Changelog

Sacramento PaintBench is a castor bench benchmark

• castor bench sacpaint — the standalone sacpaint package moves into the runtime as castor.bench.sacpaint. A robot with a pen must reproduce a photograph of Sacramento from camera feedback; any photo of the finished sheet scores offline with castor bench sacpaint score PHOTO. Install the eval stack with pip install "opencastor[paintbench]"; a plain install pulls no eval framework.
• The reference is the photograph — the model sees the original picture; the scorers read a stroke skeleton traced over its landmarks, and every EvalLog carries both hashes.
• Media as categories — every score carries medium (pen, virtual, sim), each ranked on its own. -E medium=virtual -E calibration=easel runs an arm with no paper and no pen: the canvas is inked from where the arm measured its tip. First real-arm run, an SO-ARM101 on 2026-09-09: 649 moves, 62 minutes, composite 0.82.
• The runtime’s first Python gateway client — RCAN envelopes to /v1/invoke with signed receipts kept as JSONL, a deny raised as a safety abort with the driver’s sub-code, a reach that ran but missed told apart from a fault.
• Console — /eval/reference.png serves the packaged photograph; the 3.2.0 /eval/* routes for the phone’s frames and corners are now in the changelog too.

The Microduck is the reference robot, and the ten-minute goal is a command

• castor bench ten-minutes --robot microduck — seven timestamped checkpoints from “package installed” to “robot.stop observed after silence, naming which deadman fired”, a JSON record with versions, shas, duck identity and every wire line, an EvalLog export that inspect-robots reads back byte for byte, and three targets: a mock robotd, Pollen’s own simulator running the real robotd, and a real duck. Today the sim passes C1 to C6 in 3.5 seconds; nothing has touched a physical duck yet, and the record says so.
• Five wire faults fixed — castor duck had never been run against a real robotd. Four keys read from the wrong struct, and robot.subscribe had never once succeeded. Two independent runs against the live daemon found the same fifth fault.
• castor up --archetype microduck — six units including the bridge as duckbridge.service, one token the phone and OpenCastor share, port 7788 agreed in four places.
• transport: webrtc — drive a duck over mediad’s own network transport with no SSH key, no robot group and no reboot (pip install "opencastor[microduck-webrtc]"). The guide now states plainly that mediad is unauthenticated on 0.0.0.0.
• castor doctor knows ducks and calls a mock driver a failure; the duck’s LLM tools finally register; castor duck writes a manifest castor run accepts, never prints “Duck ready.” for a duck that did not answer, and says its speed envelope out loud.
• RC car — castor doctor reads the PCA9685 back and reports a chip that forgets its configuration, found on the first hardware contact; the bring-up checklist gained that step.

Box to driving in under ten minutes, and the Microduck

Published on PyPI as 1!3.1.0: install with pip install "opencastor>=3.1". The epoch is what makes pip prefer this line over ninety-seven older date-versioned releases that had no castor up in them.

• Real wheels are one question away — castor up asks, with the wheels-off-the-ground warning, when it detects a PCA9685; rc-car-actuator is a core dependency; the template’s channels match every reference vehicle; the manifest can no longer draft a stop (duration_s has no default).
• Discovery ships — a fifth unit advertises _opencastor._tcp with rrn, ports, name and manifest path; castor discovery check proves it from the wire.
• castor doctor reads the robot it provisioned — real ports, /dev/i2c-1, simulated wheels, mDNS, USB current versus a streaming camera, every capability gap, each with a pasteable fix and a non-zero exit when the car cannot move.
• The Pi image enables I2C at first boot and ships as one file; the install and port docs agree with the code; castor init asks the robot’s shape first.

The Microduck, and one command from bare host to paired robot

• Pollen Microduck support — a 25 cm walking biped whose motion comes from RL policies executed by its own robotd daemon at 50 Hz. OpenCastor is an intent client over robotd's JSON-RPC 2.0 NDJSON socket; robotd keeps exclusive ownership of the motor bus. Hardware guide →
• castor duck — one command from a duck on the desk to a duck under an LLM: find it, prove we can reach it, ask robotd how it feels, write the config. Discovery tries five things (local socket, hostnames, duckctl ip over Bluetooth, mDNS, ARP with --deep) because Pollen's own docs admit mDNS “resolves when it feels like it”. The two things that can block you — SSH access and robot group membership — each get their exact fix printed.
• The whole vocabulary — reading duck-ipc-proto turned up twenty-two robot.* methods against the eleven the driver knew. robot.do now carries the five scripted skills (ground_pick, both kicks, sit_toggle, roulade); robot.sound carries the seven-sound voice bank including the held wheee; robot.mouth and robot.pose join the twist as continuous intents.
• Choreography — robot.do runs exactly one skill, so there is no way to say “walk to the ball, line up, knock it toward the couch, then celebrate”. Every verb exists; the sentence does not. Thirteen primitives now carry how long a move occupies the duck and whether it holds the robot exclusively, and thirteen routines compose them — fetch, greet, patrol, nudge, celebrate, settle. castor duck do "greet me, then patrol the room".
• castor up — the ten-minute contract as a command: detect hardware → pick archetype → generate the robot home (signed ROBOT.md, rcan.yaml, bearers, policy, keys, runtime) → start the systemd units → detect a brain → write the pairing QR. Non-interactive and idempotent throughout: every question it could ask is one it can answer by looking at the machine.
• Universal-link pairing — the QR is a link any camera understands, with the server half to match. The QR states what this robot can do, so a car stops arriving as an arm.
• DuckKit — the duck's brain as a zero-dependency Swift package, so a phone runs the real trained policy with no robot in the room. github.com/craigm26/duckkit
• Fixed — rc-car-actuator was a hard requirement that broke pip install opencastor for weeks; it is now the [rc-car] extra. The peripheral scanner could not see a board that had just been plugged in. Two tools shared one file and disagreed about its shape, and --force ate a live key.

A plan is a proposal. The choreographer computes no permission — every motion goes out through the driver and the SafetyLayer, the duck's own limits apply on top and come back in robot.state.limited_by, and steps abort on a fall or a flat battery. robotd remains the only thing that decides what a servo does.

Auto-detection must never offer a LoRa radio as a servo bus

• Detection put a green tick next to /dev/ttyACM0 — an ESP32 LoRa radio — labelled “controller board”, and castor arm verify therefore defaulted to it. A denylist is now checked before any allowlist match: Espressif/Adafruit/Arduino VIDs plus product-string matches on heltec|lora|meshtastic|meshcore|esp32|rak|t-beam.
• The ordering matters: pyserial asserts DTR and RTS on open, which is esptool's reset-into-download-mode sequence on an ESP32-S3 native-USB CDC. Probing such a board does not merely fail — it can knock it out of operation.

Conformance re-verified against the frozen spec

• All 195 tests across the conformance, compliance and RCAN 3 compliance suites pass against the post-freeze rcan-spec. No code changes — the release memorializes the verification outcome with a signed version-tuple envelope.

Two CLI gaps caught while recording the peer-runtime demo

• castor init now matches the v3.0 wizard. The argparse registration was still the legacy shape and did not hit the wizard's expected attribute names.
• castor validate ROBOT.md now works. The old handler always fed the target through yaml.safe_load, which errored on any ROBOT.md because the frontmatter-markdown format contains two --- documents.

ROBOT.md-native, full RCAN 3.x peer runtime

• Version scheme change — CalVer (2026.4.23.0) → SemVer (3.0.0). The major signals peer-runtime alignment with rcan-spec 3.2, rcan-py 3.3 and rcan-ts 3.4.
• Legacy .rcan.yaml input removed — runtime, register and compliance commands read ROBOT.md (v3.2 frontmatter) as the canonical source. The 2.x config path is rejected at ingress with a castor migrate hint.
• castor/rcan3/ — the peer-runtime integration package: ROBOT.md reader with runtime selection, ML-DSA-65 + Ed25519 hybrid identity (private keys written 0o600 atomically, no chmod race), dict-level signer, async RRF client, the §22–26 compliance facade, and the Harness Protocol.
• Removed — castor.config_validation (superseded by rcan-py's rcan.validate) and five legacy castor rrf subcommands.

RCAN v2.2 — Post-Quantum Primary • Dual-Brain • LLMFit • Hardware Stats

• RCAN v2.2 — ML-DSA-65 primary — Ed25519 fully removed (not sunset-dated, removed). signature.alg = "ml-dsa-65" is the sole wire format. Legacy pq_sig field removed. Q-Day timeline accelerated from 2028 to 2026.
• rcan-py v1.2.1 — 675 tests, MLDSAKeyPair, sign_message()/verify_message() with require_pq=True default. KeyPair stub redirects to MLDSAKeyPair with DeprecationWarning.
• rcan-ts v1.2.1 — 545 tests, SignatureBlock.alg narrowed to literal "ml-dsa-65", pqSig removed from RCANMessageData.
• Dual-brain architecture — brain_reactive (VLA/SLM, fast) + brain_planning (Claude/GPT, slow). Confidence gate at 0.60 — below escalates to planning brain. Wire field: brain added to action dict.
• Production data wipe + re-registration — Firestore robots collection + RRF KV cleared. Bob (RRN-000000000001) re-registered with new ML-DSA-65 key, dual-brain configured.
• castor conformance L5 — Full RCAN v2.2 compliance report across 6 domains. Score formula: max(0, 100 − failed×10 − warned×3). Bob scores 100/100. GET /api/conformance endpoint added.
• castor iso-check — ISO 13482, ISO 10218-2, ISO 42001, EU AI Act Art. 12/13/14/15/16j self-assessment. Bob: COMPLIANT (self-declared). EU AI Act deadline: August 2, 2026.
• castor llmfit — Model-fit checker: BF16 weight estimate + KV cache + overhead vs device RAM. TurboQuant 2.6× compression. MoE-aware. psutil-optional with /proc/meminfo fallback.
• castor/system_info.py — Real hardware stats in /api/status: RAM, disk, CPU temp, NPU detection (Hailo-8/8L, Coral TPU, Qualcomm HTP), GPU, active model size, context window, llmfit headroom, tok/s. Auto-flows to Firestore via 30s telemetry push.
• opencastor-client hardware stats — Robot detail screen shows live RAM, disk, CPU °C, NPU+TOPS, GPU, model+size, TurboQuant KV status, llmfit headroom, and tok/s as pill chips.
• DISCOVER iso_conformance block — RCAN DISCOVER (msg_type=1) now includes iso_conformance block. rcan_version bumped to "2.2" in all gateway endpoints.
• OpenCastor gateway: GET /api/attest, POST /api/attest/verify, GET /api/sbom, GET /api/sbom/download, GET /api/conformance.
• opencastor-client: AttestationCard + compliance report button on robot detail. /robot/:rrn/attestation alias route.

Competition Engine — Sprint, Threshold Race & Bracket Seasons

• Sprint format — time-boxed competitions with announced prize pools, tiered payout (50/30/20), anti-snipe 1h lock (castor/competitions/sprint.py)
• Threshold Race — jackpot mechanic, first robot to hit target score wins, 3× independent verification re-run with 2% tolerance (castor/competitions/threshold_race.py)
• Model×Hardware Bracket Seasons — monthly seasons, 5 hardware×model classes, class champions (2000/1000 credits), grand champion bonus 5000 credits (castor/competitions/bracket_season.py)
• Competition API endpoints: GET/POST /api/competitions, /api/competitions/races, /api/seasons/current
• Gemini 2.5 migration — all Gemini 2.0/1.5 model strings updated ahead of June 2026 deprecation
• Contribute 10× throughput — Firestore client and queue state now cached; fetch latency 4.6s → 14ms
• Auto-start contribute on gateway startup via agent.contribute config

castor contribute — Idle Compute for Distributed Science

• castor contribute start|stop|status|projects CLI — donate idle NPU/CPU cycles to climate modeling, biodiversity monitoring, protein folding, and humanitarian AI
• BOINC, Folding@home, and OpenCastor coordinator support with configurable project allowlists
• P66-safe: any incoming command preempts contribute work within 100ms via ContributeCancel
• Hardware-aware work unit matching: Hailo-8L HEF models, CPU fallback, optional camera/microphone sensor projects
• Power-capped (max_power_w) and CPU-throttled (max_cpu_pct) to respect thermal/battery budgets
• GET /api/contribute endpoint — live status, stats, and hardware profile
• Fleet dashboard integration: per-robot contribution stats, active project indicator, enable/disable toggle
• Privacy mode respected — contribute auto-pauses when castor privacy on is active

Harness autoresearch live · editor UX overhaul · hardware-profile optimization

• Autoresearch pipeline live — Gemini 2.0 Flash generates harness candidates nightly; winner pushes directly to main (no PRs). First champion: lower_cost (cost_gate_usd 0.05→0.01, score 0.9101)
• Harness editor overhaul — smart type-aware block placement; zoom controls (+/−/fit); drag-to-connect edges; block info overlay (ℹ️); auto-connect on add; list↔flow graph sync
• Connector type picker — tap any edge midpoint to label it YES/NO/loop/error/timeout/data/fallback or delete it
• Capabilities screen fix — crash on gray screen resolved (safe Map→List cast); Cloud Function 403 resolved (firebase_uid ownership check)
• Hardware-profile harness optimization — roadmap: contribute fleet runs tier-specific evals; per-profile champions for Pi5+Hailo, Pi4, Waveshare, Unitree, server. Blog post
• castor provider CLI — castor provider auth/list/status for gated model providers
• Harness per-layer provider routing — specify model: provider/name per layer; get_provider_for_layer() resolves with fallback
• PyPI v2026.3.20.4 published via OIDC Trusted Publisher; RCAN schema relaxed (40/40 configs pass)

RCAN v1.10 — Competition Protocol (types 37–40)

• RCAN v1.10 canonical MessageType table — 40 message types. Added COMPETITION_ENTER (37), COMPETITION_SCORE (38), SEASON_STANDING (39), PERSONAL_RESEARCH_RESULT (40).
• Competition protocol — robots can announce competition entry, publish verified scores, receive fleet-broadcast season standings, and log private personal research results.
• rcan-py v0.9.0 — rcan/competition.py, 32 new tests, 663 total. SPEC_VERSION 1.10.0.
• rcan-ts v0.9.0 — src/competition.ts, 28 new tests, 499 total. SPEC_VERSION 1.10.0.
• rcan-spec — competitions.astro spec page with full protocol, payload schemas, and wire examples.
• OpenCastor runtime accepts rcan_version: "1.10" / "1.10.0" in robot configs.

RCAN v1.8 — Canonical MessageType & Idle Compute Contribution

• RCAN v1.8 canonical MessageType table — 36 message types with fixed integer assignments. Single source of truth across spec, rcan-py, and rcan-ts. Zero drift.
• castor contribute — donate idle NPU/GPU/CPU compute to distributed science. BOINC coordinator, NPU-optimized runner (Hailo-8L), fleet coordination, thermal throttling, P66 preemption.
• CLI: castor contribute start/stop/status/history, castor peer-test with transport comparison
• API: /api/contribute/start, /api/contribute/stop, /api/contribute/history
• RCAN-MQTT transport wired into runtime — rcan/{rrn}/in, rcan/{rrn}/out, rcan/estop topics
• Software attestation — pip RECORD hash check, config SHA256 baseline, git cleanliness; systemd service template
• Security fixes — RCAN-Signature bypass, None-principal scope skip, WebSocket JWT bypass, SSRF, LoA default hardening
• rcan-py v0.7.0 (609 tests), rcan-ts v0.7.0 (447 tests), opencastor-client v1.1.0+2
• Fleet contribution dashboard, history timeline, project selection UI in Flutter client
• Blog: "From Robot Runtime to Climate Sensor"

RCAN v1.6 — Federation, Constrained Transports, Multi-Modal, Identity LoA

• GAP-16: Registry Federation Protocol — FEDERATION_SYNC wire format, cross-registry JWT trust chains, 3-tier registry hierarchy (root/authoritative/community), DNSSEC trust anchor discovery
• GAP-17: Bandwidth-constrained transports — RCAN-Minimal (32-byte ESTOP for LoRa/BLE), RCAN-Compact (CBOR, <200 bytes ESTOP), BLE L2CAP framing
• GAP-18: Multi-modal payloads — media_chunks[] field, SHA-256 audit trail, streaming mode, TRAINING_DATA now requires media payload
• GAP-14: Human identity LoA 1/2/3 — FIDO2 binding, registry trust tiers, min_loa_for_control in P66 manifest
• All 22 original security/safety audit gaps addressed
• castor setup wizard: interactive onboarding with inline QR codes, castor bridge setup, castor fleet-link
• rcan-py v0.6.0 (587 tests), rcan-ts v0.6.0 (405 tests)

RCAN v1.5 — Production Safety Hardening (18 gaps addressed)

• GAP-03: Replay attack prevention — 30s sliding window seen-set, 10s for safety messages
• GAP-02: Robot identity revocation — ROBOT_REVOCATION broadcast, 1h TTL cache
• GAP-01: Command delegation chains — 4-hop max, Ed25519-signed, audit trail
• GAP-05: Consent wire protocol — CONSENT_REQUEST/GRANT/DENY on-wire format
• GAP-11: ESTOP QoS exactly-once delivery — retry until ACK, SafetyHaltError on timeout
• GAP-06: Offline operation mode — 300s grace, ESTOP always allowed
• GAP-10: Training data consent — EU AI Act Art.10, DataCategory enum
• GAP-08: Cloud relay identity audit — sender_type: cloud_function
• Protocol 66 conformance: 87% → 94%
• rcan-py v0.5.0 (430 tests), rcan-ts v0.5.0 (311 tests)