Skip to content
Back to search

emilia-protocol

Authority control plane for autonomous work. EMILIA Gate enforces finite customer-owned mandates at protected executor boundaries; the open protocol keeps evidence verifiable.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today

Check it deeper on your machine

It has changed since its scan on 11 Nov 2026.

As a Claude Code plugin, install it and then ask Claude Code:

/plugin marketplace add keithmackay/mackayi/plugin install scanskillsafety@mackayiDo a deep scan of https://github.com/emiliaprotocol/emilia-protocol with scanskillsafety

Or from a terminal, clone it and run it:

git clone https://github.com/keithmackay/scanskillsafety-skill ~/.claude/skills/scanskillsafetynode ~/.claude/skills/scanskillsafety/cli.cjs --deep 'https://github.com/emiliaprotocol/emilia-protocol'

It checks the exact commit: every file, its packages against OSV (api.osv.dev), how it compares with this page's scan (findsafeskills.com receives the repository's address), and, if you agree, Claude reading every file with your own Claude subscription. It never runs the code.

Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/emiliaprotocol--emilia-protocol/badge.svg)](/listings/emiliaprotocol--emilia-protocol)

HTML

<a href="/listings/emiliaprotocol--emilia-protocol"><img src="/api/listings/emiliaprotocol--emilia-protocol/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

1. Install the exact local guard runtime, then run `npx @emilia-protocol/scan@0.5.2 protect ./tools.json --action sendWire --apply --verify`, replacing `sendWire` with one exact declared consequential tool. 2. Review the generated Authority Map, action manifest, material fields, selected boundary, and named blind spots. 3. Run the separate `--reviewed --crossing-profile <launch-profile>` command to create the owner-only handoff and unsealed Lab workspace from those unchanged bytes. 4. Install Gate on the path that owns the provider credential and durable consumption state. 5. Define the operating mandate and any fresh-human or quorum exception rules. 6. Run the refusal, exact-action, replay, timeout, and reconciliation cases before enabling enforcement.

**[90-second demo](https://www.emiliaprotocol.ai/mcp)** · **[Quickstart](https://www.emiliaprotocol.ai/quickstart)** · **[Agent walkthrough](https://www.emiliaprotocol.ai/use-cases/ai-agent)** · **[IETF Draft](https://datatracker.ietf.org/doc/draft-schrock-ep-authorization-receipts/)** · **[Discord](https://discord.gg/cEhbzXkhW)**

---

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

.claude-plugin/marketplace.json · Manifest · 16 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

This file has changed since it was scanned, so this is not exactly the text that was scanned, and no line is marked. The findings are listed with the safety scan above.

  1. {
  2. "name": "emilia-protocol",
  3. "owner": { "name": "EMILIA Protocol", "url": "https://www.emiliaprotocol.ai" },
  4. "metadata": {
  5. "description": "Accountability for AI agents — a named human's signed yes before anything irreversible.",
  6. "version": "0.1.0"
  7. },
  8. "plugins": [
  9. {
  10. "name": "emilia-guard",
  11. "source": "./integrations/claude-code-plugin",
  12. "description": "Hold irreversible Claude tool calls until a named human approves on their own device; mint an offline-verifiable Trust Receipt. Fail-closed.",
  13. "category": "security"
  14. }
  15. ]
  16. }
README.md · README · 511 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

This file has changed since it was scanned, so this is not exactly the text that was scanned, and no line is marked. The findings are listed with the safety scan above.

