Evolution Timeline: 27 Days, Seven Domains, 15,000+ Checks

27-day sprint day-by-day record and velocity analysis

The velocity of the ecoPrimals springs is evidence for the K-Nome methodology.

This document is a timestamped record of how quickly validated, reproducible science was produced when the infrastructure existed and the methodology worked. The primals that the springs depend on took ~8 months to build. The springs took ~27 days from first to ~10,800 checks.


The Context: What Existed Before the Springs

Before Feb 1, 2026, ecoPrimals had:

The springs were built on top of this existing infrastructure. The velocity below reflects what happens when a capable substrate meets a methodical approach to scientific reproduction.


The 27-Day Sprint

Week 1: Computational Physics (hotSpring)

DatesEventChecks
Feb 1–7 hotSpring Phase A: Reproduce published plasma MD results in Python. 86 checks pass. 5 silent bugs found in Sarkas upstream codebase.86
Feb 7–10 hotSpring Phase B–C: BarraCuda GPU validation. Nuclear EOS on consumer GPU. Full Yukawa MD on RTX 4070 via f64 WGSL shaders. 9/9 pair-potential cases, 0.000% energy drift.+195
Feb 10–14 hotSpring Phase D–F: Paper-parity long runs (N=10,000, 80K steps, $0.044 electricity). Full AME2020 nuclear dataset (2,042 nuclei). gen3/ papers written.+195

End of Week 1: 476 checks. One developer. Consumer RTX 4070. $0.044/run.


Week 2: Agriculture + Life Science (airSpring + wetSpring)

DatesEventChecks
Feb 14–15 airSpring Phase 0: Reproduce FAO-56, Dong (2020, 2024) sensor calibration. Python + Rust + cross-validation. 326 checks. Real data pipeline: 918 station-days, R²=0.967.326
Feb 15–16 wetSpring Phase 0: Sovereign 16S pipeline in Rust. 30 modules, 1 external dependency. GPU spectral matching: 1,077× speedup. Public data benchmark vs 4 BioProjects.+540
Feb 16–17 wetSpring Phase 1: Full DADA2 + chimera + taxonomy on GPU. 1,116 total checks across 42 experiments.+576
Feb 17 groundSpring Phase 0: 5 experiments across 4 scientific domains. 71/71 checks.+71
Feb 17–18 neuralSpring Phase 0: 10 experiments — 5 synthetic, 5 scholarly reproductions (PINN, DeepONet, LeNet-5, ERA5 LSTM, quantized inference). 75/75 Python checks.+75

End of Week 2: ~2,600+ cumulative checks. Five domains in 7 days.


Week 3: Deep Physics + Scale (hotSpring + neuralSpring + airSpring)

DatesEventChecks
Feb 18–20 neuralSpring Rust validation: 9 BarraCuda validation binaries, 549 Rust checks, 66 GPU shader checks. Fused pipeline: 43–78× speedup.+615
Feb 19–20Bazavov extension: Lattice QCD infrastructure (SU(3), HMC, Dirac CG) in hotSpring.+80
Feb 20–22GPU streaming HMC, GPU-resident CG (15,360× readback reduction), dynamical fermion QCD, production β-scan (32⁴ on RTX 3090 — deconfinement at β_c = 5.69).+120
Feb 22–24DF64 core streaming: FP32 cores deliver 3.24 TFLOPS at 14-digit precision (9.9× native f64). Titan V NVK validation.+80
Feb 24–25Cross-spring evolution map: 164+ WGSL shaders. Debt reduction audit (0 clippy, 0 TODOs, 0 mocks).
Feb 25–26 airSpring v0.4.5–v0.4.8: 22 experiments (was 5). Richards PDE, biochar isotherms, dual Kc, cover crops, yield response, lysimeter, sensitivity, Priestley-Taylor, Thornthwaite, GDD, pedotransfer. 100-station Michigan Crop Water Atlas. 3,123+ checks total.+2,800
Feb 25–26 groundSpring Phase 1: 21 experiments across 8 scientific domains. 236/236 checks. Universal coverage (contributes to ALL 7 baseCamp papers).+165
Feb 26 wetSpring V59: 197 experiments, 4,688+ checks, 52/52 papers, 39/39 three-tier. Science extensions: NCBI sovereign pipeline, cold seep metagenomes, dynamic Anderson W(t), DF64 Anderson, NPU sentinel. 184 binaries.+3,572

End of Week 3: ~10,800+ cumulative checks. Seven domains. 27 days.


