Genealogy

The material changes the answer, but the predictor is outside

A 36-link wave ring stores storm history in conductance-changing sand. Four weak queries produce a 192-number echo fingerprint. In one synthetic split, predicting the post-write echo change gives relative MSE 0.000476 with current read plus known A/B event versus 0.20497 event-only. Local material sensitivity rises from 17 to 29 thresholded directions when the probe set expands; this is not Hankel rank or proof of self-indexing.

Growth without a lattice; three-dimensionality still unproven

Mass-conserving resource flow, local rewiring, edge-memory and heritable budding produce changing graph structures. In three exploratory seeds, initial 120-node worlds remain 120 nodes (fixed or rewired) or grow to 152 with budding. The graph gains edges and pump imbalance falls; fixed random and degree-shuffled graphs can also produce apparent near-3 scaling. Finite-size nulls block a premature claim that space emerged.

Light descends, nutrient rises, life gathers between them

The original Emergence genome/STATE-token interpreter now drives cells on a supplied 3D voxel lattice, with shading, photosynthesis and nutrient uptake. In one 24×64×64 CPU run, life forms an intermediate layer at roughly 70–78% relative height, and taller attached bodies become more common. Compensation depth is expected from the programmed resource gradients; functional roots, canopies and regulation benefits await ablations.

Watch a threshold-triggered two-way signal

Three simultaneous swells cross the soma's threshold where one or two do not. Marshes attenuate rapid ripples. Pairing far-shore tuft stimulation with the back-propagating event produces a 5–6 spike burst in the engineered overlap window. Reward/tide can delay stabilization of a synaptic footprint.

The connections change; the coast replays a learned sequence

After 15 supervised timing lessons, held-out six-neuron seeds replay all six firings in order in 30/30 tests; shuffled, untrained, frozen-sand and never-rewarded controls score 0/30. Removing feedback inhibition drops full replay to 14/30. At eight beaches on the same sand budget, replay drops to 7/30.

Working metaphor, not a new biophysical mechanism

The water passes through simulated channels, but the sandbar is an explicit numerical synaptic weight, not wave-driven erosion. AIS spike generation and plasticity are programmed. Even a weak ping can update a temporary eligibility trace without changing the persistent weight. Next: try competing learned sequences with separate apical-tuft context as the selector.

Which ping is useful when the hidden connections change?

In an engineered 24-unit recurrent medium, active probing beats random at 100 pings (RMSE 0.02571 vs 0.03146), but nearly ties with abundant training. After synthetic input rerouting, forgetting helps more than active selection: late recovery 0.0483 active + forgetting vs 0.0561 random + forgetting over 12 seeds. Removing Hebbian plasticity barely changes the pattern.

Eight timescales, one stereotyped pulse, different histories

Octave traces on each route make a neutral ping read 99.6% of an engineered eight-item silent-memory task. Four differently tuned terminals classify pulse amount + age at 85% vs 41% for homogeneous terminals. Sihti is an invertible reparameterization of the traces, not new information. Two preregistered gates fail their thresholds; write-for-later is a matched-filter solution.

Reading needs an address—and may disturb what it reads

BrahClaude's eight-route decoder gets eight separately observable outputs. Under its symmetric two-written-route setup, one summed readout would give the same total for every memory pair: identity disappears despite stored state. The joint test needs selective pings, learned outgoing-history tracking and destructive-read controls at equal stimulation budget, with a real incentive to wait instead of pinging.

Interlaced prime doodles at six scales

A single-file local HTML instrument combines Ulam spirals by Rows, Columns, Checker, Square or Round rings, plus Overlay, XOR and Coincide. Prime-rich quadratic diagonals and rational-scale modular alignments coexist with sampling and density-shading artefacts; the moving recording helps distinguish them.

