Every search process, left uncontrolled, collapses into seizure (locked into one solution, too much order) or dissolves into noise (random thrashing, too much chaos). The Digital Claustrum Controller measures the complexity of the search dynamics in real time and adjusts a single coupling variable u to hold the process in the productive zone between the two failure modes.
The sensor is Lempel-Ziv complexity. The control law is band regulation. The representation is one scalar. Twenty lines of code. Zero free parameters.
Every N ticks: measure LZ of the symbol buffer → if below band, decrease u (explore more) → if above band, increase u (exploit more) → map u to system knobs (breadth, strictness, aggression). That’s it. The same loop runs in seven domains.
Lempel-Ziv scans a sequence of symbols and counts how many new patterns it must discover to describe the entire sequence. Fewer new patterns means more regularity (compressible, low complexity). More new patterns means more diversity (incompressible, high complexity).
It requires zero parameters. No window size to tune, no threshold to set, no learning rate. This is why it wins the MDL contest for “cheapest sensor” — its own description length is nearly zero.
Oscillator x-coordinate each tick → binarize (x > 0 = 1, else 0). Buffer: 64 bits.
All oscillators locked in sync. One pattern describes everything. System is dead.
Every bit is effectively random. LZ must create a new entry for nearly every position.
Repeating motifs with subtle variation. LZ finds reusable blocks but keeps discovering new ones. This is where the claustrum holds the system.
Each step: did the swap improve the tour? Emit 1 (yes) or 0 (no). Buffer: 64 bits.
Nothing improves. DCC must lower u: open search, try aggressive perturbations.
Improvements and failures alternate randomly. DCC must raise u: focus, stop wasting compute.
Calm periods, then bursts of improvement. Solver finds structure, exploits it, moves on. DCC holds steady.
Each bar: which of 4 generators (TREND, REVERSION, MOMENTUM, RANGE) wins the MDL arena? Emit winner ID. Buffer: 64 bars.
Same generator winning for too long. Regime may be stale. DCC reduces confidence → smaller positions.
Winner flips every bar, no structure. DCC raises u → stop trading, market has no exploitable regime.
Regimes persist with periodic shifts. LZ is intermediate. DCC keeps full position size. Trade with confidence.
EEG signal → binarize relative to median (above = 1, below = 0) → LZ per channel → combine with Phase-Locking Value (PLV) → S = k · Cn · Ψ(I)
All channels synchronized → Cn high, LZ ≈ 0. S collapses. The brain is locked.
Cn low, LZ low. Neither coherent nor complex. S collapses. Nobody home.
Cn moderately high AND LZ moderately high. Structured diversity. S is in the conscious band. This is what the claustrum maintains.
The same algorithm measures the same property: structured diversity of a symbol stream. Only the symbols change (bits, swap outcomes, generator IDs, neural signals). Only the response changes (coupling adjustment, search strategy, position sizing, arousal). The principle is universal.
A discovery that emerged from cross-domain transfer: the same LZ sensor requires opposite polarity interpretation at different recursive levels.
In the TSP solver at Level 1, low LZ means “stuck — loosen up.” But in trading at the meta level, low LZ means “stable regime — trust it.” Same measurement. Opposite response. The inversion isn’t a bug; it’s the calibration method for multi-level DCC systems.
This partially resolves the anchor problem in DCC-7: how do you set the target band when you don’t know the “correct” complexity for a new domain? Answer: you don’t set it. The v2.5 meta-DCC self-calibrates from the 25th/75th percentile of observed LZ history, and the polarity emerges from which level of the recursion you’re on.
On uy734 (734 cities), bands self-calibrated from the initial [0, 1] range to [0.040, 0.061] — the natural complexity scale of TSP 2-opt search is far narrower than oscillator dynamics. The system found this on its own. No human tuning.
LZ is universally known. Compression engineers use it daily. Neuroscientists publish LZC papers. Information theorists cite it as a Kolmogorov proxy. But in all cases, LZ is passive. Measure, plot, publish. Next paper.
Closing the loop — feeding the LZ measurement back into the system in real time to control its behavior — is the difference between a thermometer and a thermostat. The entire world has thermometers. The DCC is a thermostat.
People who know LZ in compression don’t think about control systems. People who build control systems use PID controllers, Kalman filters, reinforcement learning — not LZ. People who measure LZC in EEG are neuroscientists, not engineers. Nobody sits at the intersection of all three simultaneously.
Machine learning went to stochastic gradient descent, not minimum description length. Rissanen’s framework is a “solved problem” in textbooks — mentioned in the model selection chapter, then everyone moves on to neural nets. Nobody builds entire systems on the MDL kernel, because “that’s not modern.”
The AI field is obsessed with feed-forward architectures: more parameters, more data, bigger models. The idea that you need a governor — a small, cheap module that watches itself — is counterintuitive in a world where the solution to everything is “add more layers.” DCC is 20 lines of code. Nobody gets tenure for 20 lines.
The 8Z research program had to solve compression, TSP, trading, authentication, and consciousness theory with limited resources. There was no luxury of specialization, no team of 50 for each domain. The constraint forced the search for one principle that works everywhere. The constraint was the discovery mechanism.
