NRP Implementation Status

monad.ai v2.2.0 — current local implementation status

What Is Implemented

The current NRP work is no longer only an HTTP resolver. It is a mesh-aware runtime path with optional Total Monad Synthesis (MONAD_SYNTHESIS_ENABLED=1):

Single-forward (default, Phase 1–9):

  1. Parse me://namespace:read/path
  2. Discover monads that claim the namespace
  3. Filter by optional selector constraints
  4. Score eligible monads
  5. Forward to the selected endpoint
  6. Learn from the request outcome
  7. Explain and optionally log the decision

Synthesis path (Phase 10, opt-in):

  1. Parse me://namespace:read/path
  2. Discover and score all eligible claimants
  3. Select top-N candidates (MONAD_SYNTHESIS_MAX_CANDIDATES, default 3)
  4. Forward in parallel to all N candidates
  5. Synthesize: quorum → public, no quorum → contested, all fail → closed
  6. Log synthesis decision with divergence metadata and reward signal
  7. Expose _synthesis on the wire for audit
request
  → parseBridgeTarget
  → MONAD_SYNTHESIS_ENABLED?
      no  → selectMeshClaimant → forward → correlateOutcome
      yes → selectMeshClaimants (top-N)
          → forwardAll (parallel)
          → synthesize()
          → correlateSynthesisOutcome
          → _synthesis in response

Phase Summary

Phase Status Result
Phase 1 done Namespace reads/writes and claim/open lifecycle
Phase 2 done Self mapping, monad identity, __surface, disclosure envelopes
Phase 3 done Mesh index, /.mesh/monads, /.mesh/resolve, bridge forwarding
Phase 4 done Selector-aware mesh routing ([device:macbook], [host:edge])
Phase 5 done Production scoring engine with open .me claim metadata
Phase 5.5 done Score introspection, runner-up tracking, margin logging
Phase 5.6 done Decision log, biased sampling, offline analyzer
Phase 6 done Continuous reward and epsilon-greedy exploration for fragile choices
Phase 7 done Global adaptive weights and live /.mesh/weights observability
Phase 8 done Patch bay: controlled feature composition with learned weights
Phase 9 done Namespace-local weights with maturity blending and global background learning
Phase 10 done Total Monad Synthesis: multi-candidate parallel forward, quorum/contested/closed, synthesis decision log

Current Runtime Surfaces

Endpoint Purpose
GET /<path> Read a path in the namespace selected by Host
POST / Write/claim/open depending on body operation
GET /__surface Self description and monad continuity proof
GET /.mesh/monads List locally known and CLI-known monads
GET /.mesh/resolve?namespace=... Discover monads claiming a namespace
GET /.mesh/resolve?monad=frank Discover a monad by name/id
GET /.mesh/resolve/multi?namespace=... Scored candidate list for synthesis (Phase 10 observability)
GET /.mesh/weights Inspect global adaptive scorer weights
GET /.mesh/weights?namespace=... Inspect namespace-local weights and blended request weights
GET /resolve?target=me://ns:read/path Bridge a canonical NRP target (single or synthesis depending on flag)

Current Selection Model

The selector is not part of the semantic path. It constrains execution.

me://frank.local[device:macbook]:read/profile

Resolution:

  1. semantic namespace: frank.local
  2. semantic path: profile
  3. execution constraint: device:macbook
  4. winner: highest scored eligible monad, unless low-margin exploration fires

Current Learning Model

The fast index remains simple:

claimed_namespaces: string[]
tags: string[]
type: "desktop" | "mobile" | "server" | "browser-tab" | "node"

The open .me subtree holds rich learned metadata:

_.mesh.monads.<id>.claimed.<namespace> = {
  resonance: 42,
  effectiveResonance: 31.5,
  avgLatencyMs: 85,
  forwardCount: 100,
  failureCount: 4,
  geopoliticalZone: "mx-east",
  energyProfile: "low-power"
}

The index answers: who could answer?

The scoring engine answers: who should answer?

Adaptive weights now use two layers:

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

Reads blend them by namespace maturity:

maturity = min(1, sampleCount / 200)
weights = global * (1 - maturity) + namespace * maturity

Writes split the same reward gradient:

globalShare = max(0.05, 1 - maturity)
nsShare = maturity

The global prior remains alive at full namespace maturity, so cross-namespace trends still propagate.

Operational Flags

Env var Default Effect
MONAD_MESH_STALE_MS 300000 Exclude stale monads from mesh claim selection
MONAD_DEBUG_SCORING unset Log every scoring decision when set to 1
MONAD_SCORE_SAMPLE_RATE 0 Randomly log a fraction of scoring decisions
MONAD_SCORE_MARGIN_THRESHOLD 0.05 Always log fragile low-margin decisions
MONAD_EXPLORATION_RATE 0 Route some fragile decisions to runner-up
MONAD_DECISION_LOG unset JSONL path for correlated decision outcomes (single + synthesis)
MONAD_LEARNING_QUALITY_WEIGHT 0.7 Quality/latency split in reward
MONAD_DEBUG_WEIGHTS unset Log adaptive weight changes after every forward
MONAD_SYNTHESIS_ENABLED 0 Enable Phase 10 Total Monad Synthesis path (1 to activate)
MONAD_SYNTHESIS_MAX_CANDIDATES 3 Max claimants queried in parallel during synthesis
MONAD_SYNTHESIS_QUORUM_THRESHOLD 0.5 Fraction required for quorum (strict majority — tie = contested)
MONAD_SYNTHESIS_MIN_RELATIVE_SCORE 0.8 Min score relative to winner to qualify as synthesis candidate

Verification

Current regression suite:

31 test files
391 tests

Primary commands:

npm run build
npm test
npm run docs:api

Decision analysis:

MONAD_DECISION_LOG=~/.monad/decisions.jsonl npm run dev
tsx scripts/analyze-decisions.ts ~/.monad/decisions.jsonl

Not Yet Implemented

These remain design targets, not production guarantees: