The record, in order
Evolution
How JARVIS got here, from a 2024 voice prototype to 60 dated milestones across three repositories. Every date comes from a commit, a ledger or file timestamps, and periods without evidence are shown as gaps instead of being filled in.
- Milestones
- 60
- Eras
- 7
- First dated artefact
- 17 Apr 2024
- Commits, production branch
- 3,272
Era 01 / 17 Apr 2024 to 18 Apr 2024
First voice
A browser prototype with a wake word, speech in and out, and a command table.
3 dated milestones
- April 2024
- Original concept
Particle-sphere 'presence' animation
The first visual element was an animated canvas particle sphere (www/script.js) for a browser-based assistant UI.
Why it mattered. It is the ancestor of the orb, the 'intelligence core' and the HUD centre that recur in every later generation.
- Original concept
Web UI and launcher for the first prototype
main.py started an Eel web app in an Edge app window, with a mic button, a SiriWave speaking animation and a chat panel.
Why it mattered. This set the lasting product shape: a voice-first assistant with a cinematic on-screen presence.
- First voice / early architecture
First voice loop and command dispatcher
Surviving 2024 bytecode shows these pieces: cloud speech recognition with ambient-noise adjustment, SAPI5 text-to-speech, a Porcupine wake word ('jarvis'), a SQLite table of app and web commands, YouTube playback, a hosted chatbot.
Why it mattered. This was the first working voice round-trip, a plain command dispatcher with no governance, memory or verification. Every later generation defines itself against it.
Era 02 / 16 Jun 2026 to 28 Jun 2026
The toolkit
A multi-agent toolkit, a command HUD, memory consolidation and the first capability ledger.
7 dated milestones
- June 2026
- Early architecture
The 2024 prototype enters git
The prototype web UI and engine were committed to version control for the first time.
Why it mattered. It marks the end of a roughly 26-month period with no dated evidence and the restart of active development.
- Early architecture
V3repo begins: 'JARVIS toolkit through v118'
334 files (about 86.5k lines) arrived in one commit. They include a standard-library Python server that proxies a hosted cognition model, speech-to-text, vision and a hosted TTS voice, plus 100+ Python 'authority' agents (memory, desktop, research, engineering and others).
Why it mattered. This is the first large, multi-agent architecture. Versions v1 to v118 were never in git and cannot be dated.
- Cognition / Windows operation
Natural Intent Router and first desktop app-control fix
A routed natural-language intent layer was added, the desktop open_app path was repaired, and an operator manual was written.
Why it mattered. It was the first step from keyword commands toward intent routing and governed Windows operation.
- HUD
Command HUD milestone 1
A large command HUD was built, with 454 surfaces.
Why it mattered. It was the first full HUD. Later work kept cutting it down toward one truthful surface.
- Memory / capability discovery
Memory consolidation, a 560-capability ledger and a multi-model router
Three things arrived: a compounding memory consolidation job, a usage ledger covering 560 capabilities, and a multi-model router.
Why it mattered. It was the first attempt at an honest capability inventory, the idea that later became the production truth registry.
- Agents / HUD
Agent council and HUD design constitution
A 25-kernel agent council was added, along with a cinematic HUD asset system and a HUD design constitution ('function over spectacle').
Why it mattered. It introduced multi-agent deliberation and the principle that the HUD must serve function first.
- First voice / perception
Custom voice identity experiment; world model and embodied perception
An owner-gated custom voice identity was tried, and a first world-model and embodied-perception layer was added.
Why it mattered. Voice identity and embodiment became first-class concerns.
Era 03 / 2 Jul 2026 to 12 Jul 2026
Formal phases
Twelve build phases: runtime, policy, memory, voice, missions, recovery, vision, operator.
7 dated milestones
- July 2026
- Qualification
First production-certification audit; heavy asset system removed
A certification pass fixed five defects. A heavy HUD asset system was removed to cure lag, and a policy-gated MCP gateway with 15 tools was added.
