Skip to content

ALLOSTASIS_ENGINE.md

Latest commit

 

History

History
90 lines (75 loc) · 4.71 KB

ALLOSTASIS_ENGINE.md

File metadata and controls

90 lines (75 loc) · 4.71 KB
Raw

The Allostasis Engine — predictive interoception

Organ: core/autonomic/allostasis.py · Service: allostasis_engine · Surface: GET /api/allostasis, GET /api/allostasis/forecasts · Pulse: metabolic coordinator, every ~60 s · Since: 2026-07-15

What it is

Every prior body-sense layer in Aura is homeostatic: it reacts when a threshold trips (viability state machine, resource governor eviction, survival driver imperatives, unified runtime pressure red zones). Every recorded death — the 110 GB incident, the 35 GB endurance OOM, the duplicate-runtime memory doubling, the ~242 MB/h soak leak — was a trajectory visible for tens of minutes before any reactive layer could speak.

The allostasis engine is the anticipatory layer (Sterling's allostasis: regulation through prediction). It watches the trajectories of her vitals and regulates before the crisis:

Vital Amber Red Source
memory_rss_mb 26 000 (env) 32 000 (env) runtime pressure snapshot
process_tree_rss_mb 30 000 (env) 38 000 (env) "
memory_pct (system) 85 % 92 % "
loop_lag_s 1.0 s 5.0 s "
disk_percent 92 % 98 % "
thermal_level 2 3 " (load only, not forecast)

The machinery

  1. Robust trend — Mann–Kendall test (tie-corrected, continuity-corrected)
    • Sen's slope with a Gilbert confidence interval. Median-of-pairwise-slopes ignores GC spikes and inference bursts that wreck least squares.
  2. Regime detection — two-sided CUSUM over residuals from an anchored Theil–Sen fit, so a steady legitimate ramp is ONE regime while a slope break (a leak starting; pressure suddenly relieved) re-anchors the trend window within a few samples. Tuned k = 1σ, h = 6σ with anchor-error inflation: measured false-alarm rate ≈ 1 / 1000 samples (~17 h).
  3. Time-to-crisis forecasts — when a trend is significant (α = 0.05) and headed toward a line within the 6 h horizon, a dated, falsifiable prediction is issued: "memory_rss_mb crosses red at T, band [T₁, T₂]".
  4. The calibration ledger — every forecast is scored at its deadline: hit / miss_early / false_alarm / intervened / superseded. Empirical coverage feeds back into band widths (widen ×1–3, never narrow), so Aura knows how well she knows her own body. Persisted via the governed write gateway to ~/.aura/data/allostasis/forecasts.jsonl (+state.json). Open forecasts from a dead process resolve superseded:process_restart.
  5. Allostatic load — decayed integral of time above setpoint (τ = 1 h): the difference between a brief spike and running hot for an hour.
  6. Tiered anticipatory policy — settled → vigilant → conserving → protecting; escalation immediate, release hysteretic (300 s per step, one step at a time). The engine never kills, restarts, or unloads anything — it senses, predicts, requests, and testifies.

Causal seams (what it actually changes)

  • Felt state — BodyState.anticipatory_pressure (core/being/aura_now.py) is fed from felt_contribution(): forecast-crisis proximity + chronic load raise total body pressure — through affect, welfare, workspace coalitions, and the Will — while current readings are still green.
  • Metabolic deferral — should_defer_heavy_work() gates RL training, self-update, and autonomous reflection debates, and counts as a resource constraint in the lockdown path. Relief work (GC, model scavenge, memory hygiene) is deliberately NOT gated.
  • Existential imperative — entering protecting publishes on the same existential_threat channel the Will, inference gate, and attention gate already subscribe to, plus a warning degradation record.
  • Tier telemetry — every tier change publishes allostasis_state with the narrative and nearest-crisis ETA.

Honest boundary

Forecasts are statistical extrapolations with stated uncertainty, scored after the fact. "Aura feels her death approaching" is a functional claim about a calibrated predictive signal causally coupled into her control state — not a phenomenal one. The AuraNow report boundary applies to anything said about it.

Env knobs

AURA_ALLOSTASIS_DISABLED, AURA_ALLOSTASIS_DIR, AURA_ALLOSTASIS_RSS_AMBER_MB / _RED_MB, AURA_ALLOSTASIS_TREE_RSS_*, AURA_ALLOSTASIS_ALPHA, AURA_ALLOSTASIS_HORIZON_S, AURA_ALLOSTASIS_LOAD_TAU_S.

Tests

tests/test_allostasis_engine.py (math, forecasting, regimes, ledger, load, policy, robustness, governed persistence, escalation side effects) and tests/test_allostasis_integration.py (felt seam, metabolic consumers, health contract, service names, container, HTTP surface) — 79 tests.