Entropis Benchmark Suite
Public Benchmark Record
Version 2.2 | January 2026
Public Standard: Metrics are defined prior to measurement, reported with pass/fail criteria, with protected technical material held inside governed review.
Why New Benchmarks?
Existing AI benchmarks (MMLU, HumanEval, MLPerf) measure trained systems on static tasks. They are not designed to measure persistent dynamics, self-organization, or embodiment-dependent behavior.
The Entropis Benchmark Suite measures emergent properties, self-organization, and embodiment dependencies.
What This Document Provides
- ✓ Measurement summary
- ✓ Pass/fail criteria
- ✓ Scientific basis
- ✓ Published results
What This Document Does NOT Provide
- ◇ Implementation details
- ◇ Architecture specifications
- ◇ Source code
- ◇ Proprietary algorithms
Terminology
Public benchmark terms used throughout this benchmark suite:
Target Operating Metric
A controlled activity measure used to verify whether the system enters the target operating regime associated with high-information dynamics.
Response Variability
Internal variability measure used to distinguish state-dependent response behavior from deterministic output.
Target Regime
A bounded operating range used in the public validation protocol to distinguish stable activity from inactive or overloaded behavior.
Embodied Condition
Controlled embodied condition used to test whether system activity can be sustained under contrast.
Hardware Invariance
Comparable benchmark behavior on fundamentally different hardware architectures. Reduces the risk that results are platform-specific artifacts.
Structural Adaptation
Measured structural change during learning. Technical mechanism remains in controlled review.
Target Operating Regime
Entropis is reported inside a target operating regime associated with stable, high-information system behavior. Detailed metric construction remains in controlled review.
Comparison to Industry Benchmarks
Industry benchmarks test trained models on static tasks. ENT benchmarks test untrained systems on emergent capabilities.
| Industry Standard | What It Tests | Entropis Equivalent |
|---|---|---|
| MLPerf | Training/inference speed | ENT-SPEED |
| MMLU | Language model knowledge | ENT-IQ5 (emergence) |
| HumanEval | Code generation | ENT-EM5 (embodiment) |
| Mismatch Negativity (EEG) | Novelty detection in brains | ENT-NOVELTY |
| Adaptive learning theory | Adaptive threshold behavior | ENT-ADAPTIVE (L4) |
Speed Benchmark
PURPOSE
Measures processing throughput relative to biological baseline behavior.
METRIC
BASELINE
Biological baseline comparison under the public benchmark protocol.
PASS CRITERIA
Throughput exceeds biological baseline under protocol.
RESULTS
| Hardware Class | Scale Class | Throughput | Measurement | Result |
|---|---|---|---|---|
| Consumer GPU | 2B-class | >1000× biological | Internal metric | PASS |
| Consumer GPU | 470M-class | >10× biological | Internal metric | PASS |
| Mobile-class SoC | 5M-class | >1× biological | Internal metric | PASS |
Measurement Summary
Aggregate Measure
Conservative aggregate throughput measure used for public reporting.
Activity Measure
Activity-normalized measure retained in the internal validation record.
Public reporting preserves aggregate outcomes while routing detailed computation to controlled review.
Intelligence Quotient (5 Markers)
PURPOSE
Validates emergent intelligence through 5 biological markers. These markers distinguish brain-like systems from calculators and are grounded in neuroscience literature.
Adaptive Variability
What it measures: Repeated conditions produce state-dependent response variability.
Metric: Internal response-variability measure
Pass: Variability exceeds deterministic threshold
Fail: Response remains effectively deterministic
Result: Variability threshold exceeded (PASS)
Target Dynamics
What it measures: Self-organization into the target operating regime.
Metric: Controlled activity ratio used in the validation protocol
Scientific basis: High-information operating regimes in biological systems.
Pass: Activity enters target range without explicit targeting
Fail: Activity remains fixed OR never enters target range
Result: Target range reached (PASS)
Cascade Distribution
What it measures: Broad distribution of internal activity cascades.
Metric: Internal cascade-distribution measure
Scientific basis: Biological systems exhibit broad activity distributions.
