The BLOBS program — matter, chemistry, and machines from three fields

A research-post series, updated as results are audited. Latest: post 12 — the final two-island, twelve-generation v3 harvest. Partial-C9 proxy progress and added descriptor breadth are not proof of heterogeneous compartments. Each post states its measurement and validation limits. Probe code + logs: probes/blobs/.

The idea. Instead of simulating creatures or particles directly, we write down three interacting fields — think of three coupled "chemical concentrations" filling space — chosen so that the field equations themselves hold stable, particle-like excitations. Those excitations ("blobs") turn out to move, bind into molecules, come in species, and can be assembled into working machines. The program climbed that ladder one certified rung at a time: existence → motility → binding → flavors → composite dynamics → a machine.
Literature grounding (added after a 132-source review). The three-field system below is not ours: it was introduced by Schenk, Or-Guil, Bode & Purwins (Phys. Rev. Lett. 78:3781, 1997) to model planar gas discharges, and "blobs" are that field's dissipative solitons (review: Purwins, Bödeker & Amiranashvili, Adv. Phys. 59:485, 2010). Much of our phase-1 ladder reproduces known results — spot existence, the drift bifurcation and its √-law (Or-Guil et al. PRE 57:6432; Krischer & Mikhailov PRL 73:3165), tail-quantized soliton molecules (Bode et al. Physica D 161:45; measured bond forces: Bödeker et al. New J. Phys. 6:62) — and we mark those sections accordingly. What appears new here: the pair-only translation zone used as a transport selector, the nonreciprocal heterodimer rotor design (rotating bound states themselves were found by Moskalenko, Liehr & Purwins, EPL 63:361 / EPJ B 37:199), the relay-tug cargo machine, the self-written-landscape design laws, and the algebraic pre-filter for equation-space search. Full review: probes/blobs/litreview/REVIEW.md.

The series

#postwhat's in it
1What is a blob? the 3-field theory, the vacuum, dissipative solitons, the corridor between death and cancer
2Blobs move the drift bifurcation and its √-law; the lattice-pinning trap and IMEX-FFT (why our telescope needed fixing)
3Blob chemistry bonds at quantized distances, the saddle alarm, molecules; the labyrinth instability; two port-distinguishable species
4Molecules that move A=τ·Dv, wake-locked tandems and trains, the pair-only zone, the heterodimer rotor (spontaneous rotation), libration
5The living landscape b as a real field: autophoresis, trails and stigmergy, b-assembly, self-written sawtooths and self-dug racetracks, the reduction map
6Blob machines the relay tug; tow/dock/fork and the glue machine; the convoy line and the stack-safety / blade-load / buckling law cascade; composition = footprint geometry
7Searching equation space the genome, the algebraic funnel, MAP-Elites atlases, merge-beats-mutate, the 3-field engine and the plateau-bond family; stage-3 partials
8Bounded structures closed blob rings (π₁≠0), porous walls and the cross-w barrier, speed-selective membranes, cargo-in-cell, and the rigidity of one-way walls
9Accelerating blobs the GPU port: a generation as one tensor, the correctness gates (f64 parity, bond anchors, the locked assay battery), the A100 roofline, 57×/core — $0.005 per world. Plus the production sequel: the record path meets Amdahl, two fixes die by benchmark, and blob-finding moves onto the device (scatter-min labeling, exact partitions, 7.6×/field)
10Evolving at scale the v2 campaign: 9,409 evals across four compute eras → a 423-cell archive, champion at 91.2; the ×32 champion lineage (add_chan/merge/mint/delete credited per hop), the operator scoreboard, the archival ratchet (selection is archival, not populational), cap-riders and the telescope limit — with films of the five best evolved worlds
11Measuring evolved worlds BLOB2v2r2, current-first: known syllabus, hidden instances, agent-chosen ready and closed-book predictions; seven private resource guards and eight-state LRU residency. The stopped pilot reports two completions, not a clean model benchmark. Paired E1/E2 process and token audits separate record interpolation, continuation sampling, and narrower empirical response models from learned dynamics, expose parallel state-integrity failures, and set out a future-only native-controls → wiring-smoke → full-pilot → matched-panel test ladder; earlier cohorts stay separate
12Breeding spatial economies ● updated the audited final v3 harvest: two islands × 12 generations; C9 factors and partial-mode limits; at common W9 0.40, 241/1,376 baseline screens and 159/723 continuation screens qualify; 14 newly observed qualifying descriptor bins, not physical cells; best correctly linked selected partial-C9 assay 0.8504 → 0.8895, without uniform reseed improvement; corrected operator yields, provenance limits, and six labeled GPU re-simulation films. Next: identity safeguards and regional-measure validation before another campaign

Reading order is the research order, but each post stands alone and cross-links where it depends on another. The reference material below applies across the series.