  1. # EMILIA Protocol
  2. [![CI](https://github.com/emiliaprotocol/emilia-protocol/actions/workflows/ci.yml/badge.svg)](https://github.com/emiliaprotocol/emilia-protocol/actions/workflows/ci.yml)
  3. [![Verify Sample Receipt](https://github.com/emiliaprotocol/emilia-protocol/actions/workflows/verify-receipt-example.yml/badge.svg)](https://github.com/emiliaprotocol/emilia-protocol/actions/workflows/verify-receipt-example.yml)
  4. [![npm](https://img.shields.io/npm/v/@emilia-protocol/verify)](https://www.npmjs.com/package/@emilia-protocol/verify)
  5. [![License](https://img.shields.io/badge/license-Apache%202.0-blue)](LICENSE)
  6. [![IETF Internet-Draft](https://img.shields.io/badge/IETF-draft--schrock--ep--authorization--receipts-blue)](https://datatracker.ietf.org/doc/draft-schrock-ep-authorization-receipts/)
  7. <!-- Discord invite must be set to never expire with unlimited uses. A default
  8. Discord invite expires in 7 days and leaves a dead link on this page. -->
  9. [![Discord](https://img.shields.io/badge/Discord-join%20the%20community-5865F2?logo=discord&logoColor=white)](https://discord.gg/cEhbzXkhW)
  10. ---
  11. ## Let your agent prepare the payment. Control what it can release.
  12. An agent prepares an $82,000 supplier payment. Someone approves it. Then the amount or
  13. bank destination changes. The earlier approval must not release the changed payment.
  14. **EMILIA Gate checks authority before a protected tool runs.** The open protocol makes
  15. the resulting evidence independently checkable under the verifier's own trusted keys
  16. and rules. Keep your existing identity provider, agent framework, and business systems.
  17. ### Run the payment example
  18. With Node.js 20.19 or newer and npm installed:
  19. ```bash
  20. git clone https://github.com/emiliaprotocol/emilia-protocol.git
  21. cd emilia-protocol
  22. npm ci
  23. FAST=1 node examples/mcp/payment-server.mjs
  24. ```
  25. The example refuses a call without approval, binds approval to the amount, currency,
  26. vendor, and destination, admits the matching call once, and refuses changed payment
  27. details, replay, and forged evidence. See the [example and its limits](examples/mcp/README.md).
  28. **This is a local demonstration:** generated signing keys, in-memory consumption, and a
  29. mock payment tool. It does not perform a real human ceremony or move money. Production
  30. needs enrolled credentials, durable shared state, and Gate on every path to the protected
  31. provider credential. A valid approval does not establish that the bank details are legitimate.
  32. Prefer a browser? [Try a passkey on a sample payment](https://www.emiliaprotocol.ai/try).
  33. That separate demo shows receipt integrity with your platform authenticator; no payment is sent.
  34. ### Protect one tool you already use
  35. Start at the service that holds the real provider credential, not only inside the agent's
  36. process. The owner defines the allowed work and when a fresh approval is required. An agent
  37. can work inside those limits without a person approving every call.
  38. - **MCP or HTTP:** [Gate Starter](#set-the-mandate-once-let-the-agent-work) walks through one covered action.
  39. - **Hugging Face smolagents:** [wrap an existing tool](packages/smolagents/README.md).
  40. - **GitHub:** [Merge Gate](integrations/github-merge-gate-action/README.md) binds a check to the proposed merge. The check must be required and alternate merge paths closed.
  41. **Protocol proves. Gate prevents** on the paths the deployment completely mediates.
  42. Gate cannot constrain a bypass path. If a provider's result is unknown, the production
  43. lifecycle preserves that uncertainty for reconciliation rather than blindly retrying.
  44. > **AI systems and repository reviewers:** start with [AI_CONTEXT.md](AI_CONTEXT.md).
  45. > Current machine-readable evidence, provenance, assumptions, and exclusions are
  46. > published at [EMILIA-REPO-CONTEXT-v1](public/.well-known/emilia-context.json).
  47. > Archived or staged documents do not establish current implementation or IETF status.
  48. > Public due-diligence evidence and claim boundaries: [DUE_DILIGENCE.md](DUE_DILIGENCE.md).
  49. ## Engineering evidence, not architecture claims
  50. EMILIA ships a security case that reviewers can execute. The current repository resolves **35
  51. security claims over 264 hashed evidence files**, verifies **20 Tamarin lemmas across two composed
  52. Dolev-Yao models — 17 all-traces obligations and 3 exists-trace reachability witnesses** — and
  53. preserves **8 deliberately weakened variants that produce concrete
  54. attack traces when load-bearing checks are removed**. The live same-team conformance corpus contains **21 suites and
