Super Φ.Vessel (Vessie) is an experimental governed cognitive operating layer for AI systems.
CAPABILITY ≠ AUTHORITY
Vessie explores how capable AI models, local compute, hosted inference, memory, browser research, tools, evidence, and execution can live inside one system without collapsing capability into permission.
| Surface | Purpose |
|---|---|
| Project site | Public front door, architecture, release status, and project orientation |
| Portable Windows Local Console | New portable, self-contained read-only local Ollama inventory dashboard; bundled Node runtime, no Git checkout or separate Node installation |
| React Vessie Cockpit | New modular review-only React shell, full GA108 explorer, candidate models and local evidence inspector. No model gateway connected. |
| Searchable User Manual | Live searchable operator guide for chambers, models, memory, SOMA, browser tools, governance, bridges, output, and troubleshooting |
| Launch Vessie | Current live Netlify deployment |
| GitHub Releases | Frozen standalone runtime downloads and checksums |
| Current release | v2.0-alpha.11.0.54.10 — Service Models Handler Restoration |
The separate Research Council panel in Vessie's portable localhost console accepts an operator-selected, public-synthetic NBGPocketMesh E1 JSON receipt and performs strict Web Crypto SHA-256/role checks in the browser. The review is HELD_UNTRUSTED_REVIEW_ONLY, never approval or scientific verification. The panel does not contact models, use Vessie memory, write back to ThinkTank/PocketBrain, or grant action authority. Details: Council E1 review guide.
The existing localhost console can now discover and request one live PhiServe draft using a separate --enable-phiserve-desk startup switch and an operator-controlled private bearer-token file. The browser never sees the PhiServe token. Each run requires a selected registered GGUF, prompt and separate per-request approval, then checks the draft against a redacted hash receipt before human review. The default console remains read-only, existing Ollama trials are separate, and Vessie's tool and routing authority is unchanged. Operator guide.
Vessie's existing localhost console now supports human review of the operator-run PhiServe draft pair. It reads operator-selected text/receipt files, verifies local SHA-256 parity and lets a reviewer export a strictly redacted categorical assessment. It never loads a token, runs the model, changes the canonical runtime, or grants executor authority.
An isolated, explicit operator-run PhiServe client can request a bounded text-only draft from an already running local PhiServe GGUF daemon. The user supplies a local prompt file, private token file, registered model alias, and a single-run approval flag. It returns an untrusted human-review draft and a redacted hash receipt.
This does not connect the canonical Vessie runtime, browser, R1/R2 gateways, GA108 router, PV-DLAM memory or executor; it adds no action authority and requires separate physical acceptance. See operator-only setup.
The portable Windows local console
is a separately scoped read-only localhost application. It ships as a
self-contained ZIP containing a Node 22 executable, its redistribution
license, a fixed loopback-only HTTP server, the existing bounded Ollama
probe, and an offline dashboard. Extract and double-click
Start-Local-Vessie.cmd. No Git clone, developer setup, TLS certificate,
Node installation, firewall changes, admin rights or cloud API keys needed.
- Bind address:
127.0.0.1:8791only, no LAN or public listener. - The dashboard is served from the same loopback origin as its local API.
- API requires a random in-memory session bearer and rejects cross-site origins, forwarded hosts, foreign Host values, cookies, writes and unknown routes; no CORS allowlist for remote sites.
- Real Ollama inspection queries fixed
127.0.0.1:11434read-only endpoints. - Only explicit operator-triggered inventory reads are shown. No model execution, agent action, private memory, writeback or model approval.
- The hosted React
/vessie/and classic Netlify runtime remain separate; the portable local console does not magically connect those sites. - Windows package is assembled from an exact allowlist during the GitHub Pages workflow and redistributed with the Node runtime license.
Cloud-backed Ollama entries now display as CLOUD REF rather than INSTALLED, with separate local-size, cloud and unknown counts. A cloud name suffix is a hint; an on-disk size is only a reported weight size, not a GPU compatibility claim. Nothing contacts a cloud provider. See model-location evidence rules.
See local console source, Windows portable packager, and operator notes. Windows source/packaging CI cannot replace the operator's actual machine test.
Starting from the portable Windows Local Console, you can now choose between two separate double-click entrypoints:
Start-Local-Vessie.cmd: default read-only discovery and metadata only; this mode never calls Ollama's generate endpoint.Start-Local-Trial.cmd: requires you to typeENABLEat startup. This explicitly enables individual, operator-approved text-generation experiments against the fixed local Ollama/api/generateendpoint.
