Looking for a website audit against ADA Title II / WCAG 2.1 AA for municipalities or commercial portals? Go to the AIfaFocus accessibility scanner.
AIfa Cognitive Runtime (ACR)
A connectome-grounded cognitive runtime for autonomous software agents, derived from the FlyWire v783 Drosophila whole-brain connectome (139,255 neurons, 54.5M synapses). Five biological computation motifs implemented on CPU (Python/NumPy), no GPU; every number on this page comes from an open script.
Reference substrate: FlyWire FAFB v783 — 139,255 neurons, 54.5M synapses. Dual-mode architecture: Faithful Connectome Mode (topological integrity, synapse graphs & SHA-256 hash) · Distilled ACR Mode (CPU algorithms for software agents; the connectome is used for graph statistics).
Ablation Matrix — real measurement (360 episodes per configuration, 23.09.2026)
| Agent / configuration | Distractors | Task success | Steps when solved | Wrong clicks | Goal lost / click | Decision P50 |
|---|---|---|---|---|---|---|
| Naive agent (goal falls out of a 4-message window) | 35% | 0.3% ± 1.5 | 10.0 | 98.6% | 0.000 | 0.074 ms |
| Standard agent, goal pinned (system-prompt style) | 35% | 5.8% ± 3.8 | 5.71 | 97.9% | 0.000 | 0.076 ms |
| + APL novelty gate only | 35% | 4.2% ± 4.2 | 3.27 | 97.7% | 0.000 | 0.076 ms |
| + CANN focus ring only | 35% | 5.6% ± 3.9 | 5.45 | 97.9% | 0.025 | 0.097 ms |
| + APL + CANN | 35% | 1.4% ± 3.1 | 6.20 | 99.2% | 0.001 | 0.100 ms |
| + APL + CANN + CX (as shipped in aifa_sdk) | 35% | 1.4% ± 3.1 | 6.20 | 99.2% | 0.001 | 0.181 ms |
| Standard agent + visited-link memory only | 35% | 95.0% ± 5.5 | 6.16 | 53.8% | 0.000 | 0.218 ms |
| + APL + CANN + CX fixed (visited-link repulsion) | 35% | 94.7% ± 5.5 | 6.23 | 54.5% | 0.000 | 0.427 ms |
| Full ACR (fixed CX) + bilateral verifier | 35% | 93.6% ± 4.1 | 6.07 | 54.1% | 0.000 | 0.680 ms |
* Synthetic seeded websites (12 tasks × 30 seeds), lexical word-overlap embeddings. 35% distractor messages, duplicate cookie banners, decoys and loop links — identical for every agent. No outcome is drawn from a fixed probability: the code decides. Script: bench/acr_agent_real_benchmark.py. Finding: the whole gain comes from visited-link memory, which the shipped CX module lacks; a standard agent with that memory alone matches ACR.
Measured latencies (Intel i7-14700 CPU, Python/NumPy)
N = 512, measured 21.09.2026 on the aifa_sdk engine.
P50, 2,000 trials, measured 21.09.2026.
P50; the code has no loop avoidance (see card 03).
FlyHash P50, 23.09.2026. Exact FAISS: 8.8 ms.
Runs on any laptop without a GPU. Source code is in the aifa-biobench repository and on the /digital page: aifa-biobench · /digital
What works and by how much: four real measurements
1. Binary retrieval duel (N = 25,000, D = 512)
100 queries, seed 42A direct duel with classic binarisation. FlyHash uses 85× fewer non-zero projection weights (2048×6 vs 2048×512), but it loses on accuracy: 18.9% vs 80.5% for Sign-LSH at the same latency.
| Method | Recall@10 | NDCG@10 | P50 | Projection weights |
|---|---|---|---|---|
| AIfa FlyHash | 18.90% | 0.3188 | 35.46 ms | 12,288 (85× lighter) |
| Sign-LSH (BQ 1-bit) | 80.50% | 0.8715 | 36.77 ms | 1,048,576 |
| Multi-table LSH | 0.10% | 0.0022 | 0.045 ms | — |
2. Input 'claw' count sweep (d from 2 to 16)
hypothesis testTesting 'the fly's d = 6 is optimal'. Measured: 19.8% at d = 6, already 23.3% at d = 7, the highest recall is 24.9% at d = 16. On this task d = 6 is not the optimum.