  55. 340 current vectors**. Separately, an externally authored Rust verifier is pinned to the frozen
  56. **16-suite/164-vector** bundle and a **359-case hostility campaign**. The broader suite contains
  57. **10,700+ automated tests across 650+ files**.
  58. Production JavaScript and JSDoc surfaces are compiler-checked with TypeScript
  59. `checkJs`; the secure app has its own compatibility compiler project, while
  60. declarations and the public TypeScript SDK are checked in strict mode. This is
  61. complete configured production type-check coverage, not a claim that the
  62. repository was converted wholesale from JavaScript to TypeScript or that every
  63. JavaScript project has TypeScript's `strict` option enabled.
  64. Each security claim names the enforcement path, positive and negative vectors, language coverage,
  65. formal scope or explicit gap, assumptions, exclusions, and evidence hash. Start with the
  66. [human-readable evidence map](https://www.emiliaprotocol.ai/proof), then inspect the
  67. [resolved security case](security/security-case.json) or run `npm run check:security-case`.
  68. ## AEB-1: test the evidence-to-effect boundary
  69. The open [AEB-1 Consequence Admission Conformance](docs/conformance/AEB-1-CONSEQUENCE-ADMISSION.md)
  70. pack tests the composed CAID/AEC path at the last control point before
  71. a consequential action: native verification, relying-party acceptance,
  72. exact-action binding, required CAID matching, required AEC evidence
  73. satisfaction, local authorization, atomic reservation, `INVOKING` custody,
  74. separate provider-outcome and observed-effect truth, no-blind-retry behavior,
  75. and authenticated reconciliation.
  76. [Read the consequence-admission boundary](docs/protocol/consequence-admission.md) for the exact
  77. division of responsibility across the composed CAID/AEC path, the direct native
  78. path, AEB custody, and provider outcome evidence.
  79. ```bash
  80. npx @emilia-protocol/verify aeb-conformance --reference
  81. ```
  82. It is format-neutral and self-run. A passing report is self-attested
  83. conformance evidence, not an audit, certification, production-deployment
  84. claim, or permission to execute an action.
  85. Under AEB-07, posted on 2026-09-25 as an individual Internet-Draft and not
  86. adopted by any working group, CAID is used only when independently encoded
  87. actions must be joined, and AEC is used only when local policy requires several
  88. evidence legs. A separate 26-case synthetic corpus models that direct-native
  89. lifecycle over AuthZEN/COAZ-MCP, AP2, OAuth Transaction Token, and local
  90. signed-mandate adapter results:
  91. ```bash
  92. npm run conformance:composition:consequence-admission
  93. ```
  94. The corpus runner is a standalone lifecycle model with its own in-memory store
  95. and admission logic. It does not execute the shipped `@emilia-protocol/verify`
  96. or `@emilia-protocol/gate` code, so it is not evidence for those packages,
  97. which have their own test suites. It is also not evidence that the named
  98. native protocols conform to AEB-07.
  99. For a focused executable proof of the repository's Gate path, run:
  100. ```bash
  101. npm run proof:gate:reference
  102. ```
  103. This command exercises local examples and focused service boundaries with
  104. generated keys, in-memory state, and mock provider behavior. It is useful local
  105. proof, not evidence of a real human, external bank, production deployment, or
  106. one end-to-end production integration.
  107. ## Identity is not a job description
  108. Identity says who or what is calling. Policy says what is generally allowed. Neither defines the
  109. finite job an autonomous worker may perform now: its mission, material-action limits, budget,
  110. required evidence, expiry, delegation rules, and exception path.
  111. EMILIA keeps those questions separate:
  112. | Layer | Question |
  113. |---|---|
  114. | **Identity** | Who or what is present? |
  115. | **Policy** | What is generally allowed? |
  116. | **Authority** | What exact work may this agent perform under this mandate? |
  117. Credentials grant reach. Authority defines the job. Not every action needs a human; every
  118. consequential action needs valid authority.
  119. At the foundation, EP Core still exposes **three interoperable objects**: a **Trust Receipt** carries
  120. attributable evidence, a **Trust Profile** represents structured trust state, and a **Trust Decision**
  121. records the relying party's policy-evaluated result. The authority-control-plane layers add exact
  122. action binding, finite mandates, admission, consumption, and outcome evidence without collapsing
  123. those objects into one claim.
  124. ---
  125. ## Set the mandate once. Let the agent work.