For each trial, select a LOCAL_WEIGHTS_REPORTED entry, supply a prompt
of up to 2,000 characters, select 64 or 128 maximum output tokens, check
the approval box, then confirm the separate browser dialog. The server
re-probes Ollama immediately before running and refuses cloud references,
unknown origins, stale lists, extra request fields, and missing approval.
Maximum 6 attempts per process per hour, one active trial at a time,
90-second Ollama timeout, bounded model response and no automatic retry.
The local browser displays the generated text and an ephemeral redacted receipt with model name, wall time, Ollama-reported token timings and a response hash. The receipt excludes the prompt and generated text; exporting a JSON receipt is a separate human action. Neither the prompt nor reply is stored by this console, but Ollama itself may have runtime logs or implementation-specific behavior.
No BrainC/Crane Fly automatic model routing, agent execution, cloud-model
invocation, training, background tasks, governance promotion, memory access
or paid API calls are enabled. Location classification and Ollama timing
metadata are only source-reported observations, not proof of GPU placement,
security isolation, successful training or independent attestation. For
best first results use a small local model such as qwen3:4b. Close
the process and restart the default launcher to restore read-only mode.
See local Windows operator instructions and trial source.
After a future completed local one-shot trial, the operator may record a separate, categorical human answer review in the active browser session and explicitly export a small, redacted JSON review. The review references the generated answer only by its SHA-256 digest and never includes the prompt or the answer itself. Operator ratings are self-reports, not independently verified quality or a BrainC route approval. See Human Review Guide. Previously exported timing receipts cannot be given retroactive answer quality ratings unless the human also retained the actual answer.
The Local Evidence Bench can compare up to 24 explicitly selected, redacted performance and human-review JSON receipts, in browser memory only. Import is through a manual file selection, or the operator can explicitly keep the latest receipt in the bench. Exports require another click. Rows remain chronological and unranked. A matching human self-report is linked to a performance observation by model and output SHA-256, not by presumed hardware or a routing endorsement. Old performance-only receipts are supported, but no model answer quality is invented. See Evidence Bench Guide.
The bench can now be deliberately saved as a portable, redacted evidence bundle and restored in a later session by selecting that JSON file manually. The earlier descriptive-comparison export remains available, but is deliberately not reimportable. The portable bundle contains only up to 24 validated, normalized observations; it never contains prompts, model answers, bearer secrets, or routing approvals. No automatic disk save, cloud sync, retention, or import on startup. See Portable Bench Guide. New optional receipt timing fields and careful interpretation of cold load versus prompt processing and generation are documented in performance receipt guide. A trial reaching the requested token count is not automatically a complete, correct answer or a benchmark of GPU residency.
The Windows Local Console offers three public fixed-prompt cards for
governance, basic logic and code review. Clicking Load only fills the
prompt editor; it cannot execute a model. The separately enabled trial
runner still requires a locally size-reported model, a per-prompt
checkbox and a separate browser confirmation. The local server
accepts a protocol_id in its redacted receipt only after
validating the exact versioned public prompt and output-token cap.
Old custom trials remain valid and explicitly unlabeled. The evidence
bench and portable bundles preserve allowed protocol IDs for
human-reviewed, unranked comparisons. No automatic batch runs or
model routing. See
trial protocol operator guide.
The Local Evidence Bench now includes Protocol Cohorts, a pure read-only view of existing redacted observations grouped by public protocol and model. It shows sample counts, matching operator review counts, reported median timings/throughput and evidence gaps. A fixed-protocol label with a mismatched output-token limit is excluded rather than treated as a comparable trial. Custom/legacy unlabeled receipts remain in the original comparison table. The view neither runs tests nor ranks/qualifies models, and source JSON is unauthenticated. See Protocol Cohort Guide.
After a successfully completed, server-labeled public protocol trial, the local Human Answer Review panel can reveal a human-only public answer reference and three review checks by a separate operator click. Governance, sock-drawer logic and JavaScript indexing references help an operator inspect correctness before voluntarily recording the existing subjective review. There is no automatic grade, new exported rating, answer parsing, AI judge or routing permission. Custom, unlabeled or failed trials show no answer reference. See Human Protocol Reference Guide.
Version: v2.0-alpha.11.0.54.12 — BudgetGenius Canonical Outcome Handoff
Path: runtime/Super_PhiVessel_v2.0-alpha.11.0.54.12_BudgetGenius_Outcome_Handoff.html
SHA-256: a9e67b12d1c3622315b6937f1c2405acbd5b2a6bc6d32cf578c67d5178fa7212
Git blob: c9dc18eb76be9883d18014de129c07cf78c5a498
The canonical runtime was promoted from the exact standalone artifact tested in the live Netlify deployment. Runtime identity, size, version, and release invariants are recorded in runtime/MANIFEST.json.
Super Φ.Vessel treats an AI model as a capability provider, not as the final authority boundary.
The system separates:
- human intent from operational permission
- model generation from verified evidence
- observation from inference
- proposal from execution
- memory retrieval from prompt admission
- browser reading from browser interaction
- model discovery from model approval
- model approval from model activation
- evidence transport from authority transport
The result is less “one giant AI agent that can do everything” and more a governed cognitive workstation where different capabilities have explicit boundaries.
- CAPABILITY ≠ AUTHORITY
- GOAL ≠ PERMISSION
- INFERENCE ≠ OBSERVATION
- PROPOSAL ≠ EXECUTION
- RETRIEVAL ≠ PROMPT ADMISSION
- EVIDENCE TRANSPORT DOES NOT TRANSPORT AUTHORITY
- CAPABILITY TO REMEMBER ≠ PERMISSION TO SHIP MEMORY
When authorization, provenance, evidence integrity, or execution state cannot be established, governed paths are designed to prefer explicit BLOCKED / HELD / FAILED / INSUFFICIENT outcomes over invented success.
Read the full model in docs/GOVERNANCE.md.
Vessie can be used as a normal conversational interface, but its main governed work path is organized around five chambers:
| Chamber | Role |
|---|---|
| Quick Chat | Everyday conversation, questions, planning, and bounded research |
| Dreamer | Symbolic, speculative, unconventional, and exploratory thinking |
| Translator | Converts selected Dreamer material into explicit mappings and reality-bridge candidates |
| Builder | Produces governed candidate builds, patches, tests, and verification evidence |
| Ledger | Audits provenance, evidence, conflicts, execution facts, and final governed receipts |
These chambers are not a mandatory ceremony. Use only the path the task requires.
| Surface | Current status | Boundary |
|---|---|---|
| Quick Chat / Dreamer / Translator / Builder / Ledger | LIVE | Role separation does not itself grant execution authority |
| PV-MEM Vessel Memory | LIVE | Recall is context only; it does not promote truth or authority |
| Multi-Vessel Tabs / WorkObject | LIVE | Context and project state do not carry authority by default |
| Professor Φ / Genius Atlas | LIVE | Specialist perspectives remain non-authoritative |
| Brain Registry / AutoFabric / Crane Fly | LIVE | Discovery and routing remain bounded by approved model pools |
| Model Scout / Challenger Bench | LIVE | Comparison evidence never silently installs, activates, or promotes a model |
| SOMA sensory surfaces | LIVE / CONDITIONAL | Device/browser permission and operator action remain required where applicable |
| Direct Research Fabric | LIVE | Read-only public web observation; no network mutation |
| Browsallax Browser Operator | CONDITIONAL | Interactive authority remains separately governed |
| Netlify BrainC | LIVE in deployment | Hosted cognition does not inherit tool or action authority |
| PhiOS Bridge | CONDITIONAL | Proposal context does not equal execution; execution requires an existing lease |
| Local Ollama Fabric | CONDITIONAL | Installed/discovered does not mean approved or active |
| Artifact Workshop | LIVE | Output/presentation does not promote evidence or authorize action |
| SPD-W v0.1 | EXPERIMENTAL / UNWIRED | Frozen acceptance contract passes independently; not integrated into the canonical runtime |
For current subsystem truth, use docs/RUNTIME_STATUS.md and the live searchable manual.
The GitHub Pages manual is a live static documentation app:
Open the Super Φ.Vessel User Manual
It includes:
- instant client-side search
- WORK / INTELLIGENCE / SENSES / GOVERNANCE / LABS / BRIDGES / OUTPUT filters
- LIVE / CONDITIONAL / EXPERIMENTAL / LOCKED / LEGACY / UNWIRED status labels
- purpose and location for major operator-visible systems
- explicit authority boundaries
- five-chamber quick start
- Service / Models guidance
- memory guidance
- web reading vs interaction guidance
- troubleshooting notes
- shareable query/filter URLs
The manual is read-only documentation. It cannot route a model, approve an executor, install software, promote evidence, mint a grant/lease, or execute an action.
Open superphivessel.netlify.app.
The hosted deployment provides the current web experience and Netlify BrainC integration.
Download the current runtime from the v2.0-alpha.11.0.54.10 GitHub Release and open the HTML file in a modern browser.
The standalone runtime can provide a large portion of Vessie's interface and local browser behavior without an install. Hosted/server-side integrations such as Netlify BrainC require their deployment environment, and local-model execution requires a configured local model runtime such as Ollama.