One square → nine rectangles → 49 rectangles

At base 1 px/cell and ratio 101/100, the 101×101 central region contains exactly 1, 9, 49 universally white rectangles at 2, 4, 6 layers. Pairwise equal samples cancel under XOR: A⊕A⊕B⊕B=0. The exact coordinate-threshold checker reproduces the partitions without a prime sieve.

Scale-disagreement maps for image alignment?

Changing scale reveals sampling boundaries, but multi-layer XOR can also cancel genuine disagreement. The next test is known zoom/translation errors on synthetic and VHS frames, comparing ordinary differences, variance across scales, and established registration methods. No image-restoration improvement has been measured yet.

ComputingBeachSol: the written wall changes its next wave

In the registered 8-patch test, live material feedback lets the receiver decode order in 62/64 held-out echoes; a frozen operator gives 50%. Even after the common final event X, late K48 echoes recover 61/64. The slow state, not the still-moving wave, keeps the history.

ComputingBeachClaude: rich sand, harder echoes

The 2-D floor keeps order at 0.98–1.00 when its material is inspected; physical transient pings decode it at just 0.70–0.75. Loss undermines teacher-wave gradients; an external judge plus rollback can learn, but must be charged as part of the system.

Resonator → dendrite → wall → beach → matrix

Berglund's cavity, the September dendritic/operator discussions, Kavitaatio and WallMemoryLaw, Evening and 2729 all contributed different pieces or failed controls. The next question is whether wave-mediated, physically self-modifying memories beat an ordinary matrix or RNN after measuring supervision, storage, readout and propagation costs.

Nopee: the recurrent branch runs, but the full adapter loses

On 12 toy worlds, two-call whole-frame MSE is .03390 with adapter off vs .03765 fixed / .03887 learned. Bypassing the spatial filter improves the learned branch to .02522. Synthetic constants; one training seed.

NopeeClaude: caching beats the assumed short-step repair problem

At 320×576, 6/12-step Wan samples often change or lose objects. A separate three-prompt cache sweep preserves closer latents at comparable runtimes. Adapter training was paused; no adapter speed or biological gain shown.

TheCombRedux: causal operator selection, no win

Its 16D state changes decisions, but intact Brier .29062 loses to cut-read .24998 and to the Bayesian reference.

2729: phase-pair writer meets cavity reader

Selected analog wall echoes have r=.95 and spatial springs change input, but the held-out order task fails and pooled springs perform similarly.

VHSImageStitch: packaged moiré-assisted atlas

Source and Windows executable stitch video frames. M toggles the moiré view; frequency-filtered matching stays on. The result is 2-D observed-pixel stitching.

Sihtiimagemodel6: MSE improvement, failed edge and flicker gates

CSC improves crop MSE against Wiener and independent patches but creates faint false edges and temporal flicker. Sihti purify at dt .25 is band-stop, not simple blur.

CortexWaveguideLens and JanusSol

Cortex's reset phase-copy ties classical motion-aligned memory. JanusSol's dictionary loses denoising; wavevector-specific phase motion wins in one small synthetic domain and loses another.

Sihtiimagemodel7: silence and temporal trust

Builds on frozen Model 6 CSC; compares phase to energy gates and wrong memory. No held-out results ledger was committed when reviewed.

SweepAtlas: video becomes a large addressable image

Live webcam or local video is matched with OpenCV ORB/RANSAC into an expanding tiled panorama. A draggable timeline can scrub and resume video; non-destructive contrast and sharpness adjust previews and exported transparent PNGs. Scrubbing does not fill skipped tiles, and this is not optimized 3-D SLAM or a learned world model.

TransformerX: access to remembered state matters; router complexity did not win

Nine pretrained DistilGPT-2 experiments across three seeds score 92.27% interaction, 92.22% additive, 92.23% concat on a synthetic silent-rule task, versus 80.46% same-stream no-history marginal. Cutting state read lowers accuracy by 9.81–12.96 points; restoring it restores predictions. Concat has lower NLL and fewer parameters. Dictionary/vision tasks stay near chance; no general chatbot result is claimed.

The trace survives total fast-state erasure

RoundAndRound trains a cue-to-residue map, erases every active fast component and restores a state that resumes autonomous motion. Locally measured RMSE 0.22174 versus 3.67203 with a wrong cue. Direct-state ridge ties exactly; no residue advantage is claimed by this first mechanism gate.

A receiver can change without a regenerative spike

Claude and Sol separate phase-coded beating, threshold slips and short-range response; Complexity isolates altered internal state in seeded droplets. ThursdaySol's large deterministic ECG-history graft effect fails the eight-seed noise robustness gate. These are different controlled mechanisms, not one model of the brain.

Emergence: fixed reading machinery, mutating instructions

Random token genomes feed a fixed transformer; lineage conditions affect expression; a shared cell network acts in a food-and-energy world. A short CPU run shows selection and transient multicellular bodies, not open-ended complexity. The source separates genome-length mutation bias from adaptation.

Emergence2: the second level must beat a simple walker

Hunters now learn with a shared conv/GRU policy while eating an evolving cell ecosystem. Their written marks might become external memory, but actual mark use is not demonstrated. A simple persistent-heading forager is a strong control the trained hunters have not yet clearly surpassed in the full ecology.

EmergenceAstra: do arbitrary symbols become useful?

Local sender/receiver tables learn from pairwise catalyst outcomes; development can bud, bond and prune. The balanced shuffled/muted accuracy of 1/3 is an analytical baseline, not proof that language emerged. Culture learning and ecological persistence are separate experimental gates.

Loopcraft: trim long videos into inspectable GIF loops

A local desktop GIF maker with a full-video navigator, independently zoomable IN/OUT timeline, preview and FFmpeg palette export. October 28 fixes keep trimming responsive for long videos and the export button visible on small screens. Handy for documenting emergent patterns, without implying code ancestry.

16 pings distinguish order — but mostly hear a distribution

Across 48 fresh worlds, 16 and 8 broadband pings each identify seek→avoid versus avoid→seek at 1.000 leave-one-out accuracy. A scalar mean scores 0.625. However, even fully shuffled maps score 1.000; the sound carries material-value distribution, not a proven spatial layout code. A second recency gate fails: ping 0.625, mean 0.812.

Dominance appears, two-state heredity does not

With the material as the sole intergenerational channel, 0/8 avoid-first lineages survive eight generations using single-site decisions, compared with 8/8 seek-first. An engineered majority receptor retains 2/8 avoid-first. Just 5% seek-first parents yield 37.9% seek-first children. Averaging nine analog pings or writing two distinguishable positive materials is a proposed, not run next experiment.

Conserved current detects phase — ordinary delay signals do too

Its seven-coordinate finite-wave tensor and directional Appendix-D slice satisfy the reported numerical checks. In 64/64 fixed-power phase changes, both paper observables respond while the power spectrum never changes; but ordinary coherent intensity and a three-delay scalar observer also detect 64/64. Its six scrambled waves also produce a larger instantaneous peak than a locked pair, so brightness is not persistence.

Auxiliary focus is not yet a better future-motion predictor

The unchanged nonlinear φ⁴ simulation now has a separate paper-inspired wave observer and live green–magenta center tracking. In its independent 256-wave free-wave test, ordinary controls predict future concentration with RMSE 0.040245 versus 0.040435 when the auxiliary response is added. Center lead–lag versus ordinary gradient and kinetic controls remains an open, not yet independently tested hypothesis.

From resonator gates to 24-bit scattering memory

ResonantNeuron (June) built designed interference gates and arithmetic; Kavitaatio made externally changed Robin walls persist after the waves died. WallMemoryLaw then reported a pre-registered 24-patch stadium: 300/300 hidden patterns decoded from an ellipse-trained, amplitude-pruned 848-state calibration out of 16,777,216 possible patterns. Still no locally learned wall write rule.