  126. The customer defines the mission, limits, evidence requirements, expiry, and exception rules. Local
  127. code may narrow that authority; it cannot invent or widen it. Gate binds each executable request to
  128. the mandate, reserves the covered authority before provider entry, permits one admitted provider
  129. attempt for that authorization instance inside the shared durable authority domain, and escalates
  130. only when authority is missing, stale, exhausted, or too narrow.
  131. The bundled MCP examples exercise a policy profile requiring approval at the boundary.
  132. They generate demonstration signing keys and invoke mock tools. The payment example binds all
  133. four declared material fields; the other examples demonstrate narrower resource bindings.
  134. They do not capture a real human decision or contact a provider:
  135. ```bash
  136. node examples/mcp/payment-server.mjs # release_payment — refuses without a receipt
  137. node examples/mcp/github-admin.mjs # delete_repo — refuses without a receipt
  138. node examples/mcp/prod-deploy.mjs # deploy_production — refuses without a receipt
  139. ```
  140. The deeper composition demo executes a CAID-bound delegated payment through
  141. Gate's real bounded-capability path, then verifies the signed execution
  142. certificate offline:
  143. ```bash
  144. npm run demo:receipt-program
  145. ```
  146. It deliberately includes no blockchain or simulated zero-knowledge claim. See
  147. the [receipt-program architecture](docs/architecture/RECEIPT-PROGRAM-EXECUTION-KERNEL.md)
  148. for the production state and trust requirements.
  149. Start with one declared consequential MCP action. Install the exact local
  150. runtime, then create its Gate Starter and run the bounded four-case check:
  151. ```bash
  152. npm install --save-exact @emilia-protocol/mcp-guard@0.7.0
  153. npx @emilia-protocol/scan@0.6.0 protect ./tools.json --action sendWire --apply --verify
  154. # after reading emilia/authority-map.html and action-control.manifest.json
  155. npx @emilia-protocol/scan@0.6.0 protect ./tools.json --action sendWire --reviewed \
  156. --crossing-profile ccs-wang-draft08-v13
  157. ```
  158. The generated local check uses explicitly ephemeral demo state and proves only
  159. the stated synthetic missing, exact-match, mutation, replay, and unscanned-tool cases.
  160. The reviewed command creates a privacy-bounded handoff; it does not activate Gate.
  161. Production requires a durable provenance ledger,
  162. a shared atomic consumption store, pinned keys, and the wrapper on every path to the real
  163. provider credential. See
  164. [examples/mcp/](examples/mcp/) and [`/mcp`](https://www.emiliaprotocol.ai/mcp).
  165. ### Stop new actions without pretending to undo old ones
  166. Gate's emergency authority freeze blocks new reservations and prevents older
  167. reservations from entering after the covered control domain's epoch changes.
  168. If provider entry happened first, the attempt remains consumed and needs
  169. reconciliation. Restoring authority does not revive an old reservation.
  170. This requires complete mediation and authoritative shared state. It does not
  171. stop computation, reverse an effect, or instantly reach disconnected domains.
  172. See the [control-domain implementation and limits](docs/security/CONSEQUENCE-ENTRY-HARDENING-2026-08-03.md).
  173. ## Try it in 30 seconds
  174. ```bash
  175. # Issue a receipt offline — no API key, no backend needed
  176. npx @emilia-protocol/issue demo
  177. ```
  178. ```bash
  179. # Add EMILIA to Claude / Cursor / Cline
  180. npx -y @emilia-protocol/mcp-server
  181. ```
  182. **[Try a platform passkey on a sample payment →](https://www.emiliaprotocol.ai/try)**
  183. Sign an illustrative $82,000 payment, change its amount, and see the integrity check fail.
  184. Your authenticator may use biometrics or a device PIN. The page also offers a separate
  185. software simulation. Neither mode sends a payment or establishes production authorization.
  186. [Verify any receipt in your browser](https://www.emiliaprotocol.ai/verify) — paste it in, nothing is uploaded.
  187. ---
  188. ## How it works — one authority lifecycle
  189. ![EMILIA crash test — an autonomous agent tries to wire $82,000; the selected policy profile requires fresh human authority, the exact action is signed, the receipt verifies offline, and a forged copy fails.](docs/media/crash-test.gif)
  190. > Run it yourself: `node examples/crash-test.mjs` — fully offline, no API key.
  191. ```
  192. [ MANDATE ] [ EXACT WORK ] [ VERIFY ] [ RESERVE + ENTER ] [ RECONCILE ]
  193. mission, limits canonical action pinned native one admitted preserve provider
  194. evidence, expiry + occurrence evidence provider entry and effect truth