| Claw count (d) | Recall@10 | P50 | Status |
|---|---|---|---|
| d = 2 | 13.80% | 29.35 ms | baseline |
| d = 6 | 19.80% | 29.98 ms | fly value |
| d = 7 | 23.30% | 28.37 ms | higher |
| d = 16 | 24.90% | 30.21 ms | highest (2.7× more weights) |
3. Agent benchmark — real engines
360 episodes per agentReplaces the old script whose outcome was set by fixed probabilities (22.47% → 94.58%). Real aifa_sdk engines, synthetic websites, 35% distractors. Result: the gain comes from visited-link memory; ACR's other mechanisms add no measurable benefit here.
| Agent | Task success | Wrong clicks |
|---|---|---|
| Standard, goal pinned | 5.8% ± 3.8 | 97.9% |
| ACR: APL + CANN + CX as shipped | 1.4% ± 3.1 | 99.2% |
| ACR with fixed CX | 94.7% ± 5.5 | 54.5% |
| Full ACR + verifier | 93.6% ± 4.1 | 54.1% |
| Standard + visited memory | 95.0% ± 5.5 | 53.8% |
4. Robustness — real dropout
5 seedsReplaces the old formula-generated curve ('3.21× more robust'). Code bits and ring neurons are really switched off. FlyHash loses more accuracy than Sign-LSH; the CANN ring loses precision as neurons die.
| Dropped | FlyHash R@10 | Sign-LSH R@10 | CANN idle error |
|---|---|---|---|
| 0% | 58.9% | 86.1% | 0.13° |
| 10% | 52.3% | 84.5% | 7.0° |
| 30% | 39.6% | 82.2% | 15.2° |
| 50% | 25.4% | 78.2% | 14.6° |
- A set of bio-inspired algorithms (sparse projection, APL novelty gate, CANN ring attractor, CX steering, bilateral verifier) in Python/NumPy in the aifa_sdk package.
- Some mechanisms work as described (APL sparsity, CANN goal holding without distractors); some do not (CX loop avoidance, the WASM build, the deadlock check).
- Every number on this page comes from scripts in the aifa-biobench repository; result files are open.
- Not a simulation of a fly brain and not a claim of biological equivalence.
- The algorithms use random connections, not FlyWire weights; the FlyWire v783 connectome is used only for graph statistics (C, L, KS).
- No '58 μs full-loop' or '~122 ms live DOM' run exists — those numbers were removed.
No headline metric is presented without its evidence class
Biological hypothesis
Analytical complexity
Simulator / partitioner
Reproducible CPU test
Independent hardware
Production cluster
TOP-10 Implemented Connectome Stack (L0–L4 Core Production Candidates · 06–10 Research Prototypes)
Core L0–L4 cognitive pipeline and experimental research prototypes 06–10 classified under E0–E5 evidence tiers
FlyHash v783 Connectome Memory
Inspired by the fly mushroom body (Kenyon cells, ~6 input 'claws'). Projection weights are random and sparse — FlyWire data is not loaded into this mechanism.
APL Sensory Novelty Gate
Giant GABAergic Anterior Paired Lateral (APL) neuron delivering global feedback inhibition.
Central Complex CX Steering Navigation
Fan-Shaped Body (FB) and Protocerebral Bridge (PB) of the Central Complex.
CANN Focus Ring Attractor
Ellipsoid Body ring neurons (64 neurons) with continuous attractor dynamics.
Bilateral Cross-Inhibition Verifier
Lateral cross-inhibition arbitration between symmetric brain hemispheres.
Neuromorphic Energy Model (LIF Engine)
Biophysical energy model based on Drosophila event-driven dynamics under stated event-cost assumption (1 pJ/spike vs 1–3 pJ/FLOP on GPU FP16). Sparsity breakdown: Connectivity Sparsity = 85.0% structural zeros; Activity Sparsity = 96.86% quiescent neurons during benchmark execution. Physical RAPL/wattmeter wall-clock hardware testing is on roadmap.
AIfa BioMatch Score (BioMatch v1.0)
Metrological profile of the Drosophila melanogaster connectome (FlyWire v783: 139,255 neurons, 54.5M synapses, ~2.7M thresholded graph edges; Nature, Schlegel et al. 2024). Evaluates topological isomorphism of artificial memory graphs against biological reference.