Pass: Broad cascade distribution observed
Fail: Activity remains uniform or non-cascade-like
Result: Cascade-distribution threshold exceeded (PASS)
Bidirectional Learning
What it measures: Both habituation (decreased response) AND sensitization (increased response).
Metric: Internal response-change measure
Scientific basis: Biological brains show both directions; direction depends on context.
Pass: Both positive and negative adaptation observed
Fail: Only one direction, OR no adaptation
Result: Both directions observed (PASS)
Emergent Behavior
What it measures: Behaviors arise from local rules, not explicit programming.
Metric: Presence of the marker set without explicit targeting
Pass: All markers emerge without hardcoded values
Fail: Any marker achieved through explicit programming
Result: All markers emergent (PASS)
Embodiment (5 Markers)
PURPOSE
Validates closed-loop embodiment. A synthetic brain must process inputs, maintain internal dynamics under load, and produce controlled output.
Maintains target dynamics under load
Visual input processing under protocol
Audio input processing under protocol
Controlled output generation
Closed-loop embodiment under feedback
RESULTS
| Marker | Windows | Mac |
|---|---|---|
| EM5-BRAIN | PASS | PASS |
| EM5-VIS | PASS | PASS |
| EM5-AUD | PASS | PASS |
| EM5-MOT | PASS | PASS |
| EM5-LOOP | PASS | PASS |
Embodiment Contrast Benchmark
PURPOSE
Tests whether an embodied operating condition is required to maintain resting activity under controlled conditions.
METHODOLOGY
Brain receives control input condition
Expected: DORMANT
Brain receives embodied condition
Expected: ALIVE (target dynamics maintained)
RESULTS
| Metric | Windows | Mac |
|---|---|---|
| INTER-0 (Silent) | DORMANT | DORMANT |
| INTER-1 (Embodied) | ALIVE | ALIVE |
| Target-Regime Stability | >95% | >95% |
| Target Dynamics Maintained | PASS | PASS |
“Embodied operating conditions materially change resting-state dynamics.”
Validated on both platforms.
Novelty Detection Benchmark
NEWPURPOSE
Validates that the brain can distinguish familiar from novel stimuli, demonstrate habituation to repeated input, and exhibit memory through faster recovery. This supports functional information processing beyond dynamical signatures alone.
SCIENTIFIC BASIS
Based on the Mismatch Negativity (MMN) paradigm from cognitive neuroscience — a gold-standard test for pre-attentive processing, working memory, and novelty detection in biological brains.
MARKERS (3)
NOV-HAB
Habituation
Brain adapts to repeated stimulus. Shows learning over time.
NOV-DET
Novelty Detection
Brain responds differently to new vs. familiar stimuli.
NOV-REC
Memory & Recovery
Brain remembers familiar stimuli. Faster re-stabilization.
RESULTS
| Marker | Hardware Class A | Mac (M4) |
|---|---|---|
| NOV-HAB | PASS | PASS |
| NOV-DET | PASS | PASS |
| NOV-REC | PASS | PASS |
Demonstrates functional information processing beyond dynamical signatures.
Association Learning
PURPOSE
Validates classical conditioning — the brain's ability to learn that stimulus X predicts stimulus Y. This is Pavlovian learning, the foundation of all associative reasoning.
PROTOCOL
| Phase | Protocol | Expected |
|---|---|---|
| Phase A: Pairing | Paired stimulus condition repeated | Association learned |
| Phase B: Test | Partial cue condition | Anticipatory response |
| Phase C: Control | Unpaired control condition | No anticipation (baseline) |
MARKERS (3/3)
| Marker | Description | Windows | Mac |
|---|---|---|---|
| ASSOC-RESP | Stimulus response | ✓ PASS | ✓ PASS |
| ASSOC-ANTIC | Anticipatory response | ✓ PASS | ✓ PASS |
| ASSOC-SPEC | Response specificity | ✓ PASS | ✓ PASS |
Foundation for associative reasoning across chained relationships.
Expectation-Response Behavior
PURPOSE
Validates temporal sequence learning and expectation-response behavior: the brain learns sequences and generates expectation responses. This supports language-comprehension behavior.