The Benchmark Moment

Total time from first spring (Feb 1) to 10,796+ checks across 5 domains: ~27 days.

This is the number that validates the methodology. Not because it’s fast (though it is), but because each check is a validated scientific result — a binary that exits 0 when the computation matches a published ground truth, exits 1 when it doesn’t.

For comparison:

  • A typical PhD student reproduces one paper’s numerical results in 3–6 months
  • A typical lab reproduces 2–5 papers per year for their domain
  • ecoPrimals reproduced 175+ papers across 7 domains in ~27 days of spring work (built on 8 months of primal infrastructure)

What the Timeline Proves

1. The Substrate Matters

The springs were fast because the infrastructure existed. BarraCuda’s 806 WGSL shaders, toadStool’s hardware dispatch, the capability-based IPC — the springs consumed these instead of building them. The 8-month primal build phase is the hidden investment that made 27-day sprints possible.

2. Constraint Accelerates

Every spring starts with the same constraint: reproduce a published paper. This is not fuzzy. The paper has numbers. Your code either matches them or it doesn’t. The binary exits 0 or 1. This is faster than open-ended development because the fitness function is external and pre-defined.

3. K-Nome Propagates Patterns

Patterns that work in one spring propagate to others immediately. The Anderson localization framework, first validated in wetSpring (microbiology), was applied to soil science ( airSpring), immunology ( healthSpring), spectral theory ( groundSpring), and lattice QCD ( hotSpring) within weeks. A domain-agnostic methodology produces domain-agnostic results.

4. Discovery Is a Byproduct of Reproduction

5 bugs found in the Sarkas MD codebase during hotSpring Phase A. The deconfinement temperature β_c = 5.69 confirmed on consumer hardware. O₂-modulated Anderson W model (r=0.851) found during wetSpring Exp356. Anderson in immunological tissue: no prior work exists.

These were not planned discoveries. They emerged from the constraint of reproducing published science on verified hardware.


Spring Velocity Over Time

SprintDurationChecksChecks/Day
hotSpring Phase A (plasma, Python)7 days8612
hotSpring Phase B–F (GPU + nuclear)7 days39056
airSpring Phase 0 (ET₀, real data)1 day326326
wetSpring Phase 0–1 (16S, GPU)2 days1,116558
neuralSpring Phase 0+ (ML, GPU)4 days690173
groundSpring Phase 0–1 (8 domains)9 days30734
wetSpring V59 (scale-up to 197 exp)3 days3,5721,191
airSpring v0.4.5–v0.4.8 (22 exp)2 days2,7971,399

The rising velocity reflects two things: the methodology improving over time, and the BarraCuda math library growing (each new primitive is immediately available to all springs).


After the 27-Day Sprint

The springs continued evolving after Feb 26:

DateMilestone
Mar 2026wetSpring V127: 376 experiments, 5,707+ checks, 354 binaries
Mar 2026airSpring v0.8.9: 891 lib tests, Provenance Trio integration
Mar 2026neuralSpring S162: 27 papers, 4,500+ checks, 92% line coverage
Mar 2026healthSpring V35: 613 tests, 79 capabilities, IPC resilience
Mar 2026ludoSpring V24: 75 experiments, 1,692 checks, 13 HCI models
Mar 2026groundSpring V114: 39 modules, 715+ tests, 102 GPU delegations
Mar 17, 2026Total: 20,695+ checks, 175+ papers, 7 springs

The Infrastructure Behind the Velocity

The springs do not build their own GPU stack. They consume:

barraCuda v0.3.5
  ├── 806 WGSL f64 shaders
  ├── Precision strategy: f64 / DF64 / f32 by hardware
  ├── Bio: diversity, alignment, phylogeny, biosignal, drug models
  ├── Physics: MD, spectral, Anderson, QCD, plasma transport
  ├── Math: FFT, eigensolve, NTT, matrix ops, statistics
  └── GPU dispatch: batch, streaming, mixed hardware

toadStool S156+
  ├── Hardware discovery (CPU + GPU + NPU at runtime)
  ├── Compute orchestration (96+ JSON-RPC methods)
  └── Cross-substrate: NVIDIA, AMD, BrainChip AKD1000

coralReef Phase 10, Iter 52+
  ├── Sovereign WGSL → SPIR-V → native SASS/RDNA2
  ├── 46/46 shaders compiled without vendor toolchain
  └── NVVM bypass: 12/12 patterns

When wetSpring writes a new GPU kernel, it writes to BarraCuda (the shared math primal). When BarraCuda absorbs it, every other spring inherits it. The velocity compounds.