The current DCC has three replaceable components. Each has alternatives.
| Sensor | How It Works | Lopis | Trade-off |
|---|---|---|---|
| LZ (current) | Count new patterns in symbol stream | ≈ 0 | Zero free parameters. Sequential structure only. |
| SampEn | Probability of similar patterns recurring in continuous signal | +2 params (m, r) | Works on raw signals (no binarization). Needs calibration. |
| zstd ratio | Compress buffer, measure ratio | ≈ 0 | Captures all structure. Slower, version-dependent. |
| Spectral flatness | Geometric / arithmetic mean of power spectrum | ≈ 0 | Very fast (one FFT). Frequency structure only, not sequential. |
| Transfer entropy | How much agent A’s past predicts agent B’s future | +2 params | Measures inter-agent information flow. Ideal for fleet/meta-DCC. |
| Law | Mechanism | Lopis | Trade-off |
|---|---|---|---|
| Bang-bang (current) | Step u up or down if outside self-calibrating band | ≈ 0 | Zero tunable parameters. Not smooth. |
| PID | Proportional + integral + derivative response | +3 params | Smoother tracking. Must tune Kp, Ki, Kd. |
| EMA response | Exponential moving average of error signal | +1 param | Smoother than bang-bang. One free parameter (α). |
| Model Predictive | Predict N steps ahead, optimize u trajectory | +N params | Powerful when you have a dynamics model. DCC’s point is you don’t. |
| Bayesian u | Full posterior over u, updated each measurement | +prior spec | Natural uncertainty. Computationally heavier. |
| Form | What It Is | Lopis | Trade-off |
|---|---|---|---|
| Scalar (current) | One number u ∈ [0,1] | ≈ 0 | Minimal. One knob controls everything. |
| Vector | [ubreadth, udepth, uaggression, ...] | +dim | Finer control. Harder to calibrate. |
| Coupling matrix | uij for multi-agent systems | O(n²) | Exactly what the claustrum does between cortical regions. Expensive. |
LZ + bang-bang + scalar u has the shortest description length of any combination. Zero free parameters. Every alternative adds parameters that must justify themselves empirically. Per the framework’s own logic, the current architecture remains optimal until a variant demonstrates dramatically lower residual error.
The question “which DCC architecture is best?” does not require human judgment. It requires an arena.
DCC is itself a model — a model of the search dynamics. It has a description (sensor + law + representation) and a performance (how well it holds the system in the productive zone). MDL evaluates models. Therefore:
Run 12 DCC variants on wi29, dj38, qa194, uy734. MDL picks the winner. Per recursive level. Level 1 might prefer LZ + bang-bang. Level 3 might prefer transfer entropy + EMA. The arena decides, not the architect.
Now the final step. Who governs the meta-arena? Who decides how much compute to spend testing DCC variants? Who prevents the meta-search from locking into one variant too early (seizure) or endlessly cycling through all of them (noise)?
DCC.
This is Principle 17 applied to Principle 5. Never hardcode what the system can learn (P17). Let DCC control it (P5). Therefore: let the system learn which DCC to use, governed by DCC, scored by MDL.
MDL + DCC + recursion = a system that does not require an external architect to select its own architecture. The only external input is the principle itself. Everything else emerges.
If fractal MDL+DCC autonomously discovers the optimal search strategy for each problem instance, and if this discovery process scales polynomially — because the DCC prevents combinatorial explosion by governing the search budget at every recursive level — then we have an experimental program that can investigate whether a constructive path exists.
This is not a proof. It is a research direction: build the self-selecting governor, run it on NP-hard instances of increasing size, and measure whether the meta-DCC consistently finds polynomial-time search strategies for structured instances. The TSP solver is the testbed. The 8Z-RP already achieves 0.46% gap on uy734 (734 cities) with self-calibrating DCC. The question is whether this scales.
The claim is not “P=NP.” The claim is: a self-governing MDL+DCC system might empirically demonstrate that structured NP-hard instances admit polynomial-time solutions when the search itself is governed by the right recursive controller. This is testable.
The arena ran. 286 variants tested. MDL decided. Here is what actually happened.
Top 11 variants by L_total. Variants at 0.00 are exact-optimal across all tested instances.
| Sensor | Law | L_total | Source |
|---|---|---|---|
| CUSUM | BB | 0.00 | 1954 algorithm |
| LZ_binary | ADSR | 0.00 | Round 2 LLMs |
| tri_area | ADSR | 0.00 | BD, 1 AM, from bed |
| tri_compact | BB | 0.00 | BD, 1 AM, from bed |
| LZ_dual | ADSR | 1.00 | Round 2 LLMs |
| LZ_dual | ADSR+late | 1.00 | LLM feedback |
| frozen_edge | ADSR | 2.00 | Round 2 LLMs |
| LZ_binary+lk | ADSR | 3.00 | LLM feedback + literature |
| LZ_dual+cross | ADSR | 4.00 | LLM feedback |
| LZ_binary | CDPID | 5.00 | Round 2 LLMs |
| fold_dist | EMA | 8.00 | Wild combination |
The prediction was wrong. LZ-BB did NOT win across all instances. The winners included a 1954 algorithm (CUSUM), two geometric sensors proposed by a human with no CS training (tri_area, tri_compact), and combinations that no single model proposed. MDL decided. As promised.