Core definition: PiX is a universal MDL candidate generator. π is not a law, prophecy, or magic shortcut. It is one generator family that can propose sequences, offsets, phases, distances, masks, textures, rhythms, levels, and schedules. MDL then decides whether any proposal is actually useful.
Original seed: π-sequences as a discrete partner to Fibonacci. New frame: PiX as a universal candidate layer for MDL×DCC. Latest PiT weekend test completed: 2026-06-15.
PiX does not mean “π predicts X.” PiX means a generator of candidate descriptions, parameters, and structures that MDL can accept or reject.
π is rich enough that beautiful coincidences are easy to find. The antidote is brutal controls: random, shuffled-π, e, φ, √2, uniform, and no-generator baselines.
Your original Pi story belongs here first: can π-addressed data plus transform plus residual describe chunks shorter than raw or ordinary compression?
The June 2026 PiT weekend test found a weak-to-moderate positive decision-slot seed, but not a live-ready trading edge.
PiX is π-indexed structure search. Use π as a deterministic generator of sequences, bits, numbers, phases, offsets, masks, textures, rhythms, or schedules — then let MDL measure whether this description beats ordinary descriptions and controls.
A candidate can be encoded as:
MDL accepts the candidate only if:
Instead of storing chunks directly, test π offset + transform + residual. Exact long matches will be rare; residual/lossy matches are the interesting part for images, audio, sensors, textures, and structured noise.
π windows can act as deterministic texture, grain, dither, or noise beds. A candidate wins only if address + transform + residual costs less than a normal representation at the chosen quality.
π schedules for restarts, mutation amplitude, beam width, zoom, explore/exploit pulses, and worker desynchronization in TSP, Sudoku, chess, crossword, ARC, and AMR arenas.
π can propose periods, tile sizes, masks, diagonal phases, modulo rules, or offsets. MDL then selects the shortest program that explains input → output.
π is natural for angles, circles, rings, radial bins, spiral paths, polar sorting, and local-search neighborhood sizes.
Markets are only one example. The same idea can test server load, traffic, weather, sports, EEG/Muse signals, sleep/HRV, network anomalies, and energy demand.
π windows, k-mer spacings, and offsets can act as hypothesis generators or null/control families against Markov-preserving controls.
π can generate rhythms, phrase lengths, notes, color palettes, camera cuts, tiling, animation phases, and deterministic creative seeds.
π alone is not cryptography, but it can be a public carrier when the offset is chosen by a secret key or hash. It can also create reproducible AI puzzle families.
Before building a trading arena, the better first build is a general PiX arena. It should expose π as one generator family among many and test it against strong controls.
| Domain | PiX candidate | Controls | Success condition |
|---|---|---|---|
| Compression | π offset + transform + residual | raw, zstd, lzma, random stream | lower MDL cost with verified reconstruction or controlled loss |
| Image/audio | π texture/noise/grain basis | random noise, blue noise, standard codecs | smaller residual at equal PSNR/SSIM/SNR |
| Optimization | π restart / mutation / beam schedule | fixed, random, Fibonacci schedules | better solve rate, gap, stability, or runtime |
| Program synthesis | π period, mask, tile, offset, modulo candidates | standard ARC primitives + random constants | better train fit + leave-one-out + test prediction |
| Trading slot | π levels / time windows / spacing | Fib, uniform, random, no-level | lower MAE, better rescue, better profit/risk |
The first 10-worker overnight run tested the arena itself across synthetic data, small built-in real-byte samples, planted image-residual textures, and a toy optimization scheduler. This was a stability and truthfulness run, not yet a full real-corpus test on the website/project archive.
completed jobs · 10 workers · about 8 hours · 0 failed jobs
all jobs passed planted-signal recovery; false magic on random controls stayed at 0
wins vs best baseline across 686,289 chunks in the default small sample set
π average rank by mean score; competitive, but not the winner
| Test | Observed result | Meaning |
|---|---|---|
| Fleet stability | 3,341 completed jobs, 0 failures, about 79.4 serial CPU-hours compressed into one overnight fleet run. | The arena harness is stable enough for the next real-corpus test. |
| Synthetic planted π | 3,341 / 3,341 jobs passed the planted-signal recovery test; random false-magic count stayed at 0. | The detector can recover real planted π structure without hallucinating it on random controls. |
| Real byte chunks | 0 MDL wins vs raw/zlib/best baseline across 686,289 tested chunks. Among constants only, π was often competitive, but never enough to beat the baseline. | No evidence yet that exact π byte-addressing compresses ordinary real bytes. This is a useful negative result. |
| Image residual toy | Planted π/φ texture signals were recovered. π produced 3,341 wins in the planted image-residual channel. | Residual PiX can work when the data actually contains a π-derived texture; next step is real images/audio. |
| Scheduler toy | π won 351 / 3,341 jobs. φ, Fibonacci, and e won more often; π was slightly better than random by aggregate mean, but not decisively. | The scheduler idea stays alive, but π is not privileged. It must remain one candidate family inside DCC. |
The original PiX insight is simple: Fibonacci/φ already guides many natural and technical search spaces, but π may provide a different family of discrete distances, rhythms, and phase cuts.
This page keeps that seed alive while placing it inside a stricter MDL×DCC frame: build candidates first, test hard second.
Generate π-based integer sequences and compare them with Fibonacci and powers of two. This is not proof of anything; it is a candidate factory.
The chart uses a logarithmic scale so slow and fast sequences can be compared.
Trading remains a useful test bed because price data is noisy and easy to fool ourselves with. If PiX survives there against controls, that is interesting. But PiX should not be defined by trading.
1/π — shallow retracement.π/6 — radian / 30° fraction.2/π — near Fibonacci 61.8%, but not identical.π/4 — deep retracement / 45° fraction.PiT is the trading-only branch of PiX. It does not claim that π predicts price. It tests whether π-derived spacing, timing, sizing, and profit/rescue rhythms improve specific hedge-rescue decision slots against matched controls.
2026-06-12 12:00 → 2026-06-15 09:00 · offline replay only · 10 workers
out of 783,552 planned tasks · 98.84% · 15 complete cycles + 1 deadline partial
Pi candidates that passed survival, profit floor, no-slot, random-p95, and matched-control gates
USDT final-MTM/full-close for C009 mega_rescue_timing · phi_pi_cluster · survival 100% · ruin 0
| Finding | Observed result | Interpretation |
|---|---|---|
| Arena health | Self-test passed; final extractor wrote 16 roots with 0 warnings; the last cycle stopped because the wall-clock deadline was reached. | The weekend runner, resume/state, and slim extraction are healthy enough for the next stage. |
| Best Pi seed | mega_rescue_timing / phi_pi_cluster in C009 produced about +22.927 USDT final-MTM/full-close, with survival 100% and ruin 0. | This is the first serious PiT seed worth isolating, not just a pretty pattern. |
| Control pressure | Nearest random and matched controls were close in C009; in C015 partial, a geometric matched-control candidate beat the Pi candidate. | The signal is weak-to-moderate positive, not proven. Geometric and random controls remain dangerous competitors. |
| Coverage gap | Detailed policy manifests still under-covered some surfaces, especially rescue_add_sizing. | The next build must enforce balanced slot quotas before each cycle starts. |
| Trading status | No live exchange, no API keys, no orders; this was offline replay research. | Interesting research signal, but not live-ready and not financial advice. |
phi_pi_cluster.Time zones are not automatically a signal. They are windows of increased attention. The test asks whether reversal, acceleration, range break, or rescue probability increases after π time zones.
Every PiX candidate must compete against Fibonacci, e, √2, √3, Champernowne, uniform, random, shuffled-π, and no-generator controls. Without this, coincidence will look like signal.
A beautiful hit does not count. Only L(generator + params + residual) against the best baseline counts. If the residual eats the saving, PiX loses.
PiX should first compete in small slots: chunk predictor, residual basis, restart rhythm, tile/mask generator, TP ladder, rescue window, veto cone.
| Domain | PiX candidate | Control | Success meaning |
|---|---|---|---|
| 8Z compression | π offset + transform + residual | raw, zstd, lzma, random stream | smaller verified encoding or equal-quality lossy residual |
| Image/audio | π texture/noise basis | random noise + standard codecs | smaller residual at same PSNR/SSIM/SNR |
| TSP/Sudoku/ARC | π restart / period / tile / mask candidates | fixed + Fib + random | better solve rate, MDL score, or test prediction |
| Trading slot | π retracement/time/add spacing | Fib + uniform + random | lower MAE, better rescue, better profit/risk |
PiX becomes one MATH/generator family in 8Z: constant_id=π, offset, transform, residual, hash verification, MDL gate.
DCC learns when to use π, when to use φ, e, √2, random, a domain generator, or nothing. PiX must be a candidate, not a favorite.
Once results exist: “PiX: π as a falsifiable MDL candidate generator across compression, search, signals, and decision slots.”