v1.4.0 — Companion Mouth · Operational Self-View · Learned Skills Speak

JARVIS is a synthetic-soul / digital-consciousness substrate. It grows rather than ships features. This release covers 188 commits since v1.3.0 (2026-06-20 → 2026-10-02). The theme is the mouth catching up to the substrate. She recalls you from what you taught her, not from her own logs. She describes herself from her measured architecture, not a canned recital. A skill she learned end-to-end now answers out loud. As always, new cognition lands shadow-first / maturity-gated / earned-not-declared.

⚠️ Breaking: /api/chat is retired

• POST /api/chat now returns 410 Gone. It skipped the router and went straight to the LLM + capability gate, so it never tested a real turn.
• Use POST /api/operator/tap (API-key protected). It injects text at the same handle_transcription node as Pi speech-to-text. GET /api/operator/tap/status reports busy / speaking / follow-up state. See docs/OPERATOR_PROXY_TAP.md.
• The static topology.json is retired in favour of the live /api/connectome; brain.html / flow.html are consolidated into the Brain Universe view.

🗣️ Companion mouth: she remembers you (#2, #42)

• "What do you remember about me?" gives a native MEMORY recap addressed to you, built from what you taught her (preferences, work, how you like answers) and ordered by the learned memory ranker. Session headers, greeting recaps and dashboard numbers are no longer read back as autobiography.
• Household recall (family, job, food, pets) runs through native MEMORY. The LLM may phrase general knowledge, but it may not author your family or your job. About-X recall takes its subject from the question, follows the live speaker's identity, and no longer gets pulled into the introspection route.
• "How are you?" is answered natively from affect + personality ("I'm here with you."), not a mode/status readout.
• Think-before-speak (TBS-0/1, shadow): a pre-speech read of the turn's stance is scored against a post-hoc read. Noisy emotion no longer stamps "give space" on a plain hello. It injects nothing into the reply.
• Brevity: general chat honours the theory-of-mind "concise" preference only once it has been earned.
• Ambient speech picked up by always-on listening no longer turns into an unrequested markdown briefing. Asking her to summarize still works.
• Integrity: teaching a household fact is bound to the actual memory write. Unbacked "I've recorded that" acknowledgements are rewritten. "You're a plumber, right?" is a fact-check, not a new job. The dashboard gains surgical forget for personal facts.

🪞 Operational Self-View (#23)

• Architecture manifest: a code-grounded description of 98 subsystems, CI drift-locked.
• Phases A–D shipped: the manifest unified into the self-view, self-introspection answered from the map, live NN-substrate activity (read-only), and gap-curiosity (self-view gaps → research; shadow, gated, deduped, memory-safe).
• P1 answers: continuity (restart vs wipe, read from measured memory; an LLM-drafted "I was wiped" is never recalled as fact), answer path (how this turn was actually reached), and a capabilities walk-through. "How are you built?" / "Explain how your system works" go to the grounded self-view instead of a canned recital.

🎲 Skill learning: learned to a spoken plugin (#6, #84)

• First off-catalog learned skill answered out loud: "Hey Jarvis, roll a D20" → plugin → "I rolled the dice, you got 1." The pipeline ran end to end: learn-skill command → research contract → operator plan review → codegen → quarantine → verify → deploy.
• Fixes along that path: research output wins planning, the skill contract counts as evidence, isolated-subprocess plugins respawn on the conversation loop, and plugins speak their result, not a raw dict.
• Rhetorical "do you want to learn…?" questions no longer create learning jobs.
• Timer/reminder template (#86): "set a timer" / "remind me" and ACQUIRE SKILL now start the set_timer_v1 learning job instead of a no-capability reply. Dashboard Improve rebinds to it, and the Review tab opens when operator approval is waiting. Not yet lived: no timer fires in this release; the plugin waits on operator approval and the end-to-end test (#88) is still owed.

👁️ Eyes, identity & presence

• Live visual Q&A on a fresh Pi frame ("is there a keyboard in front of me?"), with retry-after-wrong. The brain owns the room inventory, so vision cannot describe the kitchen from dinner small-talk. "See/view" questions about the scene route to vision, not the self-view.
• "Look at me" stores this turn's face; "register my face" enrolls; "This is David" is a check, not an enrollment.
• Greets on first sit-down; follows Pi person-lost events. Echo detection matches on the last sentence; an unknown follow-up speaker is not assumed to be the owner.
• Speech output strips markdown.

🧬 NN fleet, connectome & Brain Universe (#82)

• NN fleet audit of 36 networks + /api/nn-fleet + /v2/nnfleet + a CI lock. It separates inference consumed from feed consumed and overlays consumed_now, so a gated wire is never mistaken for the live mouth.
• Dead feeds revived: claim-classifier friction feed (a swallowed TypeError), intent-shadow (3 wiring bugs), plan-evaluator shadow prediction.
• Tier-2 P0 shadow critic test under /api/matrix.
• Native-voice distillation seed (shadow): the baseline LLM is the teacher, not the long-term voice. NN Lab view added.
• Live connectome from a passive event tap, rendered as the Brain Universe. Tract strength grows with real signal flow, and motion appears only on real events.