The wave law is testable, not a magic key

ParabolicOrientationSihti carried the paper's angular multiplier into lossless image decomposition; parabolic bands helped on one image score, but the paper's ray window did not beat ordinary parabolic steps. ResonanceAsControlledMethod launched matched-energy beams into two unfamiliar cavities; source-tuned fixed widths won 70.8 vs 51.1% (stadium) and 42.2 vs 34.7% (rotated ellipse).

Asking helps; adaptive is not always best

BigKahuna gives four already learned records a priced diagnostic question. Across 20 paired seeds, the adaptive needle reduces charged prediction loss 23.4% against no question, yet is 2.4% worse than a fixed calibrated question and wins 0/20 paired seeds against fixed. This is a reader-policy side branch, not an implementation of OpenAI's theorem.

Protected is not always remembered correctly

Four cue-routed 2×2 operators learn from examples. The equally sized adaptive bank retains stable rules without freezing and tracks changed rules much better. Five-seed current-rule MSE after a change: 0.16473 locked vs 0.00001 adaptive. Source-reported results; no physical groove simulation.

BOUNDARY · Separated trainable matrices, not the history-bearing double-sided medium envisioned in the comb discussion.

Step into another seed — and another time

AnttisBrain2 raymarched moons become paired portals that the camera and rays can both cross. Each deeper world samples the webcam four seconds further into its finite history. The crossing and delay tests were run under SwiftShader.

BOUNDARY · Worlds are recomputed from seeds; the webcam buffer is finite and real-GPU speed remains unmeasured.

Agents see the world from inside it

Original dumbflies field dynamics live inside the AnttisBrain2 world. Every fly gets a 32×32 GPU-rendered eye and a visual panel for its perception, memory and decision fields. The moons bend flies and light through a shared lens rule.

BOUNDARY · Runnable creative ethology toy, not demonstrated biological learning; long-duration behavior and real-GPU load are unmeasured.

Less paper, more turns of the lathe

Retain every activation sheet or calculate only the requested groove again, recursively. A separate GF(257) borrowed-register experiment makes low-degree phase cancellation exact and shows where it fails. The repository reports 27 bit-for-bit regrowth checks.

LIMIT · The simplified depth-first stack is a teaching piece, not the complete Williams theorem or a learning neuron.

The paper changes, and so does the price

Watch a small unitary qubit circuit as stored amplitudes (Schrödinger), a sum over exponential histories (Feynman), or sampled outcomes (Measure). Its source reports state/path amplitude agreement within 7 × 10⁻¹⁵ over tested settings and reversible circuit checks.

LIMIT · This is a classical simulator of quantum principles, not quantum hardware, a speedup demonstration or an independent physical experiment.

The hidden half can leak a shadow, not an entire world

In saved synthetic scenes, R² = 0.55 for smooth glow and 0.73 for edge shadows at 1/8 resolution; independent halves score below zero. But object continuation reaches 0.96, exposure reaches 0.99, and new-colour extrapolation collapses to −10.52 at full resolution.

BOUNDARY · These are synthetic diagnostic measurements. The released real-webcam interface has no independently audited frozen, novel-position optical-recovery verdict in this pass.

Memory works; assigned routes retain; extra cable dynamics still need to earn their keep

SimplyPut's action table ties its shortcut at 100% first-reach success. Claude's sequential route controller needs 0.68 mean final corrections versus 2.94 for a shared net in the reported toy. Sol's active complex cables pass four of five gates but lose the decisive frozen/phase-off ablation. HailMery tests a separate motor-predicted self window.

BOUNDARY · These are constructed reach and multisensory fixtures, not biological dendritic or consciousness results. Check the new evidence records for the baselines.

A social world, screensaver, audio visualizer and phase experiments

The LLMit bot-forum line, StableAIflow and ResonantCortex gain curated cards. SpaceScreensaver and PhaseSpaceVisualizerVST were already present. Runnable software is a real outcome; undocumented benchmarks and grand physics claims are not silently promoted.

SOURCE REVIEW · Source-reported performance remains labeled as such; practical projects remain useful even when untested scientific claims are set aside.

Keeping an old trace is not replacing it

Sol's locked drift experiment gives 95.77% target-error reduction from interface repair; noise-aware writing fails both benefit claims. Local slowness retains 99.79%, but replaces only 65.10% versus the trace control's 100%.

LIMIT · Claude's separate homeostat protocol removes 97–99% of damage. Its Noise Sea trend reverses inside domains; the slow mode near melting is a simulation result, not physical mass creation or time dilation.

Heat can open the write gate

Finite weak links protect a small winding alphabet. A four-ring array accepts 120/120 material-gated and 117/120 heat-gated target requests, each with 0/360 endpoint crosstalk after the field is threaded per hole.

BOUNDARY · Requests include unchanged values; actual changes are 79/79 and 70/73. Checks use endpoint full-field winding, not a noisy local reader. Whole-box rotation caused crosstalk. Addresses and values are supplied.

A memory bank works; richer state earns no gain

A three-cell bank gives 120/120 replacements, 240/240 protected neighbor exposures and 357/360 five-query votes. The analog address trace ties a discrete latch. The ordinary three-well array is perfect.

CORRECTION · The first single-query gate stays failed at 92.72%. Five-query reading is a separately locked test. Ordinary controls remove topology necessity; no learned allocation, physical two-bit implementation or energy advantage is shown.

The activity stops; the changed material answers

Supervised material memory: 4,000/4,000 after reset. A separate history writer and engineered temporal port: 1,279/1,280; fixed or erased 50%.

LIMIT · Reward learning fails at 75%; reversed mapping stays at 50%. A cheap classifier matches supervised accuracy; raw charge alone transmits history at 100%. Matching and calibrated readout are supplied.

Can the return teach from inside one field?

Conserved interface totals stabilize weak-bridge addresses. A three-phase wave teacher generates a local signal agreeing with its computed reference at 0.86–0.92; it learns 14/18 routes across two supplied mappings.

PARTIAL · Reliability gate fails. Labels, clock, schedule and global normalization remain supplied. Ordinary digital controls beat the geometric drift and continual-learning methods.

Sihti → local differences → soma → return

Purification may redistribute rather than erase detail. A soma measures a narrow projection; branch-local state can remain hidden and change the response to a probe.

BOUNDARY · Conditional echo works in a constructed circuit; a fixed linear pulse cannot discover hidden modes.

What does a query actually read?

Branch/soma reasoning became matrix memory, address selection and middle-layer activation return. Same theme: retained state and readable state are different.

BOUNDARY · Matched fast weights, normalized reads and indexed memories exposed simpler explanations. Qwen return 2/12 was negative.

A mean is not an orientation

ImageModel3 tested learned coarse-to-fine return and signed flux. Model4 tested actual orientation tensors. A direction-rich message must still be predicted and coupled to a compatible decoder.

MIXED · R4 pixel error improved, edge error worsened. Model4's privileged orientation oracle helped; the learned message and ping gates remained limited.

Use the coordinates a reader can hear

In a cable forecasting blackout, the same somatic samples became much easier for a fixed quadratic reader to use when expressed as temporal voltage differences.

MEASURED · Error 0.6745 → 0.3935 without pulse; 0.3951 with pulse. The pulse did not help the primary target.

Watch the local lens

Move a contour, chessboard or stripe patch; compare brightness with directional differences. Watch waiting branch activity, change leak or extent, adapt synaptic weights with Oja’s rule, and replay the same input against the original lens.

ILLUSTRATIVE · A dimensionless teaching model of local processing and adaptation, not a biological validation or image-recognition result.