PROTOCOL
| Phase | Sequence | Expected |
|---|---|---|
| Phase A: Learning | Ordered sequence repeated | Sequence response learned |
| Phase B: Omission | Expected element withheld | Expectation response observed |
| Phase C: Violation | Unexpected element substituted | Violation response observed |
MARKERS (3/3)
| Marker | Description | Windows | Mac |
|---|---|---|---|
| SEQ-ENC | Sequence encoding | ✓ PASS | ✓ PASS |
| SEQ-PRED | Predictive activity | ✓ PASS | ✓ PASS |
| SEQ-SURP | Surprise response | ✓ PASS | ✓ PASS |
Activity for omitted element indicates internal prediction model.
Cross-Platform Invariance
PURPOSE
Validates that comparable benchmark behavior appears on different hardware architectures. This reduces platform-specific optimization as an explanation for the results.
PLATFORMS TESTED
| Component | Platform A | Platform B |
|---|---|---|
| Hardware class | Consumer GPU class | Mobile SoC class |
| Execution surface | Class A | Class B |
| CPU | Intel x86 | Apple ARM |
| Memory | Discrete (PCIe) | Unified (SoC) |
| OS | Windows | macOS |
| Neurons | 470,000,000 | 5,000,000 |
Cochlea & Auditory Pathway
PURPOSE
Validates audio processing behavior under the public benchmark protocol.
MARKERS (6)
AUD-FREQ
Audio Analysis
Audio feature response
AUD-SPATIAL
Spatial Organization
Biological sound mapping
AUD-ADAPT
Temporal Adaptation
Dynamic response adjustment
AUD-TIMING
Temporal Precision
Accurate timing processing
AUD-ONSET
Onset Detection
Transient sound detection
AUD-OFFSET
Offset Detection
Sound termination detection
All 6 markers validated on both platforms
Efficient Neural Processing
PURPOSE
Validates efficient activity-dependent processing under controlled benchmark conditions.
MARKERS (6)
SPARSE-SCALE
Efficient Scaling
Processing scales with activity level
SPARSE-CACHE
Cache Efficiency
Optimized memory access patterns
SPARSE-SPONT
Spontaneous Activity
Resting activity present
SPARSE-HIST
Activity Tracking
State history maintenance
SPARSE-HOME
Homeostatic Balance
Self-regulating activity levels
SPARSE-EVENT
Event Processing
Efficient event-driven computation
All 6 markers validated on both platforms
Complete Language Integration
PURPOSE
Validates language processing through the controlled benchmark protocol. This is NOT regex or pattern matching — it's neural language processing with emergent semantic representation.
MARKERS (8)
LANG-INPUT
Language Input
Text comprehension active
LANG-REC
Receptive Processing
Language input processing active
LANG-SPREAD
Activity Spread
Information propagation
LANG-SEMANTIC
Semantic Clustering
Conceptual organization
LANG-EXP
Expressive Processing
Language output processing active
LANG-OUTPUT
Language Output
Coherent language generation
LANG-EMBODIED
Embodied Language
Language affects body state
LANG-LOOP
Full Loop
Complete processing cycle
Synaptic weights change based on usage. Learning through structural modification.
Full Cognitive Integration
PURPOSE
Validates the complete embodied cognitive loop across perception, cognition, language, action, and feedback. All systems running in parallel, maintaining target dynamics under load, like a biological brain.
MARKERS (7)
EMB-PERCEPT
Perception
Visual + auditory input processing
EMB-COGNIT
Cognition
Internal state maintenance
EMB-LANG
Language
Semantic processing active
EMB-MOTOR
Action
Controlled output generation
EMB-INTER
Embodiment
Embodied condition sustaining activity
EMB-CRIT
Target Dynamics
Target range maintained under load
EMB-STABLE
Stability
Long-term operation without collapse
7/7 markers validated
All systems running simultaneously. Parallel processing like biological brains.
Parallel Brain Systems (3 Markers)
PURPOSE
Validates that multiple cognitive systems operate simultaneously without interference — like a biological brain processing vision, hearing, language, and motor control in parallel.
MARKERS (3)
PAR-SIMUL
Simultaneous Operation
All systems active at once
PAR-INDEP
Independence
Systems don't block each other
PAR-CLEAN
No Interference
Cross-system crosstalk minimal
3/3 markers validated
Visual, auditory, language, and motor systems running concurrently.
Billion-Neuron Processing (8 Markers)
PURPOSE
Validates that the system scales to billions-class substrate size while maintaining biological properties. Tests parallel processing capacity, throughput, and health metrics at unprecedented scale.
MARKERS (8)
SCALE-INJ
Throughput Rate
High-throughput activity under protocol
SCALE-PROC
Processing Rate
Sustained processing throughput
SCALE-HEALTH
System Health
Target dynamics maintained at scale
SCALE-CASC
Cascade Dynamics
Proper activity amplification
SCALE-ACTIVE
Active Population
Appropriate firing rates
SCALE-WORK
Working Set
Efficient memory management
SCALE-MEM
Memory Access
Optimized data transfer
SCALE-EPOCH
Epoch Handling
Stable long-duration operation
RESULTS (2B Brain)
| Metric | Measurement | Result |
|---|---|---|
| Neurological Health | Target-regime stability | >90% |
| All 8 Markers | Pass/Total | 8/8 PASS |
8/8 markers validated at 2B-class scale
Reported internal validation at 2B-class scale on accessible hardware.
Structural Plasticity (5 Markers)
PURPOSE
Validates biological learning cycle — new connections form during activity, consolidate over time. This supports learning without training loops.
PLAST-FORM
Synapse Formation
New connections form based on correlated activity
PLAST-STRENGTH
Connection Strengthening
Activity-linked strengthening observed over time
PLAST-HOME
Homeostatic Regulation
Self-regulation maintains stable activity levels
PLAST-PRUNE
Synaptic Pruning
Unused connections removed during consolidation
PLAST-PHYS
Physics-Based Formation
Emergent principles drive structural adaptation
RESULTS
| Marker | Measurement | Result |
|---|---|---|
| PLAST-FORM | New structural adaptations formed during learning | PASS |
| PLAST-STRENGTH | Weight changes observed | PASS |
| PLAST-HOME | Target dynamics maintained during learning | PASS |
| PLAST-PRUNE | Weak connections removed | PASS |
| PLAST-PHYS | Emergent formation dynamics | PASS |
5/5 markers validated
Learning through structural modification. No gradient descent. No backpropagation.
Physics-Based Learning (20 Markers)
NEWPURPOSE
Validates cognitive learning through physics alone. No rewards. No labels. No backpropagation. All learning emerges from exposure and structural adaptation.
LEVELS (5)
Pattern discrimination, sequence learning, habituation, association.
Repetition suppression, oddball detection, rule extraction, interval learning.
Persistence via recurrence, recognition, sequence encoding, interference resistance.
Reversal, extinction, context-dependent processing, adaptive threshold behavior.
Structural generalization, central-category formation, compositional binding, temporal abstraction.
DETAILED RESULTS
| Level | Key Measurement | Result | Structural Learning |
|---|---|---|---|
| L1 Perceptual | Pattern discrimination | PASS | Significant |
| L2 Relational | Same/different detection | PASS | Significant |
| L3 Working Memory | Persistence windows | PASS | Significant |
| L4 Adaptive | Adaptive threshold behavior | PASS | Minimal |
| L5 Transfer | Unseen central category | PASS | Minimal |
| Total | 20/20 PASS | Observed |
20/20 cognitive markers validated
Structural learning observed across all levels
Learning through exposure and structural adaptation. No rewards. No labels.
Key Discoveries from Cognitive Tests
ENT-TRANSFER / L5
Central Category Formation
Exposed brain to related exemplars around an unseen central category.
The central category was never shown during exposure.
Response to unseen central category
TARGET
Emergent statistical learning. The brain formed the central tendency without explicit training.
ENT-ADAPTIVE / L4
Adaptive Threshold Behavior
Plasticity rate depends on activity history.