Why it mattered. This was the first time the project audited its own claims, and it put performance and policy ahead of spectacle.
- Local model work / voice
Human-like voice interaction and the first local-model training path
Nineteen voice-interaction systems were added, including a turn classifier and adaptive endpointing. A local ONNX embedder tier and a training lab also arrived: persona LoRA, a model merge, and GGUF/Ollama import.
Why it mattered. It was the first local model work, the seed of the later JARVIS-LM campaign.
- HUD / engineering
One always-visible Core HUD; engineering self-rated 'PARTIALLY OPERATIONAL'
Side panels were dropped for a single Core HUD, and engineering readiness was self-certified as only partially operational.
Why it mattered. This set the direction toward one HUD and toward stating partial status plainly.
- HUD
Core HUD V2 cutover: 4 scripts instead of 1,060
The HUD was rebuilt as a small microkernel.
Why it mattered. It was a large simplification and the architecture precedent for the later ONE HUD.
- Memory
Memory database recovered; V3 formal phase programme begins
A malformed memory database was recovered. The V3 baseline (Entry 0) was recorded, and phases P01 (runtime and observability) and P03 (memory, identity, continuity) closed the same day.
Why it mattered. A real memory-integrity incident drove the rule that corruption must be refused, never replaced with an empty store.
- Agents / HUD / daily command center
V3 phases P04 to P10 close
In one day the programme closed: the intent director and persona (P04), mission control with a safety ledger and E-stop chokepoint (P05), diagnostics and recovery (P06), vision and browser operator (P07), HUD V2 polish (P08), the cognitive core and capability graph (P09), and the daily command center (P10).
Why it mattered. It was the first structured coverage of the full product surface, although later audits found the closures optimistic.
- Windows operation / CAD
V3 P11 Safe Windows Operator, tools, CAD, hardware; P12 left unmerged
P11 added governed Windows operation and CAD/hardware tooling. P12 (domain acceptance) was never merged and was recorded as blocker BLK-005.
Why it mattered. Safe computer operation became an explicit, governed capability.
Era 04 / 27 Jul 2026 to 31 Jul 2026
Proof over claims
A constitution and a truth ledger. The final gate was run honestly, and failed honestly.
9 dated milestones
- July 2026 , continued
- Qualification
V3 final gate G0 census: 0 capabilities LIVE
An honest completion-matrix census found no capability that could be certified live. The V3 line ended here.
Why it mattered. This finding produced the anti-fake-LIVE rule. The next repository was founded on it the same day.
- Early architecture (production)
Production repo founded: constitution and truth ledger
The constitution was written to disk, along with an append-only execution ledger and a capability truth registry with 19 truth states and a strict 24-hour LIVE definition. The same work set up owner-actions tracking and CI gates.
Why it mattered. It made the disk, not the model's context, the authoritative contract, and made 'LIVE' mean something checkable.
- Voice / HUD
First production HUD shell and first spoken round trip
The HUD shell came up with its first live probe, and an acceptance test drove a real spoken turn end to end.
Why it mattered. It was the first proven voice round trip in the new architecture.
- Qualification
P01 to P32 claimed in one day; the final gate 'run honestly and failed honestly'
Thirty-two build phases were committed as complete: runtime, governance, memory, voice, cognition, missions, desktop, perception, media, research, engineering, HUD and boot. The P32 final gate then failed.
Why it mattered. It fixed the architecture's skeleton (named subsystems in the jarvis/ package) and showed that 'complete' claims need independent proof.
- Qualification
Forensic audit and the named systems (Build Order 002)
Q01 audited everything forensically. Build Order 002 registered the named systems, including VERITAS, TEMPORAL, TRIBUNAL, NEWTON, PROMETHEUS and SENTINEL, each with a feasibility level kept separate from its truth state.
Why it mattered. It set the rule that the project audits before it extends, and banned describing a simulation as deployed hardware.
