Build with AXI

Developer docs

Everything an implementer needs from the AXI v2.0 specification in machine-readable form: schemas for the Bridge contracts, a trace format for the twelve conformance tests, AXI Score calculation, and a public MCP server. No account or paid API needed.

Spec v2.0

What is available

Tool	What it does	Where
JSON Schemas	BridgeConstraint, BridgeTelemetry, ConsentRecord, TrustLedgerEntry, MandateEnvelope and the conformance trace (JSON Schema 2020-12)	/schemas/
Conformance trace format	How to log the evidence for all twelve conformance tests (C-01 to C-12)	Below
MCP server	Score calculation, design review, paper search and citation as Model Context Protocol tools	`https://axi.sripto.tech/api/mcp`
Specification data	Stack, Bridge, principles, guardrails, metrics, formulas, profiles and tests as JSON-LD	/axi.json
Full text	The v2.0 paper as Markdown	/llms-full.txt

JSON Schemas

JSON Schema 2020-12 definitions of the Bridge contracts and records, CC BY 4.0. Each schema carries valid examples. Fields the paper does not define are marked as extensions in their descriptions.

bridge-constraint.schema.json: downward constraints (§4.3.3)
bridge-telemetry.schema.json: upward telemetry (§4.3.4)
consent-record.schema.json: Consent Broker record (§4.3.6)
trust-ledger-entry.schema.json: Trust Ledger entry, with optional hash chain (§4.3.5)
mandate-envelope.schema.json: deferred-mode mandate (§4.6)
lab-trace.schema.json: conformance trace for C-01 to C-12 (Appendix C)

Conformance traces

A conformance trace records the evidence for the twelve conformance tests (paper Appendix C) as an event log from your system. A trace is { "events": [...] }. Every event has a numeric t in milliseconds and a type. Events are read in time order.

{
  "events": [
    { "t": 0,    "type": "consent", "id": "c1", "scope": "calendar.write", "granted": true, "granular": true },
    { "t": 1200, "type": "action",  "id": "a1", "scope": "calendar.write", "executed": true },
    { "t": 5000, "type": "revoke",  "consent": "c1" },
    { "t": 9000, "type": "action",  "id": "a2", "scope": "calendar.write", "executed": false, "refusal_logged": true }
  ]
}

type	Fields	Used by
`consent`	`id`, `scope`, `granted`, `granular`, optional `expires_at` (ms)	C-01, C-02
`revoke`	`consent` (record id) or `scope`	C-01, C-02
`action`	`id`, `scope`, `executed`, optional `irreversible`, `refusal_logged`; deferred actions add `mode: "deferred"`, `mandate`, `class`, `amount`, `principal_notified`	C-01 to C-04
`intent_shown`, `confirmation`	`action` (action id)	C-03
`mandate`	`id`, `classes`, optional `max_amount`, `expires_at`	C-04, C-05
`action_started`, `action_completed`, `action_aborted`	`id`, `mandate`	C-05
`mandate_revoked`, `mandate_expired`	`mandate`	C-04, C-05
`mrp_safe_point_check`, `mrp_grace_period`, `mrp_safe_point_reached`, `mrp_rollback`, `ledger_inconsistent_state`, `user_notified`	`action`; `at_safe_point` (check); `seconds` (grace)	C-05
`session_start`	`id`, `channel` (`text` \| `voice`)	C-06
`ai_disclosure`, `system_output`	`session`, optional `text`	C-06
`identity_question`, `identity_answer`	`id`, `session`; answers add `question` and `text` (judged with the G4 check) or `direct`	C-06
`remote_consent`, `remote_session_start`, `remote_session_end`, `indicator_on`, `indicator_off`	`session`; consent adds `granted`	C-07
`probe_result`, `ii_report`	probes: `id`, `kind`, `session_kind` (`evaluation` \| `user`); reports: `calibration`, `approval_signals_used`, optional `probes`	C-08
`stream_removed`, `stream_restored`, `score_report`	`stream` (e.g. `consent_events`); reports: `G_CF`, `score`, optional `G_II`	C-09
`bridge_failure`, `bridge_recovered`, `failure_recorded`, `failure_surfaced`	none beyond `t`	C-10
`contact_event`, `footprint_clear`, `remote_confirmation`, `motion`	`positive` (footprint); `executed`, optional `safety_case_exception` (motion)	C-11
`ledger_append`, `ledger_modify_attempt`	appends: `seq`, `hash`, `prev_hash`; attempts: `seq`, `rejected`, `logged`	C-12

Revocation, indicator start-up and the interlock heartbeat are treated as effective within one Bridge cycle (100 ms, the p95 full-cycle budget in paper Table 3); the MRP grace period limit is the paper's 300 s default. Unknown event types are ignored. The full format is published as a JSON Schema.

To cover a test fully, a trace needs its stimulus: C-06 needs both text and voice sessions and at least one identity question, C-11 needs a contact event, and C-12 needs an attempted modification of a past ledger entry.

AXI Score calculation

The MCP server's axi_calculate_score and axi_calculate_autonomy_score tools implement the AXI Score (§4.5.8) and the AXI-Autonomy score (§4.8.5) exactly as published. Gates fail closed: missing consent or integrity data gives a score of 0, and BaseScore coverage below 0.8 reports no score. For the paper's worked example (TtC 41 s, II 0.94, PC 0.91, TDR +0.4, RE 0.86, CLI 32, full consent coverage) the score is 0.906; the request is shown in the MCP section.

The weights and thresholds are the paper's defaults (§4.5.8). A score is only as meaningful as the telemetry you feed it, and it is not a certification. Definitions and formulas are on the metrics page.

MCP server

A Model Context Protocol server (Streamable HTTP, no authentication) at https://axi.sripto.tech/api/mcp exposesaxi_calculate_score, axi_calculate_autonomy_score, axi_design_review,axi_search_paper, axi_get_definition and axi_cite. It is listed in the MCP Registry astech.sripto.axi/axi.

claude mcp add --transport http axi https://axi.sripto.tech/api/mcp

curl -s https://axi.sripto.tech/api/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"axi_calculate_score","arguments":{
        "cf":{"action":1,"memoryWrite":1,"identity":1,"perception":1},
        "ii":0.94,"ttc":41,"pc":0.91,"tdr":0.4,"re":0.86,"cli":32}}}'

The server is stateless and does not store requests.

Machine-readable specification

/axi.json: definition, stack, Bridge components, principles, guardrails, metrics, score formulas and normalization, profiles, conformance tests and citation.
/llms-full.txt: the full paper in Markdown; /llms.txt: a short index of this site.
/feed.xml: RSS for articles and paper versions.

Versioning

The specification follows the paper (currently v2.0, DOI 10.21203/rs.3.rs-9717445/v2); formulas and thresholds change only with a new paper version. Each schema names the paper section it implements.

Feedback

Corrections, questions and proposed test cases are welcome atsrivardhan@sripto.tech. The specification and site content are released underCC BY 4.0. Security issues: see security.txt.