MESH-NRP Learning Observability

src/kernel/adaptiveWeights.ts — monad.ai v2.1+

---

Start here

If you are new to the learning loop, read learning-loop.md first. It explains what the loop does, how the weight update formula works, and how to interpret the numbers you see here.

This document is the operator reference — what to run, what to watch, and what to do when something looks wrong.

---

Overview

The NRP adaptive scoring system learns scorer weights from live request outcomes. The learning loop is intentionally slow (α = 0.01) so it takes hundreds of requests before weights shift meaningfully. This guide covers how to observe the learning system in production, interpret health signals, and diagnose problems.

Running the smoke tests

Before reading live weights in production, verify the learning pipeline is wired correctly by running the smoke tests:

# All NRP tests (22 files, 235 tests):
npm test

# Only the learning loop smoke tests:
npx vitest run tests/NRP/learningLoop.test.ts

The smoke tests verify 7 properties of the learning pipeline without requiring a live server. If they pass, the pipeline from recordDecision → correlateOutcome → updateAdaptiveWeights is correctly wired. See learning-loop.md § Running the smoke tests for what each section proves.

---

Live weight monitor

tsx scripts/watch-weights.ts
tsx scripts/watch-weights.ts --port 8282 --interval 3000
MONAD_PORT=8282 tsx scripts/watch-weights.ts

The monitor polls GET /.mesh/weights and renders a color-coded table:

NRP Adaptive Weights  2026-05-05T12:00:00.000Z
http://localhost:8161/.mesh/weights  updates: 142  stable: false

Scorer       Current    Default    Delta
───────────────────────────────────────────────
latency      0.2280     0.2500     -0.0220
recency      0.3510     0.3500     +0.0010
resonance    0.4210     0.4000     +0.0210

last update: 3s ago

✓  learning loop healthy

Color key:

• Green delta — scorer is being reinforced by successful outcomes
• Red delta — scorer was associated with failed decisions
• Dim delta — movement < 0.002 (essentially unchanged)
• Bold current — weight has moved from default

---

HTTP endpoint

curl http://localhost:8161/.mesh/weights | jq

{
  "ok": true,
  "current":       { "latency": 0.228, "recency": 0.351, "resonance": 0.421 },
  "defaults":      { "latency": 0.250, "recency": 0.350, "resonance": 0.400 },
  "delta":         { "latency": -0.022, "recency": 0.001, "resonance": 0.021 },
  "updateCount":   142,
  "lastUpdatedAt": 1746412800000,
  "stable":        false,
  "health": {
    "dominantScorer": null,
    "deadScorer":     null,
    "oscillation":    false,
    "noLearning":     false
  },
  "_hint": "delta = current − defaults. Positive: scorer reinforced by good outcomes. Negative: penalized by failures."
}

---

Console logging

# Log every weight update to console after each forwarded request:
MONAD_DEBUG_WEIGHTS=1 npm run dev
# [weights] latency: 0.228 (Δ-0.022), recency: 0.351 (Δ+0.001), resonance: 0.421 (Δ+0.021) — updates: 142 reward: 0.850

---

Health signals

The health object in the weight report contains four diagnostic flags. None triggers an automatic action — they are informational.

stable: true

All deltas are within 5% of their default weight. This is expected at startup and on homogeneous meshes (all nodes behave identically, so no scorer is consistently better than another).

Not a problem unless updateCount > 100 and you expected the system to learn something. In that case, check noLearning.

---

dominantScorer: "resonance" (example)

One scorer has captured more than 70% of the total weight. The other scorers are nearly ignored.

Interpretation:

• Could be correct: if one signal (e.g., resonance) is genuinely the strongest predictor of success in your mesh, the learning loop will correctly up-weight it.
• Could be overfitting: if the mesh went through a period where all failures came from low-resonance nodes, resonance gets over-rewarded even if latency or recency also correlates.

Diagnosis: run tsx scripts/analyze-decisions.ts ~/.monad/decisions.jsonl and look at the scorer contribution by outcome table. If the delta is negative for the dominant scorer on failures, the system is self-correcting. If the delta is consistently positive for both success and failure, there may be confounding.

Remediation: use per-claim weight overrides to cap the scorer temporarily:

_.mesh.monads.frank.claimed["suis-macbook-air.local"]._weight_resonance = 0.4

---

deadScorer: "latency" (example)

A scorer's learned weight has dropped to WEIGHT_MIN * 2 (0.02) or below — it is barely contributing to selection.

Interpretation: The scorer was consistently associated with failures. This can be correct (latency is not predictive for a CPU-bound workload) or incorrect (an early burst of timeouts caused the learning loop to penalize latency forever).

Remediation:

• Inject a per-claim floor: _weight_latency: 0.1 in the claim metadata
• Reset learned weights and let the system relearn from a clean state:

  # No HTTP reset endpoint exists yet — restart the daemon to clear in-memory weights.
  # Stored weights in _.mesh.adaptiveWeights persist across restarts.

• Raise LEARNING_RATE temporarily to accelerate recovery (requires restart)

---

oscillation: true

The recent reward signal alternates sign more than 40% of the time across the last 10 rewards.