Browser WASM Search (Client Microbenchmark)
Client-side search in the browser: plain JavaScript (TypedArrays, 7.1 KB), inverted index with Jaccard similarity. No WebAssembly or SIMD is used.
Neuromorphic Graph Compiler Prototype (Loihi 2 / SynSense Target Exporter)
Software prototype: converts a connection graph into an address-event (AER) routing table and quantises weights to INT8. Core partitioning is plain consecutive blocks — no Metis, no min-cut; the 78 neuropils are not used in the code. The 'deadlock-free' check was a constant True. No Loihi 2 board telemetry has been run.
Human-AI Symbiosis Index (Mathematical Metric Φ_sym)
Formalized scalar model evaluating operator-agent alignment derived from connectome heterosynaptic plasticity principles. To strictly eliminate negative divergence values, uses bounded exponential transformation: Φ_sym = exp(-lambda * D_KL(P_intent || P_action)) * Alignment * Trust. Formalization of P_intent is established; empirical criterion validity study with human operators is pending.
Retrieval, agent and robustness — measured separately
Run: python bench/aifa_biobench.py · python bench/acr_agent_real_benchmark.py · python bench/robustness_real.pySearch over 50,000 vectors (1024d)
Compared against exact L2 brute force.
Navigation on synthetic websites
9 agents × 360 episodes, 35% distractors.
Switching off bits and neurons
5 seeds, 20,000 vectors, 512 bits.
The Complete 30 Connectome Innovations Catalog
ACR Connectome Research Catalog derived from the FlyWire v783 connectome with biological basis, mathematical models, and deployment targets
Connectome Innovation 1
Inspired by the fly mushroom body. From the literature: ~783 projection neurons pass odour signals to ~2,000–2,500 Kenyon cells, each cell receives ~6–8 random inputs, the APL neuron inhibits them all, and ~5% of cells stay active. The idea: a random sparse projection into a large space plus winner-take-all gives a noise-tolerant binary hash. Our implementation projects 1024 → 2048 units with random weights (FlyWire data is not loaded), keeps 30% of units active, shortlists by Hamming distance and reranks exactly.
Connectome Innovation 2
Inspired by the fly APL neuron: from the literature, one large inhibitory neuron per hemisphere collects Kenyon-cell activity and damps it when a stimulus resembles one already seen. Our implementation is a duplicate detector: a new vector is compared with a buffer of seen ones, and above a similarity threshold it is treated as a duplicate. Measured 21.09.2026 (bench/novelty_detector.py): buffer 5,000, 4,000 checks (2,000 new + 2,000 duplicates), 100% accuracy on this synthetic test, P50 1,069 μs, P95 2,092 μs. The earlier '3.4 μs' was not confirmed.
Connectome Innovation 3
Inspired by the fly central complex: from the literature, a ring of ellipsoid-body neurons holds the current heading and the fan-shaped body compares it with the goal and issues a turn command. Our implementation picks the next page element by the vector towards the goal instead of stepping through with the Tab key.
Connectome Innovation 4
FlyWire v783 connectome (Zenodo 10.5281/zenodo.10676866, CC BY 4.0). Measured 23.09.2026 from the release files (bench/connectome_real_metrics.py): 139,255 neurons, 54,492,922 synapses, 15,091,983 connected pairs, 2,700,513 pairs with ≥5 synapses; transmitters in the data: acetylcholine, GABA, glutamate, octopamine, serotonin, dopamine. The earlier '3,869,878 connections' was not confirmed. File fingerprints: Zenodo md5 (f48f972d… and e0e6c197…). The Merkle tree in bench/proof_of_connectome.py is built on synthetic records with the same neuron count: build 1.23 s, subtree check 1.6 ms, a forged leaf caught in 20 of 20 attempts.
Connectome Innovation 5
The idea: look at a knowledge graph the way neuroscientists look at a connectome — node degrees, clusters, hubs, short paths. The real FlyWire v783 connectome (measured 23.09.2026, 2,700,513 pairs with ≥5 synapses): clustering coefficient C = 0.160, mean path length L = 4.03. The earlier 'C = 0.34' and '2.3% hubs' were not confirmed. In our implementation the nodes are organisations, domains, phones and technologies, and the edges are links between them.
