4.3 KiB
Surf MCP Fusion Gate
Status: HOLD / external-tool fusion gate Authority: workbench integration note; not calibration proof Related:
docs/gcl/AVMFixedPointComplianceRemediation.mddocs/gcl/ThreeLayerBuilderJudgeWardenGate.mddocs/gcl/NanokernelCompressionOrchestrationUndici.mddocs/gcl/RunawayDigitalCellDivisionDoctrine.mddocs/gcl/FederatedNanokernelSwarmDoctrine.mddocs/gcl/SovereignHyperEquationSynthesisIngestion.md
Purpose
This note defines how Surf MCP or Google MCP-style external tool surfaces may fuse into the stack.
External MCP tools can help finish the orchestration layer, but they do not automatically calibrate AVM, prove the Hyper Equation, validate hardware parity, or promote claims.
Core rule
MCP tools are external route phenotypes.
They are not roots of truth.
Any MCP-sourced action or result must pass:
Builder
-> construct tool request and declare capability
Judge
-> validate schema, result shape, receipts, and claim scope
Warden
-> enforce egress, credentials, destructive-action limits, redaction, and inheritance policy
Adaptive AngrySphinx
-> scar/throttle/quarantine repeated bad patterns
Fusion placement
MCP tool surface
-> nanokernel orchestration route
-> dispatch policy
-> external call
-> compressed receipt
-> Judge replay or schema validation
-> Warden authority decision
-> ENE inheritance or HOLD/quarantine/refusal
What Surf MCP can contribute
Depending on which Surf MCP is meant, it may provide:
browser automation / web interaction
API data fetches
OpenAPI-derived tools
external search or data grounding
crypto/API telemetry
web-surface inspection
These are useful as tool phenotypes.
They are not formal receipts unless the stack records:
tool identity
tool version
request shape hash
response shape hash
credential redaction status
policy hash
timestamp / sampling context
Warden decision
claim scope
Google MCP placement
Google MCP surfaces may expose cloud/product APIs through MCP endpoints.
They must be treated as external capability routes:
Google MCP call
-> declared capability
-> credential boundary
-> request receipt
-> response receipt
-> Warden authority limit
A successful Google MCP call proves only that the call returned under the declared conditions.
It does not prove:
AVM calibration
Lean theorem validity
hardware parity
Hyper Equation correctness
formal core lock
AVM calibration boundary
The transcript claims AVM is CALIBRATED, but that status requires repo-visible receipts.
Required receipt files or equivalent:
FixedPoint.lean with Q0_16 and Q16_16 definitions
AVM step / bind semantics file
validator script rejecting float/f32/f64/double in core semantics
lake build log or CI receipt
cross-target determinism test receipt
hardware parity test plan or log
Until those are present and checked:
AVM status remains HOLD / CLAIM_BUNDLE_PENDING_RECEIPTS
External float boundary
MCP tools may return JSON numbers or platform-native numeric strings.
These must not enter AVM authority as floats.
MCP numeric output
-> boundary parser
-> declared quantization policy
-> Q0_16 or Q16_16 atom
-> conversion receipt
-> AVM core
Claim-state mapping
MCP response received
-> HOLD
MCP response with schema validation and receipt
-> V_scope
MCP response replayed / independently checked
-> REVIEWED candidate
MCP response plus Lean theorem / build / hardware logs
-> may support CALIBRATED claim only for the exact theorem/test scope
Hardware parity gate
Hardware parity is not complete until the router/FPGA target produces comparable receipts.
Required:
AVM reference trace
Verilog/router trace
input vector set
state hash comparison
fixed-point arithmetic equivalence
UART packet boundary test
partial-commit rejection test
metastability-safe boundary documentation
Operating sentence
Surf MCP and Google MCP can finish the fusion only as external route phenotypes: they may provide tool access, data, browser/API interaction, and receipts, but every result must pass Builder/Judge/Warden, fixed-point boundary conversion, compressed receipt capture, and scoped claim promotion before it can influence AVM, ENE, GCL, MOIM, or hardware parity.