High activity → reduced subsequent plasticity.
Baseline
High
Saturated
Reduced
Rested
Recovered
Significant reduction after saturation under the reported protocol.
ENT-SEQUENCE
Predictive Processing
Trained on an ordered sequence.
Presented a withheld-element condition.
Activity under withheld-element condition
Expectation response
The brain predicted an element that was never displayed.
ENT-INTER
Embodiment Dependency
Compared brain with and without the embodied operating condition.
No embodied condition
DORMANT
With embodied condition
TARGET
Embodied operating conditions sustain activity.
Complete Validation Score
| Benchmark | Markers | Windows | Mac |
|---|---|---|---|
| ENT-SPEED | 3 processing speed markers | 3/3 | 3/3 |
| ENT-IQ5 | 5 intelligence markers | 5/5 | 5/5 |
| ENT-EM5 | 5 embodiment markers | 5/5 | 5/5 |
| ENT-INTER | 5 embodiment-contrast markers | 5/5 | 5/5 |
| ENT-ASSOC | 3 association learning markers | 3/3 | 3/3 |
| ENT-SEQUENCE | 3 expectation-response markers | 3/3 | 3/3 |
| ENT-NOVELTY | 3 memory/recognition markers | 3/3 | 3/3 |
| ENT-AUDIO | 6 auditory processing markers | 6/6 | 6/6 |
| ENT-SPARSE | 6 efficient processing markers | 6/6 | 6/6 |
| ENT-LANGUAGE | 8 language integration markers | 8/8 | 8/8 |
| ENT-EMBODIED | 7 cognitive integration markers | 7/7 | 7/7 |
| ENT-PARALLEL | 3 parallel processing markers | 3/3 | 3/3 |
| ENT-SCALE | 8 billion-neuron markers | 8/8 | N/A (2B only) |
| ENT-PLASTICITY | 5 structural adaptation markers | 5/5 | 5/5 |
| ENT-XPLAT | 11 hardware invariance markers | 11/11 cross-validated | |
| ENT-PERCEPT | 4 perceptual learning markers (L1) | 4/4 | 4/4 |
| ENT-RELATION | 4 relational learning markers (L2) | 4/4 | 4/4 |
| ENT-WORKING | 4 working memory markers (L3) | 4/4 | 4/4 |
| ENT-ADAPT | 4 adaptive behavior markers (L4) | 4/4 | 4/4 |
| ENT-TRANSFER | 4 generalization markers (L5) | 4/4 | 4/4 |
| ENT-TOTAL | All validation markers | 101/101 | 101/101 |
101/101
Three scales validated: 5M → 470M → 2B neurons
Including 20/20 physics-based cognitive tests (L1-L5)
Same benchmark family across materially different scales.
Neurons
2B-class
Speed
Faster than biological
Markers
101/101 PASS
20 benchmark categories • 5 cognitive test levels (L1-L5) • accessible hardware class
Falsification Record
Defined falsification criteria with observed results.
| Claim | Falsification Condition | Actual Result | Status |
|---|---|---|---|
| Faster than brain | Biological baseline not exceeded | Baseline exceeded | Not falsified |
| Self-organization | Target range never reached | Target range reached | Not falsified |
| Non-deterministic | Deterministic output | Variability observed | Not falsified |
| Bidirectional learning | One direction only | Both directions | Not falsified |
| Hardware invariant | One platform fails | Both 101/101 | Not falsified |
| No reward/backprop training loop | Reward, label, or backprop loop used | Not used in reported protocol | Not falsified |
| Embodiment required | Active with zero input | Goes dormant | Not falsified |
| Active sensors | Any sensor = 0 | All > 0 | Not falsified |
8 falsification tests. 0 falsified.
Tested across a large scale difference and two materially different hardware classes.
Benchmark Suite Evolution
Development history of the Entropis Benchmark Suite:
Each version adds markers while maintaining backward compatibility with prior validations.
Verify Live
Request a live demonstration to observe benchmark tests in real-time.
Request Demo→Governed access for live demonstration.