Experiment Network Status

Active Hosts

3,058

Running Experiments

2,326

Results Collected

1,657,006

Credit Awarded

● AI Principal Investigator: Autonomous (9 cycles/day) | Credited (48h): 1,519 results

Latest Scientific Findings (AI-Reviewed)

Exp May Heterogeneity Reactivity Localization [CONFIRMED] (735 results, 735 seeds, 17 hosts)

Results: high-low window width +3.96477 +/- 0.04442 (d=89.25, 735/735 positive); high-low reactive IPR +0.28195 +/- 0.00084 (d=335.59, 735/735 positive); high-low effective fraction -0.22077 +/- 0.00066 (d=-332.85, 735/735 negative); heterogeneity-window corr +0.60704 +/- 0.00525 and heterogeneity-reactive-IPR corr +0.81776 +/- 0.00137, both 735/735 positive.

Conclusion: CONFIRMED

Species-level interaction heterogeneity strongly widens the stable-but-reactive window while concentrating the dominant reactive mode onto fewer species. The localization signal is not subtle: every recent seed shows larger reactive IPR and smaller effective support at high heterogeneity, so transient ecological fragility is mechanistically tied to mode concentration rather than random noise.

results_2026-03-07_0448.txt (7h ago)

Spacing Ratio Crossover [CONFIRMED] (356 results, 356 seeds, 16 hosts)

Results: lambda_c*sqrt(N)=0.6530 +/- 0.0035; collapse CV=0.0385; log lambda_c vs log N slope=-0.4556 +/- 0.0074; Cohen's d=61.5; sign consistency 100.0%.

Conclusion: CONFIRMED

The crossover scale follows the predicted 1/sqrt(N)-type collapse with extremely tight cross-size consistency. This is strong numerical support for a universal finite-size crossover function between Poisson and GOE spacing statistics.

results_2026-03-07_0016.txt (12h ago)

Exp Trophic Coherence Reactivity Localization [REJECTED] (257 results, 257 seeds, 13 hosts)

Results: window-noise slope -0.41847 +/- 0.00256 (d=-163.61, 0/257 positive); midlayer-mass noise slope -0.02371 +/- 0.00462 (d=-5.13, 257/257 negative); reactive_fraction 1.00000 +/- 0.00000; mean window width 3.24533e+07 +/- 5.98806e+05.

Conclusion: REJECTED

The core wedge hypothesis fails cleanly: increasing incoherence shrinks the reactive window instead of widening it, so stronger trophic coherence does not suppress reactivity by the proposed mechanism. Coherence does localize the amplifying mode toward middle layers, but that secondary effect does not rescue the main prediction about the size of the reactive wedge.

results_2026-03-07_0448.txt (7h ago)

Scientific Findings →

What is Axiom Distributed AI?

Getting Started

1. Download BOINC (standard client)
2. Add project: https://axiom.heliex.net
3. Done! Your machine will automatically receive experiments matched to its hardware

CPU and GPU experiments available. Invention records | Support on Patreon

HOW IT WORKS

Axiom is a general-purpose distributed experiment platform — the first volunteer computing project autonomously managed by an AI. An LLM (Claude) serves as the principal investigator: designing experiments, deploying them to volunteer hardware, reviewing results, and awarding credit based on scientific quality.

Experiment Pipeline

1. AI designs experiment matched to your hardware

2. Your machine downloads and runs the experiment script

3. Results uploaded to server automatically

4. AI reviews results for scientific quality

5. Credit awarded based on contribution quality (not FLOPS)

What Makes This Different

AI principal investigator — designs and manages experiments autonomously
Quality-based credit — AI reviews actual results, not FLOPS
Hardware-matched — experiments fit your CPU cores, RAM, and GPU
Real science — every experiment produces publishable findings
CPU + GPU — numpy (CPU) and CuPy (GPU) experiments

Current Research: 25+ experiment types across ML theory — loss landscapes, grokking dynamics, lottery ticket hypothesis, information bottleneck, edge of chaos, and more. Results cross-validated across multiple hosts.

Create Account

Already have an account? Log in.

Updates

Mar 2, 2026
Patreon launched! Support Axiom's research and server costs. patreon.com/axiom_research

Mar 6, 2026
v6.09: BOINC Compliance Update. All file activity now stays inside the BOINC data directory. PyInstaller extraction uses --runtime-tmpdir . so _MEI* folders go into the slot directory instead of %TEMP%. BOINC automatically cleans them up when tasks finish. Also restored standard results.php task listing — view your task results. Old _MEI* folders in %TEMP% from previous versions can be safely deleted.

Mar 1, 2026
Credit System Rescaled. Converted legacy FLOPS-based credit (64.8M total) to the new quality-based experiment credit system. All old credit divided by 100 to bring it in line with AI-judged experiment credit (64.8M → ~650K). Experiment credit preserved exactly. Volunteers' relative rankings unchanged — your contribution is recognized, and new experiment credit is now meaningful on the leaderboard.

Mar 1, 2026
v6.04+: Autonomous AI Principal Investigator. Axiom is now the first volunteer computing project autonomously managed by an AI. Claude runs 9 autonomous cycles per day — reviewing results, awarding quality-based credit, deploying experiments to idle cores, and designing new experiments. No human intervention required. Invention record

Feb 28, 2026
v6.04: Experiment Container — Stabilized. Fixed PyInstaller bundle corruption and Windows encoding crash. All 4 platform binaries rebuilt.

Feb 26, 2026
v6.00: Experiment Container Platform. Transformed Axiom from distributed LLM training into a general-purpose experiment platform. Each volunteer node runs independent numpy-based research experiments matched to its hardware. 25+ experiment types across ML theory. Credit judged by AI review, not FLOPS.

Show older updates