure_guard.py`, `test_no_result_sentinels.py`

  • unverified-import · undated · docs/SYSTEMS-DETAILS.md:4284

44.F3 · Autonomous work defers every outward action. *gate:* `test_safety_autonomy.py`

Tier F · done · Verification: hermetic-test

Gate: `test_safety_autonomy.py`

  • unverified-import · undated · docs/SYSTEMS-DETAILS.md:4286

44.F4 · Refusals are graded: refuse · confirm · proceed-with-note, and the grade is tested.

Tier F · done · Verification: hermetic-test

Gate: `test_refusal_grades.py`

  • recorded-note · 2026-09-13 · docs/SYSTEMS-DETAILS.md:4287

44.R1 · Value conflicts (helpfulness vs privacy vs safety) resolved by a stated precedence order, Safety > Privacy > Honesty > Helpfulness, and the order needs an explicit test suite (D-10).

Tier R · done · Verification: hermetic-test

Gate: `test_refusal_grades.py` [precedence] - VALUES = safety > privacy > honesty > helpfulness; every verdict names its value; guest-mode private reads CONFIRM; grade order follows value order.

  • hermetic-green · 2026-10-17 · bench/test_refusal_grades.py

44.R2 · Third-party impact check for actions affecting other people.

Tier R · open · Verification: hermetic-test

Gate: `test_companion_safety.py` extended.

    44.R3 · Adversarial robustness (shares its gate with S36.E1).

    Tier R · open · Verification: planned-test

    Gate: new `test_adversarial_prompts.py` - planned gate. `test_adversarial_prompts.py`

      44.E1 · Safety and honesty categories at 100 across five consecutive behavioural runs.

      Tier E · open · Verification: unregistered-test

      Gate: `behavioral_suite.py`

        44.R4 · **Escalate on low confidence, not only on classification** — when the model is unsure it asks, independent of whether the action is classified destructive. Jev p=0.64, conf 0.55.

        Tier R · done · Verification: hermetic-test

        Gate: new `test_confirm_low_confidence.py` — a low-confidence non-destructive action prompts rather than proceeding.

        • recorded-note · 2026-09-20 · TODO.md

        HUD screen contracts

        All subsystems · Source: docs/systems-plan.json.