Connectome Innovation 6
According to the literature, a fly brain uses on the order of 10 microwatts, and only a small fraction of neurons is active at any moment. The model's idea: count energy by the number of events (spikes) instead of by all multiplications, as in a dense neural network. This is a calculated model, not a measurement: no wattmeter or RAPL counters were used.
AIfa BioMatch Score (BioMatch v1.0)
Architectural prototype: Metrological profile of the Drosophila melanogaster connectome (FlyWire v783: 139,255 neurons, 54.5M synapses, ~2.7M thresholded graph edges; Nature, Schlegel et al. 2024). Evaluates topological isomorphism of artificial memory graphs against the biological reference.
Browser Client-Side Search
Architectural prototype: client-side associative search engine (public/aifa_connectome_web.js). 21.09.2026: card previously claimed 'compact WebAssembly binary (Wasm SIMD128, 126.7 KB)'. DIRECT FILE INSPECTION found: pure JavaScript, zero WASM bytes, zero SIMD instructions — the file's own comment honestly says 'pure JavaScript / TypedArrays'. Actual file size is 7,121 bytes, not 126,700 (17.79x discrepancy). Real algorithm is an inverted posting-list index with Kenyon Cell activation and Jaccard similarity, not popcount over bit-packed hashes.
Neuromorphic Graph Compiler Prototype (Loihi 2 / SynSense Target Exporter)
Architectural prototype: Software prototype translating connectome synaptic graphs into asynchronous Address Event Representation (AER). Metis-based graph partitioning across 78 neuropils, INT8 synaptic quantization, and wormhole routing simulation. 21.09.2026: the claim '0 deadlocks across 1,000 trials' was checked — the underlying engine returns is_deadlock_free as a HARDCODED constant True, never measured. The claim was unverifiable by construction. An honest replacement (bench/neuromorphic_compiler_run.py) searches for cycles in the inter-core NoC dependency graph across 1,000 synthetic graphs: 1,000/1,000 runs found a cyclic dependency (expected for dense random graphs), which does NOT prove real hardware deadlocks but honestly shows the original 0/1000 number was never computable.
Human-AI Symbiosis Index (Mathematical Metric Φ_sym)
Architectural prototype: Formalized scalar model evaluating operator-agent alignment derived from connectome heterosynaptic plasticity principles. Phi_sym = exp(-lambda * D_KL(P_intent || P_action)) * Alignment * Trust. 21.09.2026: the variables Alignment and Trust were never defined numerically on this card. Found a pre-existing engine (aifa-biobench/aifa_sdk/symbiosis_index.py) implementing a DIFFERENT, fully-defined 4-factor formula. Honestly run (bench/symbiosis_index_run.py) over 1,000 synthetic interaction turns — mean Phi=0.1547, P50=0.0974, per-turn latency 49.7-54.7 μs (not 39.7 μs, but same order of magnitude). This is an honest synthetic test, not an empirical study with live operators.
Connectome Innovation 11
Measured 23.09.2026 on the real FlyWire v783 connectome (bench/connectome_topology_extra.py; 134,181 connected neurons, 2,511,789 pairs with ≥5 synapses): clustering C = 0.155, mean path L = 4.02. A random graph of the same size, measured the same way, has C = 0.00029 and L = 3.65. Small-world index σ = (C/C_rand)/(L/L_rand) = 490 — the fly brain is highly clustered with short paths. The card's earlier figures (C = 0.284, L = 3.82, σ = 8.42 or 79.81) were not confirmed. For a synthetic memory graph (500 nodes, p = 0.08) the 21.09.2026 run gave σ = 19.7, not the claimed 7.15.
Connectome Innovation 12
Measured 23.09.2026 on the real FlyWire v783 connectome (bench/connectome_topology_extra.py, 134,181 connected neurons): ⟨k²⟩/⟨k⟩ = 190.9, hence the random-failure breakdown threshold f_c = 1 − 1/(⟨k²⟩/⟨k⟩ − 1) = 0.995. Direct test: after removing 30% of neurons at random the largest connected part still holds 68.5% of all neurons in the graph; after removing the 30% most connected — 57.5%; at 50% — 48.1% and 28.0%. The earlier '⟨k²⟩/⟨k⟩ ≈ 42.6, f_c ≈ 0.976' was not confirmed.