A fresh public Vessie must not ship with prior operator-derived:
- chat or Dreamer history
- committed Vessel memory
- prior session/work state
- receipts
- persisted provider credentials
- private browser-derived state
The public data epoch introduced in .54.8 is retained by .54.10:
PV-PUBLIC-BLANK-2026-09-29-V1
The .54.10 hotfix does not establish a new epoch, so it does not intentionally repeat the legacy-state reset for users already on the current epoch.
The source-level blank-state contract and CI gate remain active. Full browser-storage deployment verification is tracked separately in docs/BLANK_STATE.md.
.54.10 repairs the complete known-good Service → Models handler set that remained referenced by the UI after later runtime merges dropped the implementations:
- exportKeyring
- autoAssignOllama
- applyOllamaStabilityPreset
The exact artifact was verified in production before canonical promotion.
CI now contains a regression gate ensuring those handler definitions remain present before their Service / Models bindings.
core-integrity.yml verifies:
- canonical runtime path from runtime/MANIFEST.json
- runtime byte size
- SHA-256
- version marker
- blank-state migration contract
- Service / Models handler restoration contract
- frozen SPD-W acceptance contract
The repository also contains an automated canonical-release workflow. A release is published only after the runtime identified by the manifest matches its expected digest.
Release chain:
production-tested runtime
↓
runtime/MANIFEST.json
↓
deterministic CI
↓
main
↓
GitHub tag + prerelease
↓
standalone HTML + checksum
↓
GitHub Pages status/manual
Human Operator
│
▼
Intent / Work Context
│
▼
Governance + Authorization
│
▼
Routing / Cognitive Roles
│
▼
Model / Tool Capability
│
▼
Evidence + Verification
│
▼
Effect Boundary
│
▼
Receipts / Ledger / Persistent State
Read docs/ARCHITECTURE.md for the architectural overview.
SuperPhiVessel/
├── runtime/ Canonical standalone runtime + integrity manifest
├── site/ GitHub Pages project site + searchable user manual
├── docs/ Architecture, governance, runtime status, blank-state contract, release notes
├── protocols/ Governed contracts and protocol specifications
├── packages/ Extracted reusable components and experiments
├── tests/ Deterministic tests, fixtures, and acceptance harnesses
├── deploy/ Deployment adapters/configuration
├── examples/ Bounded examples and reference integrations
├── research/ Design notes and research artifacts
├── legacy/ Selected historical artifacts, not current runtime truth
└── .github/ CI, release publishing, Pages deployment, contribution surfaces
- Project site
- Searchable User Manual
- Architecture
- Governance
- Runtime Status
- Blank-State Release Contract
- Roadmap
- Security
- Contributing
- Release notes
Super Φ.Vessel is an alpha research and engineering project.
Interfaces, protocols, file layout, model support, bridges, and runtime contracts may change rapidly. A named feature, source file, package, or historical note is not evidence that a subsystem is currently active, authorized, integrated, or production-ready.
Do not commit:
- API keys
- access tokens
- provider credentials
- private exports
- browser storage
- local memory databases
- operator-specific runtime state
See SECURITY.md.
Source code is released under the MIT License, except where a file or asset is explicitly marked otherwise.
Created and maintained through an ongoing human–AI engineering collaboration led by MichaelWave369.
The existing GitHub Pages front door and searchable manual remain intact.
The new React shell is built from apps/vessie-web/ into site/vessie/ only during Pages deployment.
Preview-only: shows real GA108 roster and repository manifest, explicit candidate model statuses, and browser-only bounded receipt inspection. Does not run models, accept secrets, validate signatures, or replace the canonical executor. Physical Observer .54.13 remains a separate unpromoted review candidate.
The React Model Fabric tab can optionally pair with a locally trusted HTTPS gateway started by the operator. Pairing is a one-use secret from the local terminal and a short-lived in-memory read-only session, never a model execution grant. See R2 local certificate setup and field limitations.
The canonical .54.12 runtime and unpromoted .54.13 Physical Observer are untouched. R2 is source/test qualified only; real browser/private-network behavior requires an explicit Windows pilot.
The local Python R3-A routing trace exporter uses read-only SQLite to verify P1-C route receipts and optional P3-A shadow outcomes before projecting only safe, bounded routing metadata. The React receipt inspector recognizes that projection without claiming a signature or authority.
No live BrainC telemetry, model execution or browser gateway trace endpoint is enabled. R2 field qualification remains an operator-owned prerequisite.
R3-B source audit
pins the actual BrainC v1 model/chat code and provides a fixed-loopback local
probe for /models and /models/active. The CLI defaults to a redacted
review report. This is not an executed chat-model or Crane Fly route receipt,
and it does not connect BrainC to the React HTTPS gateway.
If this is your first time using SuperPhiVessel on a Windows PC,
nothing is installed by merging GitHub pull requests. Use the
Windows starter page
and download the portable SuperPhiVessel-Windows-Starter.zip
built and hosted automatically by the GitHub Pages workflow.
Extract the ZIP and double-click Start-Vessie.cmd (no administrator
rights required). The read-only launcher checks optional local Ollama
at 127.0.0.1:11434 and opens the hosted React cockpit. No Node.js,
Git clone, gateway credentials, TLS certificate or private key are
required for this first step.
This is a first-run starter, not a full local SuperPhiVessel installation. It does not install a model, execute inference, import private memory, or make the hosted cockpit read your local models. The secure browser pairing gateway remains a separate optional R2 field pilot requiring Windows trust/browser qualification.
The package is built from three allowlisted files in
windows-starter/ with
build-package.py. No download
scripts, package-manager installs, self-updaters, server exposure,
certificate bypasses or privileged actions are included.
The R2 field-pilot guide
includes Invoke-R2FieldPilot.ps1 with explicit Collect and
Assess stages. It combines the existing no-key-read preflight,
fresh operator trial ID, Windows strict TLS/refusal checks and later
offline redacted-browser comparison. An actual Windows browser run is
still mandatory; no CI test or local helper self-certifies a machine,
changes route weights, or grants model execution.
Before physically testing browser-to-localhost pairing, the Windows R2 operator preflight can check Node version, the public TLS certificate's IP SAN/validity, and the existence of the private key outside the repository, without opening that key or changing Windows trust. It deliberately reports no real machine, gateway, browser or routing qualification. The physical Windows R2 pilot still requires separate operator testing.
The R2 physical field pilot now generates one random, non-secret trial ID for the Windows HTTPS report and React browser report. The assessor requires the same label and a bounded timestamp relationship, preventing accidental comparison of unrelated runs. Browser receipts use schema v0.3. Matching reports still do not attest a machine, prove inference quality, qualify R2 physically, approve BrainC, or grant execution authority.
The Windows R2 physical-pilot guide
now requires the browser to observe both an acknowledged read-only
session deletion and subsequent 403 SESSION_DENIED for that same
old bearer. The redacted browser report is now v0.2; older reports
cannot pass the updated field comparator. A browser-observed refusal
does not equal independent Windows/Chromium field qualification,
live routing permission or runtime attestation.
The R3-F signed batch auditor
reviews multiple operator-signed R3-E receipts together and flags
repeated packet IDs, indistinguishable redacted summaries, non-monotonic
operator signing times and mixed terminal outcomes. Its results remain
operator-custody only, not a runtime trace, independent quality
metric, cross-batch replay prevention or P4 learning reward. Production
runtime .54.12 remains frozen; R2 physical field qualification and
Physical Observer .54.13 hold remain unchanged.
An optional R3-E Ed25519 operator envelope
wraps the redacted R3-C/R3-D projection for independently checkable
operator-key custody, using an externally supplied trusted public key.
It does not attest canonical runtime source, completed inference, answer
quality, physical R2 pairing, or grant action/learning authority. The
canonical .54.12 production HTML and .54.13 promotion hold remain
unchanged. React recognizes imports but cannot validate signatures.
R3-D hardens the R3-C projector and React inspector against misleading
mixed terminal outcomes. When the same imported run contains both native
FAILED and COMPLETED rows, the report now declares
MIXED_TERMINAL_RECORDS_UNRESOLVED instead of selecting completion by
precedence. This does not prove a final outcome, a valid retry sequence,
or suitability for routing-learning rewards. The canonical runtime and
authorization chain remain unchanged. See the
R3-C/R3-D evidence contract.
R3-C receipt projector
joins native .54.12 PhiRunCapsule, BrainRouteReceipt,
ExecutorAuthorizationReceipt, attempt-ledger and optional BudgetGenius
canary influence records. Selection, authorization, attempt and recorded
completion are separate claims; completion is not independently verified.
The React inspector recognizes the resulting redacted schema without treating
it as live or as an executor grant. No production HTML bytes are changed.
The React Vessie cockpit includes a SCOUT HANDOFF tab to review operator-pasted phibot.scout-vessie-handoff.v0.1 JSON from a real Windows PhiBot Scout qualification. The UI recomputes source freshness but does not verify the SHA digest (it has no original receipt bytes), authenticate model execution, or grant any capability. No network, model, routing or memory connection is implemented. See operator guide.