Reference: the dial table

All measured windows below are from the certified probe campaigns (fresh-seed audited). "Death" = blob decays to vacuum; "replication" = blob splits into spreading spot soup — the two cliffs that bound every corridor in this world.

dialmeaningbaselinemeasured effect & window
λdrive / nonlinearity scale2.0sets activator plateau (±√λ) and overall energy scale; not scanned as a control dial
k₁global bias ("vacuum depth")−0.7THE existence dial: −0.9 death · −0.7 blob · −0.5 replication (M0). Traveling corridor k₁∈(−0.75,−0.68). Binding sits ~0.1 from the replication edge. Positional shifts of k₁ = force on blobs (see isok)
k₃v-inhibition strength1.0held fixed; k₃/k₄ balance sets how much of the inhibition is "memory" vs "halo"
k₄w-inhibition strength1.51.35–1.65 binding window; 2.5 = spot soup at M0; ≥1.7 replication cascades when pairs form. In flavors, k₄ is the species dial (A:1.40, B:2.15) along the iso-line
τslow-inhibitor time constant3.0THE motion dial. Single-blob drift onset τ_c=4.78 (A≈3.1); pair onset 5.636 (A=4); pair-only zone (5.636, 5.748); replication ≥6.2 (A=4). Statics blind to τ at fixed A
θfast-inhibitor time constant0.70.35 kills the M0 blob (halo reacts too fast); windows ≥2.2× wide around baseline elsewhere
Duactivator diffusion1.0blob core size scale; flavors species use 0.65
Dvslow-inhibitor diffusion1.0with τ only via A=τ·Dv for statics: A=5 deep bond d*=15.70 (never travels) · A=4 shallower bond d*≈15.4 that CAN travel · A≈3.1 single-blob motility. Dynamically: narrower v (smaller Dv) = easier drift
Dwhalo range20long-range repulsion & wall cushion; also the stiff term that forces implicit integration (explicit Euler needs dt<dx²/(4Dw))
σ (noise)additive field noise2·10⁻³blob survives to 0.075, dies 0.09; bonds outlive 4000 tu at 0.075. Too stiff for noise ratchets: positional diffusion ≈ 0 (measured; honest negative)
ε (isod/isok)slope of the zero-footprint load field—two-species world: cargo drift v=−0.906ε (linear to 0.03, safe ≤0.02). Single-species near-onset (τ=5.7): susceptibility amplified ~50× (v up to −0.08) — measured, used as the machine's adversary. Machine tooth ε=5·10⁻⁴; pair reverses at b≈0.005; 3-train climbs to b≈0.0104; replication cliff b≤−0.09

Vocabulary alignment (full table in probes/blobs/litreview/NAMING.md): our "encounter table" = scattering outcomes organized by unstable "scattor" states (Nishiura et al. Chaos 13:962); our isok landscapes = the heterogeneous-media program of Nishiura et al. (bump/jump heterogeneities, pinning/depinning — Chaos 17:037104), whose taxonomy we adopt for dock/pin/rebound behaviors; our replication cascade = self-replicating spots (Pearson Science 261:189).

"Rails", precisely

A rail is not an object — it is a static, zero-footprint potential in the cross-track direction: bchan(y) = 0.002·min(|y−y₀|, 24) added to the same iso-line shift that makes the track. It is a V-shaped valley in parameter space, invisible to the vacuum, that gently pushes any blob back toward the channel centerline (measured: cargo injected 6 px off-axis is centered to ±0.5 px while being conveyed). Rails were forced by an honest failure: without them, the growing train buckles sideways into the saw's trough valley after 1–2 pickups. (In the two-species world we also certified self-assembled walls — a blob parked on a gradient ridge destabilizes into a static stripe that blocks or channels cargo — defect turned tool; the machine uses the simpler potential rails.)

Reference: what's still missing (mapped, not hidden)

missing piecestatus
Unload / releaseSOLVED in phase 3 (section 9a): η(x,y) null zones release towed cargo — coupling geography, not bond-breaking
RotationSOLVED in phase 2 (M7 heterodimer rotor, section 8) — same-species rotors remain knife-edged
Flavor-selective machinesneeds B–B binding in the two-species world (never searched); currently cargo = carrier species
Big continuum speciesspecies A is lattice-stabilized; a truly continuum large blob is parked until a machine needs the 6.8× size contrast
Max-train-length lawthe 6-train sheds its rear blob (power ceiling) — worth mapping if longer hauls are wanted

Reference: provenance & audit trail

Program spec and gates: probes/blobs/PROGRAM.md. Milestone searchers ran as independent agents with locked metrics; the controller re-derived every headline on fresh seeds before certification (M1 speeds within 5%; M2 bond from fresh stamps on a larger box: 15.67/15.70; M4 speed law within 3% + out-of-window 6.3%; M5 machine rerun: 6/6 blobs upstream, efficiency 5.8×, null clean). Convention-faithfulness matters: three audit attempts initially "failed" by omitting a searcher's documented protocol detail (symmetric ICs; the exact gain eps_g; species stamps) — all three resolved on faithful replay. Raw run logs, job specs, and every failed candidate are in the per-milestone results.json files. Full chronological narrative incl. retractions: REPORT.md addenda 29–32.