Facts first
Code computes exact features and checks data sufficiency before an uncertain Judge is asked anything.
A research architecture · Interactive explainer
Follow the evidence.
Then decide what thinks next.
Dynamic Judge Network asks whether small, diverse evaluators can form a sparse reasoning path: each wave of evidence helps choose the next Judges, or tells the system to stop.
The High/Low benchmark and the network controller are research plans. The interactive model below is a hand-written illustration; it makes no Jev calls and reports no measured accuracy, latency, or savings.
01 / INPUTMarket state
02 / CODEValidity gate
03 / JUDGESWave 0
04 / SUMMARYEvidence State
05 / CONTROLStop or ask more?
06 / OUTPUTHIGH · LOW · SKIP
Conceptual path · later waves run only when the evidence calls for them
The idea
A network of judgments,
not a deeper single model.
DJN studies when additional judgment is useful, which specialist should run next, and when the evidence supports no action.
Judges in waves
Independent Judges can share a request. Their answers form a compact Evidence State that guides the next wave.
Stopping is a choice
Counter-evidence may change a direction. Missing data, disagreement, or low actionability can produce SKIP.
Interactive toy model
Put the decision path under pressure.
Change synthetic inputs and compare a fixed routing rule with a proposed adaptive threshold. Both are illustrative, deterministic rules, not calibrated probabilities or benchmark results.
NO API CALLS · NO REAL MARKET DATA
02 / RESULTWave 0 stop
Synthetic decision
HIGHThe evidence supports a directional decision.
This result illustrates a rule. It is not a trading signal or a measured prediction.
03 / PATH
1 waveWhat ran, and why
GATEData checkPassed
WAVE 0Initial Judges3 activated
STATEEvidence StateSummary ready
WAVE 1Counter-evidenceNot needed
Wave 1 uncertainty check: 0.25 versus threshold 0.55.
04 / SUMMARY
Evidence State
Wave 0 forms a compact summary. A later Judge can receive that snapshot with selected raw features.
After Wave 0 · used for routing
After the final wave · used for decision
Direction +0.78
Disagreement 0.08
Uncertainty 0.25
Direction score+0.78
Disagreement0.08
Uncertainty0.25
Actionability0.91
05 / ACCOUNTING
Judge accounting
Question and path counts differ when an answer is fetched speculatively.
3Requested
3Fetched
3Activated
3Used
1 simulated provider request across 1 wave.
06 / BASELINE
Fixed set vs. this path
A toy fixed set asks six questions on every valid input: Trend, Momentum, Noise, Mean Reversion, Counter-HIGH, and Counter-LOW. The dynamic path asks only the questions shown above. Counts are not accuracy or cost measurements.
Fixed set
6
This path
3
07 / LOG
Decision trace
Inspect the toy routing rules
These deliberately simple rules make the illustration reproducible. Their thresholds and weights are not calibrated to market data.
- Gate: completeness below 40% returns SKIP before any Judge question.
- Wave 0: Trend, Momentum, and Noise run together. Direction = 0.55 × Trend + 0.45 × Momentum; signed inputs are scaled to −1…+1. Disagreement = |Trend − Momentum| ÷ 2. Uncertainty = clip(1 − |direction| + 0.25 × Noise + 0.125 × disagreement).
- Route: the fixed policy runs Wave 1 when uncertainty reaches 0.55, reversal risk reaches 0.55, or disagreement reaches 0.45. The adaptive hypothesis changes only the uncertainty threshold to clip(0.55 + 0.05 × |direction| − 0.12 × reversal − 0.10 × disagreement, 0.40, 0.65). The two risk triggers remain. A one-wave budget returns SKIP when more evidence is needed.
- Wave 1: a Counter-HIGH or Counter-LOW Judge challenges the preliminary direction. Mean Reversion joins when reversal risk is at least 0.45 or noise at least 0.55. Counter strength = clip(0.5 × Reversal + 0.25 × Noise + 0.25 × disagreement); it reduces the preliminary score, with an additional reversal adjustment when Mean Reversion runs.
- Stop: the final direction is used only when its score magnitude is at least 0.34, uncertainty at most 0.68, and actionability at least 0.55; otherwise the result is SKIP. Actionability falls as noise, reversal risk, disagreement, missing data, or counter-evidence rises.
Read the counts carefully
A question asked is not always evidence used.
Requested
A question was sent to the illustrative provider boundary. A failed call would still count here.
Fetched
An answer returned. In this toy, every request succeeds; real experiments must log failures.
Activated
The controller placed the Judge on the reasoning path. A prefetched answer may never be activated.
Used
The evidence affects routing or the final decision. All activated Judges in this toy are used.
Research status
The architecture follows evidence.
The repository currently has a provider-agnostic Judge contract, a Jev stage executor, and SQLite records for provider calls and Judge answers. It does not yet have the network controller, benchmark pipeline, or end-to-end experiment results.
A SpikingBrain-inspired research input proposes adaptive Judge thresholds, group routing, and bounded extra waves as separate future tests. The interactive toy now previews one deterministic adaptive threshold alongside its fixed baseline. Group routing and burst waves remain future hypotheses; the preview does not implement spiking neurons or claim measured gains.
NOWFoundation
Judge, executor, and storage contracts are implemented.
NEXTFixed baselines
Experiment 1 will establish a simple comparison before dynamic routing.
RESEARCHDynamic paths
Test waves, counter-evidence, stopping, Evidence State, and cost one mechanism at a time.