Connectome Innovation 13
Inspired by fly olfaction: according to the literature, the first 'dangerous / edible' decision is made by hard-wired receptors, without lengthy computation. The IT idea: handle simple cases — a domain, a title, an error text — by fast pattern matching, and call a heavy model only where patterns do not decide. Earlier figures about server heat, model latency and '12–18% hallucinations' were not measured and have been withdrawn.
Connectome Innovation 14
Inspired by the fly ellipsoid body: according to the literature, a ring of 'compass' neurons holds a bump of activity that turns with the fly and keeps the heading. The idea: keep the current phase of a dialogue in such a ring. Our implementation is a ring attractor (Amari equation, 'Mexican hat' weights); the run used 16 neurons, as stated on the card (the engine default is 64).
Connectome Innovation 15
The FlyWire connectome labels each connection with a predicted transmitter. Measured 23.09.2026 from the release data (bench/connectome_transmitters.py, 54.46 M synapses with a prediction), share of synapses by most likely transmitter: acetylcholine 55.8%, GABA 23.3%, glutamate 17.8%, dopamine 1.4%, serotonin 1.2%, octopamine 0.6%. The earlier shares on the card (acetylcholine ~45%, dopamine ~5%, serotonin and octopamine ~3% each) were not confirmed. The model's idea: excitation and inhibition balance so the network keeps a set level of activity.
Connectome Innovation 16
Inspired by fly sensory adaptation: according to the literature, a constant background stops being perceived while a rare important signal is amplified. The idea: give rare features more weight and prune frequent connections. Formula: w(f) = log(1 + N/df) · (1 − e^(−λ·Δt)), pruning threshold 5% of the largest weight.
Connectome Innovation 17
The idea: describe an AI system's architecture as strictly as FlyWire describes a brain — every element has an ID and a type, every link a type and a weight. The CADF format: S = ⟨V, E, T, W⟩, where the link type T ∈ {Sync, Async, Inhibitory, Modulatory} and the weight W > 0.
Connectome Innovation 18
The idea: just as FlyWire gave everyone one open reference brain, an open set of website accessibility measurements can give researchers a shared reference. Our set is a keyboard crawl of US websites: which of eight typical pages open and can be navigated with a keyboard.
Connectome Innovation 19
According to the literature (Caron et al., Nature 2013; Litwin-Kumar et al., Neuron 2017), a fly Kenyon cell receives inputs from about 6–8 projection neurons, a number considered close to optimal for separating odours. We tested this on our own tasks.
Connectome Innovation 20
The idea: show activity spreading through the fly brain visually — in a terminal and in a browser, with sound driven by the number of spikes. Projection of points to the screen: [u, v] = R(α, β)·[x, y, z]; sound: f(t) = f0 + k·Σ s_i(t).
Connectome Innovation 21
Inspired by the fly central complex: according to the literature, P-EN and P-FN neurons link the ellipsoid body and the protocerebral bridge and help turn towards a goal. The card's formula: V = α·∇Φ_goal − β·Σ repulsion from visited nodes. The engine code (aifa_sdk/cx_steering.py) implements only the first part — attraction to the goal; the visited-nodes field is declared but never used.
Connectome Innovation 22
Inspired by fly neuromodulation: according to the literature, dopamine, octopamine and serotonin switch the brain between sleep, quiet wakefulness, foraging and escape. The idea for a crawler: change the number of parallel workers and the pauses by successes and errors, the way modulators change the brain's mode.
Connectome Innovation 23
Inspired by the APL neuron: according to the literature, one large inhibitory neuron per hemisphere receives input from all active Kenyon cells and inhibits them all proportionally, so only a few stay active. The card's formula: A = ReLU(X − quantile(1−k)), then normalisation by the sum.
Connectome Innovation 24
The C1-FFL motif (Alon, 2007): node X switches on Y, and the output Z fires only when both X and Y are active (logical AND). A short pulse has no time to pass through Y and dies out; a long signal passes. How enriched this motif is in FlyWire v783 specifically has not been checked by us.
Connectome Innovation 25
According to the literature, T4 and T5 neurons of the fly visual system detect the motion of light and dark edges by the Hassenstein–Reichardt model: one photoreceptor's signal is delayed and multiplied by its neighbour's. The idea: detect dangerous flicker and intrusive animation on a page the same way. The earlier 'a million times more efficient than graphics cards' is not a measurement and was withdrawn.