Interpretation: The system is receiving contradictory signal — successful and failed requests are alternating. This makes weight learning unstable (each update partially cancels the previous one).

Common causes:

• Two similarly-scored nodes with opposite reliability: the learning loop wins with node A, then loses with node B, then wins with A again
• MONAD_EXPLORATION_RATE is too high relative to mesh size: forced exploration through the runner-up triggers failures which flip sign
• Transient infrastructure instability (nodes rebooting, network flapping)

Remediation:

• Reduce MONAD_EXPLORATION_RATE temporarily
• Check scripts/analyze-decisions.ts — "runner-up on failure" section shows if alternating node selection is the source
• If infrastructure is the cause, wait for stability before drawing conclusions

---

noLearning: true

More than 10 gradient updates have been applied but no weight has moved more than 0.002 from its default.

Interpretation: The bridge is calling updateAdaptiveWeights but the breakdown contributions are all near zero. This usually means:

1. No mesh-claim selections: all requests are resolved via name-selector, not scored claimants. Check MONAD_DEBUG_SCORING=1 output — if you see no [scoring] lines, no scored decisions are being made.
2. Zero scorer values: all claimants have zero latency, recency, and resonance scores (fresh nodes with no history). Contributions = value × weight; if value is 0, delta is 0.
3. Bridge integration gap: correlateOutcome is not being called after forwards. Check bridgeHandler.ts for the if (decisionId) guard.

Verification:

MONAD_DEBUG_WEIGHTS=1 npm run dev
# Should print [weights] lines after each forwarded mesh-claim request

---

Reward formula

Every forwarded request produces a reward that drives the weight update:

rewardQuality = ok ? 1.0 : −1.0
rewardLatency = ok ? max(0, 1 − latencyMs / 5000) : 0

reward = 0.7 × rewardQuality + 0.3 × rewardLatency

Outcome	Latency	reward
success	0 ms	1.000
success	2 500 ms	0.850
success	5 000 ms	0.700
failure	any	−0.700

The 0.7/0.3 split ensures correctness errors move weights more decisively than latency variance. Override with MONAD_LEARNING_QUALITY_WEIGHT:

# Weight quality at 90%, latency at 10%:
MONAD_LEARNING_QUALITY_WEIGHT=0.9 npm run dev

---

Weight update rule

Δweight = α × reward × contribution
new_weight = max(WEIGHT_MIN, old_weight + Δweight)

• α = LEARNING_RATE = 0.01 — controls convergence speed
• WEIGHT_MIN = 0.01 — no scorer falls below 1% influence
• contribution = scorer_value × normalized_weight — how much this scorer influenced the winning selection

Weight resolution priority (highest first):

Priority	Source	Notes
1	meta._weight_<name>	Per-claim explicit override
2	ctx.adaptiveWeights[name]	Globally learned prior
3	scorer.defaultWeight	Hardcoded fallback

---

Environment variables

Variable	Default	Effect
MONAD_DEBUG_WEIGHTS=1	off	Log weight update after every forward
MONAD_DEBUG_SCORING=1	off	Log every scoring decision to console
MONAD_SCORE_SAMPLE_RATE=0.01	0	Sample ~1% of decisions to console
MONAD_SCORE_MARGIN_THRESHOLD=0.05	0.05	Always log fragile decisions
MONAD_EXPLORATION_RATE=0.15	0	Route ~15% of fragile decisions to runner-up
MONAD_DECISION_LOG=/path/decisions.jsonl	unset	Enable JSONL decision log
MONAD_LEARNING_QUALITY_WEIGHT=0.7	0.7	Quality vs latency blend in reward
MONAD_MESH_STALE_MS=300000	300000	Staleness cutoff for claimants

---

Offline analysis

MONAD_DECISION_LOG=~/.monad/decisions.jsonl npm run dev

# After accumulating traffic:
tsx scripts/analyze-decisions.ts ~/.monad/decisions.jsonl

The analyzer complements the live weight monitor: the monitor shows the current weight state, the analyzer explains why weights moved (which scorer dimensions correlate with success vs failure).

See scoring.md for full analyzer output documentation.

---

Phase 9 — namespace-maturity blending (implemented)

The adaptive learner now uses a global prior plus namespace-local posterior weights:

_.mesh.adaptiveWeights          global prior
_.mesh.nsWeights.<namespace>    namespace-local posterior

Reads use a maturity blend:

maturity = min(1, nsSamples / 200)
selectionWeights = global × (1 − maturity) + namespace × maturity

• samples = 0 → 100% global (bootstrap)
• samples = 100 → 50% blend
• samples = 200+ → 100% namespace for selection; global still receives 5% background signal during attribution

The global prior is never fully disabled so cross-namespace trends (e.g., a latency regression affecting all routes) still propagate upward.

Observe a namespace-specific blend:

curl "http://localhost:8161/.mesh/weights?namespace=suis-macbook-air.local" | jq
tsx scripts/watch-weights.ts --namespace suis-macbook-air.local