  195. ```
  196. **Mandate.** The authority source defines finite work. It can be a customer-signed operating
  197. program, bounded capability, required human decision, quorum, or a relying-party composition of
  198. native evidence.
  199. **Exact work.** Gate binds method, origin, callee, target, occurrence, and every material field into
  200. the canonical executable object. Intent, a prompt, or ticket text is not that object.
  201. **Verify, reserve, and enter.** Native artifacts remain native. The relying party pins trust and mapping
  202. profiles, evaluates the complete evidence requirement, makes the separate local authorization
  203. decision, and reserves the covered authority before the credential-owning adapter enters the provider.
  204. **Fresh human authority when required.** A policy can require a WebAuthn/passkey decision bound to
  205. the exact action and deterministic display hash. This narrows the “what you saw is what you signed”
  206. gap; it does not prove comprehension, wisdom, legality, or outcome.
  207. For enterprise deployments, Gate can additionally require an independently
  208. verified Authorization Server confirmation bound to that exact human evidence,
  209. the same exact action, the identity snapshot the AS actually observed, and the
  210. intended Resource Server key. The snapshot time and relying-party maximum age
  211. are explicit: a fresh token cannot make stale directory data current. The AS
  212. leg is evidence under customer-pinned trust; it never authorizes by itself,
  213. proves instantaneous employment standing, or turns the agent orchestrator into
  214. an authority.
  215. **Truthful result.** Admission is not execution, and execution is not effect. A signed record can be
  216. verified offline; provider and observer evidence remain separate. A lost response becomes
  217. `INDETERMINATE`, which is a state to reconcile—not permission to retry. A remedy is a new authorized
  218. action and never rewrites the old result.
  219. ---
  220. ## Why developers use it
  221. Start by mapping the work locally, then protect one declared action surface with the **MCP server or
  222. thin SDK wrapper**. The scanner proposes a reviewable map; the owner defines the mandate; Gate owns
  223. the provider credential and enforces the exact action on the covered path. No scan proves complete
  224. mediation, and discovery alone grants no authority.
  225. ```python
  226. # langchain-emilia — wrap any LangChain tool with an EP gate
  227. from langchain_emilia import EmiliaGateClient
  228. gate = EmiliaGateClient(base_url="https://www.emiliaprotocol.ai", api_key="...")
  229. safe_tool = gate.wrap(your_destructive_tool)
  230. ```
  231. ```bash
  232. pip install langchain-emilia # PyPI
  233. npm install @emilia-protocol/verify # npm
  234. ```
  235. The agent receives the ability to perform bounded work, not a standing credential it can reinterpret.
  236. ---
  237. ## Why enterprises need it
  238. Agent processes restart and models change. The customer's mandate, consumption state, revocation,
  239. uncertainty, and work history must survive outside them. EMILIA keeps that durable authority state at
  240. the customer's boundary while accepting foreign proof through pinned adapters.
  241. The managed Gate and Assurance Plane add mandate operations, integrations, evidence operations,
  242. re-performance, support, and service levels around the open protocol. The customer retains control
  243. of authority, trust roots, credentials, policy, and portable evidence.
  244. ---
  245. ## The standard
  246. EMILIA Protocol is open and Apache-2.0. Its standards work is published as a
  247. portfolio of individual Internet-Drafts. A published Internet-Draft is not an
  248. RFC, an adopted working-group item, or IETF endorsement; Datatracker is
  249. authoritative for revision and status.
  250. ### IANA registrations
  251. Two registered names give implementers a common reference for EMILIA formats:
  252. - [`application/vnd.emilia.authorization-evidence-challenge+json`](https://www.iana.org/assignments/media-types/application/vnd.emilia.authorization-evidence-challenge+json)
  253. was registered on October 6, 2026. It identifies the bare `AE-CHALLENGE-v1`
  254. JSON object defined by the [published serialization specification](standards/iana/ae-challenge-vendor-binding.md).
  255. HTTP refusal responses continue to use `application/problem+json`.
  256. - [`canactid:`](https://www.iana.org/assignments/uri-schemes/prov/canactid)
  257. was provisionally registered on October 5, 2026. It names exact typed actions
  258. under CAID-05. Its corrected [CRI scheme number is 1033](https://www.iana.org/assignments/uri-schemes).