- Voice
Production speech output comes up with a pinned voice identity
Production speech output came up with the pinned voice identity, and the connection handshake was taken off the latency path.
Why it mattered. JARVIS had a consistent spoken identity from this point.
- Qualification
Verification Mandate 004, Voice Path Directive 005, AUDIT 03
The verification mandate introduced a nine-point certificate per capability. The voice-path directive defined how JARVIS speaks. AUDIT 03 demoted 70 capabilities, then 20 more in a cascade, because no production entry point reached their code.
Why it mattered. 'Reachable from a real entry point' became a precondition for LIVE. The registry still carries these demotions.
- HUD
ADR-001: the HUD shell
The HUD shell architecture was decided, and a twelve-category HUD fidelity gate replaced the earlier ten.
Why it mattered. HUD work became measurable, with 'the surface that cannot lie' as the organising idea.
- HUD / qualification
Z08 Unified Completion Campaign anchored
The Z08 order was committed. It made one designed HUD the only HUD and required everything to be wired into it. The Z09 (n8n workflow suite) and Z10 (adaptive intelligence) orders were queued the same day.
Why it mattered. It became the longest campaign, about 1,400 commits across phases P01 to P14, re-doing the July phase claims with real proof.
Era 05 / 1 Aug 2026 to 30 Aug 2026
Space and forge
Engineering solvers, spatial cognition, and the printer placed behind the approval path.
9 dated milestones
- August 2026
- Spatial scene / perception
Z11 Spatial Cognition order registered
The spatial cognition, gesture language and direct manipulation order was queued behind Z08.P10.
Why it mattered. It opened the spatial and gesture era.
- Engineering workspace / CAD
Engineering stack: run registry, solvers and FreeCAD geometry
Z08.P09 added an engineering run registry, solver paths and FreeCAD-backed geometry.
Why it mattered. It was the first governed engineering and CAD work in production. Later corrections record a spoken request driving a real CalculiX solve to the engineering panel in tests.
- Cognition
Z12 The Composer, Z13 The Character, Z14 The Observer
Three owner-authorised orders were registered. The Composer wired tool use between understanding and doing; the owner's own audit had found the cognition path was prompt-in, text-out with no tool calling. The Character covered persona behaviours, and The Observer covered noticing everything while speaking only when it matters.
Why it mattered. JARVIS moved from a talker to an actor: typed tool calls, epistemic honesty ('never state something with more confidence than the evidence supports'), and quiet observation.
- Fabrication
Z15 The Forge: the printer under the approval path
The 3D printer transport, telemetry and control were built. The print start gate was built whole but refuses every caller until a pause has been observed on a real running print, and E-stop reaches the heater.
Why it mattered. It was the first irreversible physical capability. PolicyEngine, EStop and ApprovalRegistry became load-bearing in the physical world.
- Spatial scene
Governed spatial cognition complete in software; ADR-003
The camera pipeline landed with an ephemeral authority lease, calibrated targeting, temporal gestures, governed confirmation, exact desktop postconditions and truthful HUD states. Architecture decision ADR-003 was recorded.
Why it mattered. Gesture-driven control was built so that POINT never becomes effect authority and the physical hand runbook stays with the owner.
- Local model work
Z16 JARVIS-LM native cognitive engine order
A future order was filed for a JARVIS-native language model as the primary cognitive engine behind the same persona, HUD, memory and safety systems. The builder transition directive and the n8n rescope (n8n is not the authority for identity or reasoning) came with it.
Why it mattered. It started the local-model campaign: cognition would become a replaceable engine under the existing governed body.
- Spatial scene
Z11 safe local live proof
The real core and HUD were booted on isolated loopback ports with disposable stores. Gesture read paths were exercised, and unauthenticated mutation controls were refused with 403. No camera was opened and no window was moved.
Why it mattered. It showed a live proof that deliberately takes no physical action.
