Decision Layer

Signals, scoring, and decision metrics for a self-observing validator

TSUNAGI continuously measures its own behavior and the network it operates on. Every forge opportunity is evaluated by a layered pipeline that normalizes raw observations, tracks their drift and confidence, and produces decision-theoretic metrics.

This is an observability layer, not autonomous control. Currently observe-only: decisions are logged but never override the deterministic forge path. Future modes are documented below.

Metrics exposed per leader slot: posterior (Beta(α,β) probability), EV (expected value), Kelly (allocation fraction), LLR (log-likelihood ratio). Correlation with real outcomes is being measured.

EVENTS  →  SIGNALS  →  AGGREGATION  →  DECISION  →  (POLICY — future)

Each layer is deterministic, observable, and independently verifiable.

Signals

Six normalized signals in [0,1], each with a traceable raw source, a named normalization method, and per-signal confidence derived from the variance of its recent history window.

Signal	Meaning	Normalization
parent_freshness	How far our selected parent is behind the probe-observed tip	linear_gap_to_20
latency	Wall-clock from probe header to parent selection	linear_ms_to_1000_inverse
acceptance	Recent block acceptance rate	rate_div_1000
probe_advantage	Percentage of updates where probe beat primary	advantage_div_100
fork_density	Recent fork events in last 100 slots	linear_forks_to_10_inverse
stability	1 − rejected/built	1_minus_rejected_over_built

Every signal is logged with its raw value, the normalization method applied, and the resulting confidence score. Drift is tracked against a 50-sample rolling mean and emitted when it exceeds the alert threshold. Cross-signal conflicts (e.g. latency low but gap high) are flagged explicitly.

Execution Quality

Real-time measurement of forge readiness. Combines three metrics into a single score with a HIGH / MEDIUM / LOW rating attached to every pending leader slot.

INPUT

Parent Gap

Distance in blocks between the probe-observed tip and the selected forge parent. Lower is better.

INPUT

Probe Advantage

Rolling win-rate of the secondary probe over the primary follower across the last 100 header updates.

INPUT

Latency

Milliseconds from probe header arrival to forge-path parent selection. Low values indicate fresh data reached the forge.

OUTPUT

Score & Rating

score = (advantage × 2) − gap − (latency / 10). Bands: >50 HIGH, >10 MEDIUM, else LOW.

Peer Scoring

Up to 32 peers are tracked with three weighted components. The best candidate is logged as advisory; actual parent selection remains deterministic (max block_no across sources).

Component	Weight	Source
freshness	0.50	% of updates that were fresh wins
latency	0.30	Inverse-normalized average latency
reliability	0.20	connects / (connects + disconnects)

Decision Metrics

Outcomes feed a Beta(α, β) posterior that updates with every success or failure. Four decision-theoretic quantities are derived and logged periodically.

POSTERIOR

Bayesian α / β

α counts accepted blocks, β counts rejected. Posterior = α / (α + β). Starts from a uniform prior Beta(1,1).

Expected Value

EV = posterior × reward − (1 − posterior) × cost. Negative values trigger an unfavorable advisory.

KELLY

Kelly Criterion

kelly = posterior − (1 − posterior) / odds. Clamped to safe bounds. Indicates how strongly conditions favor action.

LLR

Log-Likelihood Ratio

LLR = log(posterior / (1 − posterior)). Symmetric around zero; positive favors the success hypothesis.

Current Mode

OBSERVE-ONLY All decision outputs are emitted to logs and exposed via /status/full. The forge path uses the same deterministic eligibility check it always has — VRF threshold evaluation, KES signing, and parent-hash linkage. The decision layer never vetoes or alters a forge.

Future Modes

Two further modes are defined but not activated:

Mode	Status	Behavior
observe	ACTIVE	Emit signals and decisions, no action taken.
adaptive	FUTURE	Self-tune operational parameters (forge delay, peer preference) within deterministic bounds.
autonomous	FUTURE	Trust-native participation: weighting received data by measured peer reliability.

Exposed Interfaces

Full state is available at /status/full. Upcoming leader slots are available at /schedule?hours=12&limit=32.

The decision layer is not a consensus change. It is an observability and decision-reporting surface on top of a deterministic forge pipeline.