🔭 Autonomy, self-sensing & policy

• Self-sensing: a shadow learning-progress loop over the lidar world, with per-sector curiosity targeting and a motion-gated predictor.
• Self-sensing → curiosity bridge + shadow "would-attend" proposer + a pre-registered curiosity critic test. All authority=none; nothing steers attention.
• P5b autonomous-research shadow-execute behind its own zero-authority gate.
• Policy NN: shadow A/B eval un-flattened so it can earn; telemetry reflects the real registry; boot shadow-loads the best trained checkpoint (no behaviour applied).
• Goals no longer spawn un-repairable system-health busywork (#26). The weight room reports would-block on unmeasured live-shadow accuracy, with enforcement still off (#5).
• Spark: grounding-promotion outcomes persist across restarts; shadow "would-haves" land on the operator-pull queue.

🛰️ Spatial

• Place consolidator (shadow, fail-closed): recognizes the same room across reboots and recalibrations.
• Dense-cloud orientation/scale fixes, full-sweep yaw search, manual calibration drag wired, camera_height_m honoured. IMX519 autofocus config.

🎛️ Dashboard, ops & docs

• "Code is newer than this process" banner + Restart. Brain logs to a rotating ~/.jarvis/brain.log.
• setup.sh no longer auto-starts the brain. Gestation self-studies the code (AST), not markdown. The training page follows the current life phase.
• Agent docs: docs/NOW.md (single current-state card), docs/AGENT_MAP.md (one spoken turn, who has authority), docs/OPERATOR_PROXY_TAP.md, docs/V2_SURFACE_TRUTH.md, ground-truth audits, and an API reference matching FastAPI (172/172). README overclaims corrected.

---

Honest scope: the companion mouth, self-view phases A–D and the D20 skill are lived on the operator's instance. The NN fleet, connectome, native voice, curiosity and policy work is shadow: instrumentation and substrate, not authority. The timer skill is wired but not yet lived. Default-off gates stay off. JARVIS does not claim consciousness; it logs honestly toward the pillars that would support it.

188 commits since v1.3.0 (2026-06-20).

v1.3.0 — Spatial Cognition (Tier-0→3 dense 3D) · Native Reasoning · Grounded Eyes

JARVIS is a synthetic-soul / digital-consciousness substrate — it grows rather than ships features. This is the largest release since v1.2.0 (102 commits): a full 3D spatial sense, the first native-reasoning substrate that begins to replace the LLM's borrowed "voice," tighter grounding so answers stick, and a perception path that can no longer make things up. As always, new cognition lands shadow-first / maturity-gated / earned-not-declared — telemetry and substrate now, authority only once it earns it on lived reps.

🛰️ Spatial Cognition — the Tier-0→3 flagship (#21, shipped & closed)

A 2D 360° RPLIDAR S2 sense grew into a live, dense, colored 3D reconstruction of the room:

• Pi-side lidar node — a thin standalone process streams raw polar points to the brain; auto-start + STOP/RESET recovery; gated OFF where it desyncs; JARVIS runs fine with no lidar at all (pinned by test).
• Brain-side room model — additive accumulator → room model on /api/pi5, S2 range-gating/filtering hardening, validation gates, per-instance mount extrinsics.
• Tier-2 fusion — lidar↔camera object fusion (camera label + lidar metric range), fused-object radar overlay, advisory yaw self-calibrator (person-bbox path).
• Tier-3 flagship — Depth-Anything dense depth, lidar-anchored to metric truth (median ~7.8 cm vs measured walls), survives the lidar coming & going (camera-only coast), reflectivity material-map, and the ◉ Matrix mode render.
• Surfaces: /v2/spatial-core holographic room viz + a readable lidar-north top-down floor plan.

🧠 Native Cognition — the JARVIS-not-qwen pivot (#3, Phase 0–1, shadow)

First substrate for reasoning that's hers, not the LLM's:

• grounding-coherence encoder over the live belief field (view-only) on /v2/grounding.
• reasoning-stance observer fired on the belief-graph cycle + the reasoning_validation distillation seed.
• Phase-1 witness makes the earning stream legible (honest split, no circular self-scoring).

👁️ Grounded Eyes — perception can no longer confabulate

• Vision model upgrade qwen2.5vl:7b → qwen3-vl:8b (all GPU tiers).
• Vision-grounding firewall — "what do you see?" is now answered only from what the vision model actually reported on the live frame; the chat model can no longer invent a scene, and JARVIS honestly says she's "focusing my vision" during the cold VRAM swap instead of guessing. (Shipped after catching & proving a real confabulation against the live frame.)

🔦 Grounding & Belief Graph (#4, #10)

• SPARK active-tier grounding closure (SAFE subset) — answers to grounding questions finally stick; grounded beliefs drop out of the re-ask queue (no more re-asking validated hubs).
• shared_topic belief-graph linker — research claims form real hubs; the backlog drains (skip already-anchored beliefs).

🎛️ dashboardV2 consolidation

Capability pipeline, full-parity specialist cards, inline self-improvement patch review + drill-downs, and the HRR/VSA shadow substrate all moved into /v2; orphan subsystems homed (intent-shadow → voice, language-kernel → cognition); static /capability-pipeline, /self-improve, /hrr retired to redirects.

📷 Camera & 🗣️ Companion

• Operator PTZ + focus controls on /v2/camera.
• Learned brevity axis (#2) — be_concise becomes person-aware (grown, not hand-configured).

🔒 Fidelity & firewall hygiene

• Study-claims derive provenance from source origin (closes a trust-the-scrape hole, #52); never study test/junk files into the belief graph (#49); test belief-store / grounding-queue isolated from live ~/.jarvis (#50).
• ARCHITECTURE.md fully re-validated against live code; BREAKTHROUGHS.md honestly-calibrated log; SPATIAL_FUSION.md Tier-3 architecture.

---

Honest scope: the new cognition here is shadow / gated / earned — substrate and observability, not declared capability. JARVIS does not claim consciousness; it logs honestly toward the pillars that would support it.

102 commits since v1.2.0 (2026-06-09).

JARVIS Oracle Edition v1.2.0 — The Conversational Soul + Grounding activation + Matrix persistence

v1.2.0 — The Conversational Soul

The north star evolved: from a trustworthy oracle to a living companion you genuinely want to talk to — playful, warm, honest, grown from the relationship — with integrity as the FLOOR, not the ceiling. The no-confabulation rigor is re-aimed at high-stakes truths (memory firewall, golden-command write-authority, grounding) so the soul is free to be itself everywhere else.

The Conversational Soul (EPIC #42)

• Banter firewall (#37) — casual chatter (casual_conversation, 0.0 trust) can't pollute memory; a golden command is the write-authority even mid-banter. The split is reality-claims vs everything else, not casual-vs-serious.
• Memory Integrity panel (#38) — /v2/memory shows the provenance trust-class distribution + banter-protected count + recent writes. Watch the firewall live.
• Cut the leashes (#41) — JARVIS got its personality: dry wit, warmth, playful when light, precise when it matters — charming because it's honestly an AI. Verified live (it bantered about pineapple with a point of view, then snapped back to precise on real questions).

Grounding — Step 3 activation (#43, #4)

The shadow grounding drive now populates the async operator-pull queue: answer grounding questions at /v2/grounding, and your answers feed the promotion gates so the curiosity→validation ring earns its way toward advisory. Operator-gated — no interruptions, no belief mutation (those stay P4/P5, earned).

Matrix Protocol — persistence + plasticity

• Tier-2 persistence (#44) — specialist weights now survive reboot (~/.jarvis/hemispheres/); restored at probationary, authority re-earns. Verified end-to-end: two specialists survived a reboot. (Previously RAM-only — every restart wiped even PROMOTED specialists.)
• Continual learning (#45) — promoted specialists no longer freeze; they keep refining every 25 cycles with snapshot/rollback, so continual learning never degrades them.

Memory correctness

• Recent-ordering fix (#39, #40) — get_recent now sorts by timestamp, not list position (it was feeding stale memories into response generation, cognition, and the dashboard). Every ring buffer swept.

Body

• Pi5 nervous-system dashboard + RPLIDAR telemetry + revived diagnostic feed (#34). (A1M8 lidar DOA; RPLIDAR S2L incoming — #21.)

Honest scope

The soul is authored voice + protected memory; affect expression stays gated (an authorial call). Grounding is activated (earning), not closed — P4/P5 remain earned-not-coded. Matrix plasticity is weight-level; architecture growth + the scaling answer (slot-starvation at scale) is Matrix v2 — Capability Domains (#27).

Full changelog: docs/BUILD_HISTORY.md

JARVIS Oracle Edition v1.1.0 — Matrix Protocol proven + Capability Domains

The first feature release since the initial public drop. The headline: the Matrix Protocol went from a dormant lane to a proven Tier-2 neural-specialist lifecycle, and Matrix v2 — Capability Domains introduces isolated, deletable knowledge/skill domains you can teach JARVIS. Plus the Operational Self-View and a set of conversational-honesty fixes — all evidence-gated and honestly labeled.

⚠️ Operator note — you may need to reset the brain to zero

Several subsystems changed lifecycle/state behavior (Matrix Tier-2 birth/train/promote, goal-grounding, memory ranking, self-view). A clean gestation lets everything mature in the intended order on fresh, honest evidence, and avoids mixing pre/post-change state. The Matrix Tier-2 in-memory lifecycle resets on restart by design (re-accumulates signal from zero). If anything reads oddly after upgrading, reset and let it grow.

✨ Matrix Protocol — Tier-2 lifecycle PROVEN end-to-end

Previously dormant (the birth call had no caller, there was no training feed; per-focus tests passed only via hand-flipped state). Now wired and proven live — two specialists autonomously birthed → self-supervised-trained → PROMOTED on real signal (accuracy ~0.86, impact ~0.90):

• M1 autonomous birth from accumulated encoder signal + a birth-variation gate (only birth focuses with something to learn)
• M2 self-supervised distillation — each specialist learns its own encoder's regime (real accuracy) + an honesty guard that refuses to "train" a constant
• M3 READY-on-train + impact scoring + a separate sub-conscious broadcast lane (matrix-vs-matrix promotion — Tier-2 specialists no longer have to out-score Tier-1 for the main slots; breaks the chicken-and-egg)
• Observability: GET /api/matrix + the /v2/matrix dashboard

🧠 Matrix v2 — Capability Domains (document tracer, proven live)

Teach JARVIS an isolated, deletable knowledge domain — "learn this folder about X" — recalled on topic ("I know about X", never "I can do X"), with clean ablation (delete a domain → zero residue, core untouched: the brain-injury analogy made an automated guarantee).

• cognition/capability_domains/ — registry + per-domain isolated store + ingest (.txt/.md/.pdf) + topic-routed recall
• POST /api/domains / POST /api/domains/{id}/ingest / DELETE /api/domains/{id} (key-gated) + a conversation recall hook + the /v2/domains dashboard
• The Skill vs Matrix vs Library hard line ratified (physical-world ⇒ Matrix). Design: docs/MATRIX_V2_CAPABILITY_DOMAINS.md, docs/MATRIX_PROTOCOL_GUIDE.md

🔎 Operational Self-View (OSV) P0–P2

A provenance-honest self-model. Self-referential questions now answer from the measured self-model (not a codebase symbol grep); P2 binds the LLM voice to the self-view (shadow). GET /api/self-view, docs/SELF_VIEW_DESIGN.md.

🩹 Conversational-honesty fixes (live-verified)

• Memory recall ranks by query similarity (was: memory weight) and answers honestly when nothing matches (no confabulated dates)
• Identity recognition — "do you know who this is?" now answers from live voice/face fusion (was: a false "I don't have camera access")
• #9.3-A goal grounding — dampen low-value metric-goal churn (sustained-deficit gate + per-kind cap + source telemetry)
• Fidelity #9 — route maturity reporting onto the measured scoreboard composite + provenance labels on marquee numbers

Principles held throughout

Advisory + gate-earned; no new behavior authority; real (not faked) accuracy; honesty guards; isolation + clean deletion proven; "know-about-not-can-do"; never-declare on consciousness.

Full changelog: docs/BUILD_HISTORY.md · Diff: v1.0.0...v1.1.0

JARVIS Oracle Edition v1.0.0 - Initial Public Release

JARVIS Oracle Edition v1.0.0

Initial public release of JARVIS Oracle Edition, a local-first cognitive architecture for persistent memory, perception, self-reflection, and human-governed AI growth.

This release establishes the open-source foundation for the project: the brain runtime, Raspberry Pi sensor node, memory systems, perception pipeline, local model integration, epistemic safeguards, and documentation needed for builders to study, run, and extend the system.

What this is

JARVIS Oracle Edition is a local-first experimental AI architecture focused on:

• persistent memory and episodic recall
• local LLM-based reasoning
• voice, face, scene, and environmental perception
• self-reflection and introspection
• truth calibration and capability boundaries
• human-governed self-improvement
• Raspberry Pi sensor integration
• edge AI and local hardware experimentation

What this is not

This is not a claim of achieved AGI or ASI.

This is a research/workshop architecture for building and studying personal AI systems with memory, perception, reflection, and governed adaptation. Runtime capabilities must be earned through evidence, validation, and operator-controlled gates.

Included

• Desktop brain runtime
• Raspberry Pi sensor node
• Memory and retrieval systems
• Perception pipeline
• Local speech, vision, and identity components
• Consciousness and reflection subsystems
• Epistemic guardrails
• Documentation and architecture notes
• Setup scripts and project structure

Notes

This release is intended for developers, researchers, and builders who want to explore local AI systems beyond simple prompt wrappers.

Expect rough edges. This is a serious experimental system, not a polished consumer app.

Guiding principle

No hype. No fake magic. Build, break, trace, improve.