The SCOUT HANDOFF React tab can now read a previously qualified, redacted PhiBot Scout receipt from the existing optional local HTTPS pairing gateway. The operator must explicitly configure one local scout-* receipt directory at gateway startup. Default disabled; no model, tool, memory, routing or agent invocation. Pair under MODEL FABRIC, then perform one explicit read in SCOUT HANDOFF. Browser session remains in memory across tabs for at most 15 minutes. See Windows operator guide.
SuperPhiVessel now includes a small offline read-only Node.js 20+ consumer for PhiLearn Skill Passport v1:
node interop/verify-passport.mjs path/to/skill-passport.json
It is not connected to Vessie routing, local-console APIs, memory, agent identities, deployment gates or tools. No network calls occur. The command reads a locally supplied JSON file capped at 256,000 bytes, replays six simulated PhiLearn R6 actions, verifies R7 and R8 SHA-256 checksums, and outputs a JSON report.
Pass statuses are explicitly REPLAY_PASS_UNTRUSTED; failing data returns REPLAY_FAIL_UNTRUSTED. In both cases authority = NONE and permissions = []. Checksums are not digital signatures and cannot prove model identity or that a claimed observation happened. A JSON author can fabricate and rehash synthetic traces. Imported passports never grant capabilities or authority.
Conformance checks:
node --test interop/skill-passport-verifier.test.mjs
This adapter establishes wire-protocol compatibility, not live interchange or any trust promotion. PhiLearn's interop/README.md documents the contract.
R10 adds a read-only offline signature-verification layer to the existing R8/R9 portable Skill Passport. The outer format is philearn.signed-skill-passport.v1, enclosing an unchanged R8 philearn.skill-passport.v1 record. Signature checking is optional and requires an independently operator-pinned public key; receiving a public key from the sender is not sufficient authentication.
Run on a trusted operator-owned machine:
node interop/r10-cli.mjs keygen /secure/operator-dir field-lab
node interop/r10-cli.mjs sign passport.json /secure/operator-dir/issuer.private.pem field-lab signed.json
Keygen creates a private Ed25519 PEM and a candidate public trust file. The unencrypted private key is saved with permission mode 0600, but must be protected using local access controls, excluded from Git, and kept out of shared or cloud-synced directories. For higher-risk signing prefer hardware-backed key protection. The recipient separately authenticates the issuer identity and public key fingerprint through an out-of-band channel, then pins that key in their own local trust-store copy. A key delivered along with an untrusted passport must not be automatically accepted.
Verify using an already authenticated receiver-controlled local trust store:
node interop/r10-cli.mjs verify signed.json /local/operator-pinned-trust.json
Optional --review is an explicit inspection acknowledgment, not a permission change. Unknown keys, revoked/inactive keys, wrong scope, modified payloads, forged signatures or missing evidence fail closed. Without a trust store the signature remains UNPINNED_SIGNATURE. Key revocation is manual at the receiver and never auto-synchronized.
The Ed25519 signature covers a versioned domain separator, issuer ID, SHA-256 DER-SPKI public key fingerprint, claimed signing time and validated R8 passport payload digest. The verifier always validates the six frozen synthetic R6 tool traces and the R7/R8 hashes before accepting provenance.
- PhiLearn: offline Node 20+ signer/verifier and conformance checks.
- BrainC: offline Python 3.11+ verifier using cryptography; no FastAPI integration.
- SuperPhiVessel: offline Node 20+ verifier; no routing or local-console integration.
In all outcomes, even after a valid pinned-key signature and manual inspection: authority NONE, permissions [], and actionAllowed false. A valid signature confirms use of the matching operator-pinned private key, not the historic truth of an editable synthetic benchmark, real-world model ability, changes to weights, an externally attested clock, or permission to execute actions. No signing keys are committed to the repository.
node --test interop/r10-signed-passport.node-check.mjs
R11 keeps signing-key identity, demonstrated synthetic benchmark evidence, and operational permissions as three different concepts. The last remains completely outside this module. The receiver cannot enroll anything for action execution, even when every test succeeds.
A standalone receiver-controlled inspection enrollment registry records the operator's explicit approval to inspect one exact R10 Ed25519 signed R8 Skill Passport. The review entry is bound to the entire signed envelope's SHA-256, issuer ID, signing key fingerprint, model label, the exact R6 synthetic-only scope, the reviewing operator, and a short lease (1–168 hours). The system replays the six synthetic R6 traces and validates R10's operator-pinned public key every time. The receiving machine's local registry is not a cryptographic credential. Keep it outside shared/untrusted locations.
Node.js 20+ (PhiLearn or SuperPhiVessel):
node interop/r11-cli.mjs init local-registry.json
node interop/r11-cli.mjs check signed.json /local/pinned-keys.json local-registry.json
node interop/r11-cli.mjs approve signed.json /local/pinned-keys.json local-registry.json operator-id 24 reviewed-registry.json
node interop/r11-cli.mjs check signed.json /local/pinned-keys.json reviewed-registry.json
To revoke, copy the enrollmentId from your locally approved registry and run:
node interop/r11-cli.mjs revoke reviewed-registry.json <enrollment-sha256> operator-id revoked-registry.json
BrainC Python 3.11+ with cryptography installed:
python3 scripts/r11_enrollment.py init local-registry.json
python3 scripts/r11_enrollment.py approve signed.json /local/pinned-keys.json local-registry.json operator-id 24 reviewed-registry.json
python3 scripts/r11_enrollment.py check signed.json /local/pinned-keys.json reviewed-registry.json
python3 scripts/r11_enrollment.py revoke reviewed-registry.json <enrollment-sha256> operator-id revoked-registry.json
All approve and revoke operations require explicit CLI invocation. Updates are written to new files rather than silently overwriting existing registries; operators decide whether and when to replace the active receiver-controlled file.
- SIGNATURE_REJECTED: R10 signer pin or R8/R9 source replay failed.
- REGISTRY_REJECTED / NOT_ENROLLED: absent, malformed, mismatched, or unapproved receiver-local registry.
- NOT_YET_VALID / EXPIRED: local enrollment clock outside the bounded lease.
- REVOKED: operator revoked this exact envelope; same envelope cannot be silently renewed.
- APPROVED_FOR_INSPECTION_ONLY: exact payload, pinned key, valid evidence replay, and receiver's manually approved unexpired inspection lease.
Invariant: Every result carries authority NONE, permissions [], actionAllowed false. Even approval authorizes only local passive evidence inspection. This is not executable agent enrollment, an access token, deployment approval, a persisted learned skill, or a privilege escalation. No API endpoints, tool bridges, model calls, agent-memory writes, shell execution from evidence, key generation, network exchange, or runtime permission modifications were added by R11. No signing key is checked in.
The registry can be edited by whoever controls the host, and local timestamps depend on that machine's clock. It therefore does not establish an independently verified human approver or trusted time. A more advanced release would require authenticated operator decisions, tamper-evident log storage, key lifecycle operations and a separate real-world action policy gate before permitting any tool execution.
R12 adds a separate, offline operator event history on top of R10 pinned-key provenance and R11 inspection-only enrollment. It is intentionally not wired into live BrainC or Vessie APIs, agents, model routing, memory, authorization or tooling.
The event log's scope is limited to inspection of evidence. It records an operator-signed event for APPROVE_INSPECTION, REVOKE_INSPECTION, or EXPIRE_INSPECTION, with an incrementing sequence number, SHA-256 of the prior signed event, exact signed passport envelope hash, pinned issuer key fingerprint, model label, claim scope, reviewing operator's pinned Ed25519 key fingerprint, timestamp, and bounded expiry (at most seven days).
A receiver independently validates every Ed25519 event signature using a separate, operator-controlled R12 public-key trust store, reconstructs the hash-linked chain, validates state transitions, compares the entire event chain with an independently preserved checkpoint (event count + head SHA-256), and then rechecks the underlying R10 signed passport and R6 synthetic replay. A changed, reordered, forged, truncated or stale log is rejected when the receiver has the authentic latest checkpoint. A signed revocation or explicit expiry is terminal for that exact enrollment.
Critical limitations: The chain alone does not prevent deletion or rewriting. If an attacker can also replace the receiver's checkpoint or pinned keys, a local verification cannot detect all tampering. Keep the latest known head independently from the writable ledger, ideally replicated to independent protected storage. Ed25519 authenticates possession of a pinned operator signing key, not an independently verified human identity or honest historical action. Local timestamps are untrusted. All CLI operations create new files, not silent in-place overwrites. Private operator keys must be generated and stored outside Git or synced folders.
Invariant: Even a valid checkpointed and operator-approved report returns authority NONE, permissions [], actionAllowed false. It authorizes passive evidence inspection only. No runtime permissions, agent tool grants, live deployment approval, memory-writing, or model execution exist in R12.
The full offline signing and event-authoring CLI is included with PhiLearn:
node interop/r12-cli.mjs operator-keygen /secure/outside-repo/dir operator.alpha
node interop/r12-cli.mjs init new-ledger.json new-checkpoint.json
node interop/r12-cli.mjs approve signed.json R10-pinned-keys.json R12-operator-keys.json ledger.json operator-private.pem operator.alpha 24 prior-checkpoint.json next-ledger.json next-checkpoint.json --confirm
node interop/r12-cli.mjs revoke enrollment-sha256 R12-operator-keys.json ledger.json operator-private.pem operator.alpha prior-checkpoint.json next-ledger.json next-checkpoint.json --confirm
node interop/r12-cli.mjs expire enrollment-sha256 R12-operator-keys.json ledger.json operator-private.pem operator.alpha prior-checkpoint.json next-ledger.json next-checkpoint.json --confirm
node interop/r12-cli.mjs check signed.json R10-pinned-keys.json R12-operator-keys.json ledger.json independent-checkpoint.json
The operator-keygen command creates an unencrypted mode-0600 private key PEM and a candidate public trust-store JSON. Recipients must verify and pin the public key through a separate authenticated process before accepting it. Do not commit either real private keys or approvals as authorization credentials. Explicit event expiry is optional because elapsed lease expiry is enforced during verification even without a new event.
BrainC has its own Python 3.11+ standard-library/cryptography implementation in scripts/r12_governance.py. SuperPhiVessel includes a passive read-only Node verifier in interop/r12-check.mjs, with the matching shared Node event-chain implementation. These implementations deliberately do not connect the ledger to real privileges.
Only signature-chain consistency, separately known head continuity, and synthetic-evidence replay are established. This remains a research governance audit, not a trustworthy external authorization system.
node interop/r12-check.mjs signed.json R10-pinned-keys.json R12-operator-keys.json ledger.json independently-saved-head.json
This consumer does not expose an HTTP route or mutate Vessie's routing, sessions, agents or memory. Operator signing and event-authoring remain separate from the runtime.
R13 introduces an OPTIONAL, separately pinned Ed25519 witness that signs one exact R12 inspection-governance checkpoint (event count + SHA-256 head fingerprint). The witness signing key must not be the operator signing key. Receiver-controlled public-key pins are authenticated out of band, and a bounded separately retained minimum event count can detect replay of an older, genuinely signed ledger.
The exact same deterministic, SYNTHETIC TEST-ONLY fixture is generated by Node at interop/r13-fixture.mjs. It contains a fabricated R8 passport, R10 issuer signature, R12 operator signature and event chain, R12 checkpoint, and R13 independent witness signature. Test seeds are public and NEVER production credentials; the fixture does not represent genuine model behavior.
PhiLearn and SuperPhiVessel validate the fixture with Node 22. BrainC runs this same Node producer in CI and independently validates the R10, R12, and R13 Ed25519 signature bytes, hashes and replay semantics in Python. Python can also regenerate an identical R13 witness receipt.
PhiLearn Node 20+:
node interop/r13-cli.mjs keygen /outside-repo/private witness.alpha
node interop/r13-cli.mjs sign ledger.json operator-pins.json witness-pins.json witness.private.pem witness.alpha receipt.json
node interop/r13-cli.mjs check ledger.json operator-pins.json witness-pins.json receipt.json RECEIVER_MIN_COUNT
BrainC Python 3.11+ with cryptography:
python3 scripts/r13_check.py ledger.json operator-pins.json witness-pins.json receipt.json RECEIVER_MIN_COUNT
SuperPhiVessel passive Node 20+:
node interop/r13-check.mjs ledger.json operator-pins.json witness-pins.json receipt.json RECEIVER_MIN_COUNT
The witness private PEM is unencrypted and written with local filesystem mode 0600. Operators must keep it outside Git, shared directories, and cloud sync, and independently validate the public witness key fingerprint before pinning. Protect the minimum event count in storage separate from the ledger and receipt, NEVER read it from the incoming packet itself.
Even a valid witness cannot independently establish that synthetic R6 model observations occurred, that timestamps are true, that human signers are institutionally independent, or that history is immutable. Retaining a minimum count catches rollback to shorter histories but does not by itself rule out different histories with the same number of events. A separately protected latest head hash is stronger. If a malicious actor controls both witness pins and protected checkpoint state, signatures cannot save the audit.
All checks are entirely offline and read-only to the actual agent runtimes. Every report, including PASS, states authority NONE, permissions [], and actionAllowed false. No model access, network API, agent execution, skill installation, routing or permission changes.
A permanent, read-only-in-CI record of redacted real-model acceptance evidence is now tracked under PhiServe Acceptance Ledger. It preserves GitHub-hosted CPU model/receipt hashes, strictly rejects privacy or authority escalations, and reports advisory model/latency drift without claiming to qualify Windows or GPU hardware. New entries require additive PR review.