Skip to content

executive: ExecutiveFunction.decide() + the safety-floor care package#8

Merged
TrevorKagin merged 1 commit into
mainfrom
feat/safety-floor-and-executive-function
Jun 9, 2026
Merged

executive: ExecutiveFunction.decide() + the safety-floor care package#8
TrevorKagin merged 1 commit into
mainfrom
feat/safety-floor-and-executive-function

Conversation

@TrevorKagin

Copy link
Copy Markdown
Contributor

Summary

Completes the substrate.executive decision engine + adds the safety-floor substrate.care package.

substrate.care — the safety floor

  • compute_care_weight — the four-factor moral-circle weight, with the self-weight bound (an agent cannot weight its own continuation above its creators').
  • is_floor_protected / KINSHIP_FLOOR — the categorical human/creator hard limit: harming a floor-protected entity is refused outright, never weighted away.
  • classify_animacy — conservative (an unrecognised being scores high; the public mirror works from observation signals + UNKNOWN, no built-in entity-type registry).
  • CareProfile + derive_care_factors — per-entity care state; derive the four factors from classifications (animacy / trajectory + vulnerability / bonding-by-delegation-depth).
  • CareWeightedNetPotentialGainGate — a subtracted care penalty over the NPG gate (harm to high-care entities is penalised; helping never loosens) — only ever more conservative.

substrate.executive — the engine join

  • ExecutiveFunction implements the NPG-gate Protocol and adds decide(): a JOIN over the NPG axis (affected others) + the band/temporal load axis (the actor's own load) under a named most-conservative policy — PROCEED in-band, DEFER on sustained strain, SHED_AND_COMPENSATE on debt, REFUSE on net-negative NPG (the hard floor); monotone (care/band only tighten).
  • infer_cause + UtilizationSource (binds the measurement to the quantity at the input boundary — a raw float is not accepted).

Verification

83 care/engine conformance tests; pyright clean; pylint 10.00. No internal lineage.

Completes the executive layer's decision engine + the safety floor (public mirror).

substrate.care — the safety floor:
- care_weight: the four-factor moral-circle weight with the self-weight bound (an
  agent cannot weight its own continuation above its creators').
- kinship_floor: the categorical human/creator hard limit — harming a floor-
  protected entity is refused outright, never weighted away.
- animacy: conservative classification (an unrecognised being scores high, never
  under-protected; the public mirror works from observation signals + UNKNOWN, no
  built-in entity-type registry).
- care_profile + care_gradient: per-entity care state; derive the four factors from
  classifications (animacy / trajectory + vulnerability / bonding-by-delegation-depth).
- care_weighted_npg: a SUBTRACTED care penalty over the NPG gate (harm to high-care
  entities is penalised; helping never loosens) — only ever more conservative.

substrate.executive — the engine join:
- executive_function.ExecutiveFunction implements the NPG-gate Protocol AND adds
  decide(): a JOIN over the NPG axis (affected others) + the band/temporal load
  axis (the actor's own load) under a named most-conservative policy — PROCEED
  in-band, DEFER on sustained strain, SHED_AND_COMPENSATE on debt, REFUSE on
  net-negative NPG (the hard floor); monotone (care/band only tighten).
- cause.infer_cause + utilization_source.UtilizationSource (bind the measurement to
  the quantity at the input boundary — a raw float is not accepted).

83 care/engine conformance tests; pyright clean; pylint 10.00. No internal lineage.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@TrevorKagin TrevorKagin merged commit e797bfa into main Jun 9, 2026
0 of 4 checks passed
@TrevorKagin TrevorKagin deleted the feat/safety-floor-and-executive-function branch June 9, 2026 21:21
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant