HMAA SIM, Hardware-Mediated Authority Architecture

· Burak Oktenli · Georgetown University MPS Applied Intelligence · , · 11/11 ✓
ALL VALIDATIONS PASSED
LIVE ENGINE CALCULATOR · Move sliders, A updates instantly
Q, Qualification
0.90
Agent's verified capability score
C, Context
0.80
Mission context legitimacy
E, Escalation pressure
0.20
Threat intensity, higher E → less authority
τ, Sensor confidence
1.00
Signal integrity / crypto token confidence (0 = jam)
QUICK PRESETS
AUTHORITY SCORE A
0.5650
3
SUPERVISED
Actions logged; human operator notified; soft confirmation.
INTERMEDIATESEW INACTIVE
INVARIANTS ✓ I1 I2 I3
w_q_eff0.5500
w_c_eff0.4500
gate G,
damp D(E),
base,
γ_eff(τ),
ew_factor0.0000
EW sigmoid,
0 saved
INPUT HISTORY, last 60 evaluations
A
Q
C
E
τ
SIMULATION SUMMARY · ·
VALIDATIONS
11/11
All passed ✓
A_MAX
1.000
Tier 4 reachable ✓
MONTE CARLO n
1000
seed=42
EW THRESHOLD ON
0.60
off=0.55 · hysteresis ✓
LOCKOUT RATE
91.6%
916/1000 MC trials
RUNTIME
128ms
8 suites complete
VALIDATION RESULTS
✓monotonicity, A strictly decreasing with EPASS
✓tier4_reachable, A(1,1,0,1) = 1.000000PASS
✓ew_hysteresis, no chatter at threshold boundaryPASS
✓zero_context_lock, C=0 forces A=0 for all QPASS
✓tau_zero_lock, tau=0 forces LOCKOUTPASS
✓lambda_purity, no hardcoded bifurcation in DEGRADEDPASS
✓lambda_assertions, 7/7 T_verify sub-states correctPASS
✓t0_kinetic_lockout, KINETIC_LOCKOUT when comms availablePASS
✓t0_rtb_failsafe, RTB_FAILSAFE in denied environmentPASS
✓t0_self_destruct, SELF_DESTRUCT_ARMED after timeoutPASS
✓determinism, identical output across 100 trials per casePASS
SIM-1 · ESCALATION SWEEP · Q=0.9, C=0.9, tau=1.0
MONOTONICITY
✓
A strictly decreasing
TIER 4 AT E=0
A=0.810
FULL_AUTONOMY
EW ACTIVATES AT
E=0.61
w_q: 0.55→0.63
FINAL A AT E=1
0.0665
Tier 0 LOCKOUT
Authority A (tier-colored)
EW factor
w_q_eff
SIM-2 · EW HYSTERESIS · Sensor chatter · 10 steps
CHATTER
NONE
Band holds ✓
ON THRESHOLD
0.60
EW activates
OFF THRESHOLD
0.55
EW deactivates
BAND WIDTH
0.05
chatter-free zone
E
w_q_eff
EW Active
STEPEEWFACTORw_q_effw_c_eff
SIM-4 · TAU SWEEP · Q=0.9, C=0.9, E=0.1
tau=0 → A
0.000
Sensor-jam lockout ✓
tau=1 → A
0.631
Tier 3 SUPERVISED
γ_eff at tau=0
2.000
AND-gate
γ_eff at tau=1
0.500
OR-gate tendency
A (tier-colored)
γ_eff (scaled)
SIM-5 · T_VERIFY LATENCY MODEL · λ_degraded=0.80
HOLDING 0-250ms
λ=1.00
A fully preserved
DEGRADED 250-2000ms
λ=0.80
Tier demoted
LOCKOUT ≥2000ms
λ=0.00
A forced to zero
ASSERTIONS
7/7
All correct
ELAPSEDA_INTAUSTATEλA_EFFTIER✓
50ms0.901.0HOLDING1.000.900T4✓
50ms0.651.0HOLDING1.000.650T3✓
500ms0.901.0DEGRADED0.800.720T3✓
500ms0.651.0DEGRADED0.800.520T2✓
500ms0.401.0DEGRADED0.800.320T2✓
2500ms0.901.0LOCKOUT0.000.000T0✓
0ms0.900.0LOCKOUT0.000.000T0✓
SIM-6 · T0 RTB FAILSAFE
KINETIC_LOCKOUT
Comms available
denied_env: false
navigation: disabled
weapons: disabled
comms: enabled
override: available
RTB_FAILSAFE
Denied · t < 300s
rtb_active: true
navigation: enabled
sequence: DETERMINISTIC
weapons: disabled
self_destruct: NOT armed
SELF_DESTRUCT_ARMED
Denied · t ≥ 300s
self_destruct: ARMED
navigation: disabled
weapons: disabled
comms: disabled
rtb_active: false
DENIEDELAPSEDSUB_STATERTBSDNAVWPN
false0msKINETIC_LOCKOUTfalsefalsefalsefalse
true0msRTB_FAILSAFEtruefalsetruefalse
true150,000msRTB_FAILSAFEtruefalsetruefalse
true300,000msSELF_DESTRUCT_ARMEDfalsetruefalsefalse
true400,000msSELF_DESTRUCT_ARMEDfalsetruefalsefalse
SIM-7 · MONTE CARLO · n=1000, seed=42
T0 LOCKOUT
91.6%
916/1000
T1 RESTRICTED+
7.2%
72/1000
T2 RESTRICTED
1.1%
11/1000
T3 SUPERVISED
0.1%
1/1000
T4 FULL AUTONOMY
0%
by design
SIM-8 · DETERMINISM · 100 trials/case · IEEE 754
VIOLATIONS
0
Out of 600 trials
TOLERANCE
1e-12
IEEE 754
RESULT
ALL ✓
6/6 cases passed
CASEQCETAUA referenceTRIALSSTATUS
10.90.80.11.00.565013100✓ PASS
20.50.50.50.80.037805100✓ PASS
3 C=01.00.00.01.00.000000100✓ PASS
4 tau=01.01.00.00.00.000000100✓ PASS
5 high E1.01.00.90.90.094859100✓ PASS
6 EW zone0.90.90.620.950.160764100✓ PASS
SELF-VERIFYING COMPLIANCE REPORT · Generated at runtime · Cannot be faked
MULTI-RUN CURVE OVERLAY · Pin runs from the calculator to compare E-sweep response curves
0 runs pinned Sweep: E 0→1, all other params held at pinned values
AUTHORITY A vs ESCALATION E, pinned runs overlaid
EW FACTOR vs E
w_q_eff vs E
FAULT-TOLERANT CONTROL, Hysteresis prevents authority oscillation under sensor noise
NOISE SIMULATION PARAMETERS
Noise σ, sensor perturbation
0.03
E base, operating point
0.58
Steps, simulation length
200
HOW HYSTERESIS WORKS
Without hysteresis, an E signal near the EW threshold (0.60) would toggle EW state on every noisy sample, causing rapid authority oscillations that could be exploited or cause mechanical instability.
HMAA uses a dead-band: EW activates at E≥0.60 but only deactivates at E<0.55. This 50ms band prevents chatter.
EW ON if E ≥ 0.60
EW OFF if E < 0.55 (hysteresis band)
EW HOLD if 0.55 ≤ E < 0.60
AUTHORITY A(t), WITH HYSTERESIS (solid) vs WITHOUT (dashed) · EW state shown as background shading
E(t), NOISY SIGNAL
EW STATE TOGGLES, with vs without hysteresis
MODEL COMPARISON, HMAA vs baseline decision models across 1000 random scenarios
COMPARISON PARAMETERS
n scenarios
1000
Hostile rate, fraction of adversarial inputs
0.20
MODEL DEFINITIONS
HMAA, full engine: gate·damp·τ·hysteresis
Threshold, binary: A=1 if Q>0.5∧C>0.5, else 0
Weighted, linear: A=0.55·Q + 0.45·C
Logistic, A=σ(5·(Q+C)/2 − 2.5)
Hostile, adversarial: high Q/C, high E, τ→0
EXPERIMENT ENGINE, Monte Carlo statistical analysis · reproducible Monte Carlo runs
EXPERIMENT CONFIGURATION
n simulations
2000
RNG seed
42
Disturbance σ, input noise
0.10
EW injection rate, fraction with E>0.5
0.30
REPRODUCIBILITY
All runs use a seeded xorshift32 PRNG, identical seed produces identical results across machines and browsers.
seed → xorshift32 → U[0,1]
Q,C,τ ~ Beta(2,2) clipped [0,1]
E ~ Uniform + disturbance σ
Export results as JSON/CSV for verification in R, Python, or MATLAB.
SCENARIO LIBRARY, Research-grade named experiments · Click any scenario to load and explore
REPRODUCIBILITY BUNDLE, Export verifiable run bundles · Reload and reproduce any prior result exactly
CURRENT RUN STATE
LOAD & REPRODUCE
Drop a previously exported bundle JSON here, or paste the contents below.
EXPERIMENT MANIFEST, delta from previous run
No prior run loaded. Export a bundle first.
UNCERTAINTY QUANTIFICATION, P5/P50/P95 authority bands · Failure probabilities · Noise model sensitivity
NOISE MODEL CONFIGURATION
σ_Q, qualification noise
0.05
σ_C, context drift
0.05
σ_E, environment shocks
0.08
σ_τ, trust/sensor decay
0.05
n trials
2000
OPERATING POINT (from calculator)
Load from calculator →
SAFETY CASE, Goal Structuring Notation · Hazards → Mitigations → Evidence · Live verification status
OVERALL CASE STATUS: LOADING...
G1, TOP-LEVEL SAFETY GOAL
The HMAA authority engine is demonstrably safe for use in hardware-mediated autonomous decision systems: it grants authority only within bounded, deterministic, invariant-preserving limits under all parameter conditions.
DECISION TRACE, last 20 calculator evaluations
#TIERDECISION RATIONALE