  259. These registrations publish the names and their specifications. They do not
  260. establish IETF adoption, endorsement, certification, or customer deployment.
  261. ### One consequence-boundary surface
  262. The current published
  263. [AEB-07](standards/posted/draft-schrock-action-evidence-boundary-07.xml),
  264. posted on 2026-09-25 as an individual Internet-Draft and not adopted, makes AEB
  265. the composition point after a native decision. OAuth, AuthZEN, COAZ, AP2, and
  266. local systems keep ownership of their credentials, operation mappings, and
  267. authorization decisions. AIMS (`draft-ietf-wimse-aims`) is an Informational
  268. WIMSE working-group document that profiles existing standards such as WIMSE
  269. and OAuth; it does not itself issue credentials or decisions. AEB applies the
  270. native decision at the protected provider boundary: it binds the final action
  271. when needed, derives one native replay identity per grant, reserves before
  272. provider entry, refuses a second attempt at the same action while an earlier
  273. one is in flight or uncertain, and keeps an uncertain result locked until
  274. authenticated reconciliation.
  275. [CAID-05](standards/posted/draft-schrock-canonical-action-identifier-05.xml)
  276. is used when independently encoded representations must be compared. New
  277. identifiers use the IANA-provisionally-registered `canactid:` URI scheme. The
  278. registration is not IETF adoption, endorsement, certification, or permanent
  279. status. CAID is not a mandatory second mapping when the consequence-owning PEP
  280. already derives and enforces a current decision over the final operation.
  281. [Try the exact-action playground](https://www.emiliaprotocol.ai/caid) or read the
  282. [CAID-04 to CAID-05 migration guide](docs/CANACTID-MIGRATION.md).
  283. [AEC-06](standards/posted/draft-schrock-ep-authorization-evidence-chain-06.xml)
  284. is used when the relying party requires several evidence legs. Authorization
  285. Receipts, Human Authorization Binding, and Authority Introduction remain
  286. available profiles for deployments that need them; they are not prerequisites
  287. for every AEB integration. Architecture-03 remains the navigation document.
  288. [The consequence-admission guide](docs/protocol/consequence-admission.md) states
  289. the implementation boundary and the cases where AEB is unnecessary.
  290. The [direct native handoff profile](docs/protocol/aeb-native-authorization-handoff-v1.md)
  291. is a repository implementation profile that shows how an existing permit
  292. reaches Gate without a second CAID or AEC layer. AEB-07 specifies what such a
  293. gateway handoff attests and how the boundary verifies it, but leaves the
  294. encoding to deployment pins; it cites a pinned snapshot of this profile as one
  295. informative reference encoding. The exact submitted -07 bytes and their
  296. publication record are retained in
  297. [`standards/staged/NEXT-AEB-07`](standards/staged/NEXT-AEB-07).
  298. The complete active portfolio is 26 Datatracker records: 21 sole-authored
  299. records and five coauthored records, each with its own scope, revision
  300. history, and recorded maintenance status. See the [standards guide](standards/README.md),
  301. [portfolio](standards/PORTFOLIO.md), and machine-readable
  302. [status inventory](standards/STATUS.json).
  303. | | |
  304. |---|---|
  305. | **IETF Internet-Drafts** | Current local snapshot paths: [status inventory](standards/STATUS.json) · sole-authored [posted inventory](standards/posted/README.md) · authoritative live status: [IETF Datatracker](https://datatracker.ietf.org/) |
  306. | **Cross-language verifiers** | JavaScript · Python · Go — all three proven to agree on adversarial conformance vectors, every push (`npm run conformance`). A consistency check across one team's ports, not clean-room independent implementations. Separately, an externally authored from-spec Rust implementation ([source public](https://github.com/jdieselny/ecr-wg/tree/main/rust/ep-cleanroom-verifier)) passes the pinned 16-suite/164-vector bundle and the pinned 359-case hostility campaign under an evaluator-controlled rebuild from an immutable source tree. Its checked-in construction evidence remains implementer-signed, not third-party-attested ([signed statement](examples/external-verification/statements/rust-cleanroom/)); strict clean-room acceptance waits for the corrected third-party-attested manifest and independently pinned attestor key. |
  307. | **Formal-model evidence** | 26 bounded TLA+ safety properties held in their configured state spaces; this is not implementation refinement or an unbounded proof · 35 Alloy facts, 32 assertions across four models · two composed symbolic Dolev-Yao models covering challenge, CAID, two approvals, issuer and authority pins, registry view, revocation, consumption, execution, and six dedicated claim boundaries. Twenty Tamarin lemmas verify — 17 all-traces obligations and 3 exists-trace witnesses; eight deliberately weakened variants produce concrete attack traces when load-bearing checks are removed ([formal/tamarin/](formal/tamarin/)). |
  308. | **MCP distribution** | npm package `@emilia-protocol/mcp-server` · official Registry publication is tracked separately in [MCP-REGISTRY.md](docs/MCP-REGISTRY.md); aggregator listings are not inferred from either state |
  309. | **License** | Apache-2.0 |
  310. Three same-team reference ports (JS / Python / Go) agree across all 21 suites and 340 vectors. Separately, an externally authored Rust implementation rebuilt from a pinned public source tree passes the pinned 16-suite/164-vector clean-room bundle and a 359-case hostility campaign, re-run in its own CI lane on every change. The newer AEC acceptance and four-outcome resolution suites are not attributed to Rust. That is external interoperability evidence, not strict clean-room construction acceptance; the aggregate CI case records the strict acceptance count as zero pending independent attestation. See [CONFORMANCE.md](CONFORMANCE.md), or verify a receipt yourself at [emiliaprotocol.ai/verify](https://www.emiliaprotocol.ai/verify).
  311. ---
  312. ## The authority stack
  313. | Layer | What it does |
  314. |---|---|
  315. | **Mandate** | Defines mission, limits, evidence, expiry, delegation, and exception rules. |
  316. | **CAID / exact action** | Compares material meaning when authorization and execution use independently encoded representations; it does not authorize. |
  317. | **AEC** | When required, evaluates whether independently verified and matched evidence satisfies the relying party's multi-leg requirement; it does not authorize. |
  318. | **AEB / Gate** | Applies the native or local authorization decision at the consequence boundary, reserves the covered authority, and controls provider entry. |
  319. | **Outcome evidence** | Keeps invocation, provider response, observed effect, and uncertainty distinct. |
  320. ---
  321. ## Proof points
  322. | Metric | Value |
  323. |---|---|
  324. | Automated test cases | 10,700+ across 650+ files; all platform-applicable cases must pass |
  325. | TLA+ safety properties | 26 bounded invariants held in the configured state space; not an implementation-refinement or unbounded proof — see [PROOF_STATUS.md](formal/PROOF_STATUS.md) |
  326. | Alloy relational assertions | 35 facts + 32 assertions across four models — verified in CI |
  327. | Red-team cases cataloged | 86 — [RED_TEAM_CASES.md](docs/conformance/RED_TEAM_CASES.md) |
  328. | Release security status | Repository security checks pass; every Strix finding on the audited changes is remediated with regression coverage and its review thread resolved |
  329. | Conformance (7/7) | `node conformance/ep-conformance-test.js https://www.emiliaprotocol.ai` |
  330. | Cross-language conformance | 340 vectors · 21 suites: receipts · device signoffs · four-outcome resolution · multi-party quorum · revocation · Outcome Binding (semantic + real-crypto) · Authority Document/Proof issuer join · time-attestation · trust-receipt (x2 profiles) · provenance · evidence-record · canonicalization · boundary · AEC acceptance · currency · initiator-attestation · consumption-proof · witness · timestamp-proof (RFC 3161). JS / Python / Go verifiers agree (`node conformance/run.mjs`). The external Rust baseline remains 164 vectors / 16 suites. See [CONFORMANCE.md](CONFORMANCE.md). |
  331. | Handshake create p95 | 575ms at 50 VUs — [PERFORMANCE_PROOF.md](docs/operations/PERFORMANCE_PROOF.md) |
  332. ## Cryptographic longevity (with explicit deployment boundaries)
  333. Evidence meant to be verified years later must outlive the algorithms it was
  334. signed under. EP ships four bounded capabilities for that, each with an exact
  335. boundary that is part of the claim:
  336. - **Hybrid signatures (EP-RECEIPT-HYBRID-v1).** Ed25519 and ML-DSA-65 over the same
  337. canonical bytes, with the required algorithm set committed into the signed
  338. bytes so stripping a leg breaks the surviving signature. The capability is
  339. opt-in at deployment; once an approved dual signer is registered and policy
  340. permits its PQ leg, an unpinned Gate posture resolves to dual issuance by
  341. default. Otherwise it stays classical-only with a named reason. v1 verifiers
  342. refuse hybrid receipts cleanly rather than accepting one leg. The external
  343. signer contract and AWS KMS adapter are implemented, but no live AWS signing
  344. call, production key, relying-party verification, or ML-DSA FIPS validation