Connectome Innovation 26
A k-core is the largest subgraph in which every node has at least k neighbours inside it; it is found by repeatedly removing nodes of degree below k. Measured 23.09.2026 on the real FlyWire v783 connectome (bench/connectome_topology_extra.py, pairs with ≥5 synapses): k_max = 57, with 258 neurons in the innermost core. The earlier 'k_max = 78, 1,420 neurons' was not confirmed.
Connectome Innovation 27
According to the literature (Turrigiano, 1998), a neuron scales all its input synapses so that their total strength stays constant — otherwise Hebbian learning drives it into saturation. The idea: keep the 'volume' of memory constant in the same way and remove weak unused links.
Connectome Innovation 28
The idea: use a real brain graph as a test for graph databases, because synthetic graphs do not reproduce its structure. Measured 23.09.2026 on the real FlyWire v783 graph: 139,255 neurons, 15,091,983 connected pairs, 54,492,922 synapses; at a threshold of ≥5 synapses — 134,181 neurons and 2,511,789 edges. The earlier '3,869,878 connections' and '3.87 M edges' were not confirmed.
Connectome Innovation 29
According to the literature, the fly brain is mirror-symmetric and the hemispheres exchange signals through commissures. The idea: two independent 'hemispheres' score an answer, and it is accepted only if they agree. In the code (aifa_sdk/bilateral_verifier.py) agreement is √(s_L·s_R)·(1 − 0.4·|s_L − s_R|) with a threshold of 0.52 (the card stated 0.95).
Connectome Innovation 30
According to the literature, neurons of the fly central complex (E-PG, P-EN, P-FN, Δ7) form a continuous ring attractor: the activity bump can sit anywhere on the ring and holds without external input. Our engine (aifa_sdk/cann_focus.py, 64 neurons) implements this correctly: 'Mexican hat' weights, position read out by population vector.
Open Research Benchmark & Enterprise Core Licensing
Reproducible open-source evaluation framework under Apache 2.0 alongside sovereign closed-core binary licensing for autonomous robotics, agents, and air-gapped perimeters.
AIfa BioBench Suite v3
Complete reproducible evaluation suite for connectome LSH hashing, CANN continuous attractors, and multi-agent associative retrieval.
43.2 ms (P50, 23.09.2026); the query is a noisy copy of a database vector (σ = 0.08), not an independent search.
43.9 ms (P50, 23.09.2026); queries are independent of the index — a representative search. Exact FAISS: 100% in 8.8 ms.
- Measurement suites: retrieval 50,000 × 1024, method duel, claw sweep, agent test on real engines, dropout test, FlyWire v783 connectome statistics.
- CPU, no GPU: runs on any computer with Python and NumPy; in-browser search is JavaScript (not WebAssembly).
- Open Apache 2.0 licence: free use in scientific publications, comparisons and research.
AIfa Cognitive Runtime Core (.aci)
Closed-source high-performance binary runtime core for autonomous agents, industrial robotics, aerospace, and sovereign air-gapped perimeters.
Search over 50,000 × 1024 vectors on CPU, measured 23.09.2026.
Delivery: Linux (.so) / Windows (.dll) / C++ SDK with Merkle audit — under contract.
- Custom distillation: training and calibration of projection matrices for the customer's domain.
- Cryptographic audit: file fingerprints and a Merkle tree with an OpenTimestamps Bitcoin timestamp.
- Dedicated R&D co-development: a dedicated solutions architect, direct contact with the Chief Architect and a guaranteed SLA.
Legal Status, Copyright & Intellectual Property Protection
Sole Creator, Author, and Chief Architect of the AIfa Cognitive Runtime (ACR) architecture and all 30 connectome innovations: Maksim Valentinovich Galatin. All exclusive rights strictly reserved.
IP Demarcation: Drosophila connectome substrate is open science (FlyWire Consortium, Nature 2024, CC-BY-4.0). Our proprietary IP is the mathematical CPU-native distillation of 6-claw projection, k-WTA APL gating, steering navigation, and runtime packaging. Benchmark script aifa_biobench.py is open under Apache 2.0. The compiled binary AIfa Core runtime (.aci) is strictly proprietary under CODE Eternal Commercial EULA & Trade Secret.