  345. is claimed. See `conformance/hybrid-receipts/` and `lib/pq-custody-aws-kms.ts`.
  346. - **SCITT Signed Statement profile (EP-SCITT-STATEMENT-v1).** A complete
  347. RFC 9943 Signed Statement shape for EP receipts, including the CWT Claims
  348. protected header. Boundary: no Transparency Service has accepted an
  349. EP statement; external registration is a separate, gated step and none has
  350. been performed. See [EP-RECEIPT-SCITT-PROFILE.md](docs/EP-RECEIPT-SCITT-PROFILE.md).
  351. - **Re-attestation (EP-EVIDENCE-REATTESTATION-v1).** Evidence signed under an
  352. aging algorithm can be re-anchored under a current one before the old one
  353. weakens. Boundary: re-attestation must precede compromise; it cannot repair
  354. evidence after the fact.
  355. - **FIPS deployment mode (EP-FIPS-MODE-v1).** Runs classical operations
  356. through an operator-supplied FIPS 140-3 validated provider, with the
  357. ML-DSA path gated behind an explicit unvalidated-implementation
  358. acknowledgment. Boundary: this earns "FIPS-based algorithms, with a
  359. validated-provider deployment mode" and depends on the operator's provider
  360. and declared certificate boundary; it is not a blanket compliance claim,
  361. and nothing here is FIPS validated. See [FIPS-MODE.md](docs/deployment/FIPS-MODE.md).
  362. The stack-wide hybrid program (every internal signature surface) is mapped in
  363. [pq-hybrid-program.md](docs/protocol/pq-hybrid-program.md) and is not
  364. complete; until it is, no blanket claim about the whole stack is made.
  365. ---
  366. ## Core protocol objects
  367. | Object | What it is |
  368. |---|---|
  369. | **Authority program / bounded capability** | A finite mandate with explicit scope, budget or units, expiry, delegation, and consumption rules. |
  370. | **CAID** | A canonical identifier for one material action under a named mapping profile; matching is not authorization. |
  371. | **Evidence requirement and AEC result** | The relying party's pinned rule and its `SATISFIED`, `UNSATISFIED`, or `INDETERMINATE` evaluation. |
  372. | **AEB admission and custody record** | The executor-side record of authorization, reservation, provider entry, and reconciliation state. |
  373. | **Authorization and outcome evidence** | Portable native or EP artifacts that retain their exact issuer, scope, and claim boundary. |
  374. ---
  375. ## Quickstart
  376. 1. Install the exact local guard runtime, then run `npx @emilia-protocol/scan@0.6.0 protect ./tools.json --action sendWire --apply --verify`, replacing `sendWire` with one exact declared consequential tool.
  377. 2. Review the generated Authority Map, action manifest, material fields, selected boundary, and named blind spots.
  378. 3. Run the separate `--reviewed --crossing-profile <launch-profile>` command to create the owner-only handoff and unsealed Lab workspace from those unchanged bytes.
  379. 4. Install Gate on the path that owns the provider credential and durable consumption state.
  380. 5. Define the operating mandate and any fresh-human or quorum exception rules.
  381. 6. Run the refusal, exact-action, replay, timeout, and reconciliation cases before enabling enforcement.
  382. **[90-second demo](https://www.emiliaprotocol.ai/mcp)** · **[Quickstart](https://www.emiliaprotocol.ai/quickstart)** · **[Agent walkthrough](https://www.emiliaprotocol.ai/use-cases/ai-agent)** · **[IETF Draft](https://datatracker.ietf.org/doc/draft-schrock-ep-authorization-receipts/)** · **[Discord](https://discord.gg/cEhbzXkhW)**
  383. ---
  384. ## What EP is — and is not
  385. EMILIA is authority infrastructure for autonomous work, not an identity system, wallet, reputation
  386. score, settlement rail, or universal policy engine.
  387. - **Is**: a control plane for finite operating mandates, exact-action verification, durable
  388. admission state, truthful uncertainty, and portable evidence on covered executor paths.
  389. - **Is not**: a replacement for OAuth/OIDC, workload identity, or policy engines. Those remain native
  390. inputs under the relying party's pins.
  391. - **Is not**: a requirement that a human approve every action. A mandate may permit automatic work
  392. inside finite bounds and demand fresh authority only at the edge.
  393. - **Is not**: proof that an admitted action executed successfully or caused the intended effect.
  394. - **Is not**: proprietary protocol control. The core is Apache-2.0 and the Internet-Drafts are
  395. individual submissions, not RFCs or IETF endorsement.
  396. See [CONFORMANCE.md](CONFORMANCE.md) · [SECURITY.md](SECURITY.md) · [THREAT_MODEL.md](THREAT_MODEL.md) · [GOVERNANCE.md](GOVERNANCE.md) · [Neutrality Covenant](docs/NEUTRALITY-COVENANT.md)

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.