- Cognition
Z10 Adaptive Intelligence closed
The adaptive-intelligence and workflow campaign (35 canonical workflows, nothing activated) closed, and Z08.P13 became active.
Why it mattered. Background workflows were defined but deliberately left inactive.
- Qualification
Z11 closes; Z08.P14 'hostile certification and release' opens
The spatial lane handed off, and P14 opened on a requirement matrix that had measured itself.
Why it mattered. P14 is the release and qualification row. It is still ACTIVE on 2026-09-22.
Era 06 / 5 Sept 2026 to 21 Sept 2026
Body and brain
The ONE HUD, the approved voice, 44 body rows, and a local language model held to hard gates.
16 dated milestones
- September 2026
- HUD
ONE HUD, slice A
One page at '/': the intelligence core at the centre, identity above it, and four permanent indicators. Workspaces became internal layout contexts, not windows.
Why it mattered. It was the end state of the HUD story: one surface where nothing renders without a source and an observed instant.
- LM training
Local-model adapter pipeline and qualification v1.1
An isolated adapter-training pipeline for a 14B open-weight base model (Ministral family) was preserved. A dynamic tool router, curriculum and qualification suite v1.1 were added.
Why it mattered. It was the start of governed LM training in the production repo.
- LM training / qualification
Post-LM gate opens; first sealed LM qualification FAILS
The Post-LM gate opened with a latency lab, typed inlet and HUD widget. The JARVIS-LM sealed qualification recorded a FAIL on hard axes, and hosted cognition stayed in production (Path 1).
Why it mattered. The failure was recorded plainly and the post-LM body build went ahead without waiting on the model.
- C013 voice
C013: the owner-selected final local voice
Voice-lab candidate C013 was integrated, a local neural TTS package (F5-TTS based) running on the owner's GPU. It is verified against its manifest digest, fails closed on mismatch, and keeps the hosted and Piper voices as rollback. First audio was re-measured within contract the next commit.
Why it mattered. The voice identity moved onto local hardware and is 'honest about who spoke': any fallback voice is announced.
- Spatial scene / HUD
Spatial HUD layer inside the ONE HUD
A three.js spatial layer was added inside the ONE HUD, with the pre-spatial baseline measured first. Acceptance journeys 9, 14 and 16 reached LIVE.
Why it mattered. 3D became part of the single surface rather than a separate app.
- Body architecture
Owner body audit and the interaction capability model
The current body was classified category by category (PROVEN_LIVE, IMPLEMENTED_UNPROVEN, BROKEN, BLOCKED, NOT_IMPLEMENTED), with live drives of the desktop, browser, files, research and missions.
Why it mattered. It was an honest map of which parts of the 'body' actually work end to end on the owner's machine.
- Qualification
Critical path scoreboard: full suite 23,984 tests; 9 of 341 LIVE
The full suite at bb54722e collected 23,984 tests: 23,588 passed, 277 failed, 34 errors, 85 skipped. 27 of 28 mandatory gates passed, 9 of 341 capabilities were LIVE, and 1 of 44 A-rows was formally closed.
Why it mattered. It was the most complete recorded measurement of the tree. The failures were partitioned rather than hidden.
- Qualification
Tier 1 mutation testing: 449 of 611 mutants killed (73.5%)
Every one of the 44 Tier 1 modules was measured for mutation score.
Why it mattered. It measures whether the tests can catch real defects. 73.5% is still below the harness's own 90% target.
- Agents / research / fabrication
Post-LM waves three to five: fabrication software, modes, document intelligence, shadow model
Several rows were merged: A01 fabrication software, A22 durable owner modes, A43 local document intelligence, A31 gesture personalisation, A04 cited research viewer, A08 contextual widgets, and A38, where a local candidate model answers real turns in shadow and acts on nothing.
Why it mattered. The body grew around the model, and the local model could be compared without being given authority.
- Windows operation
Governed semantic application operation extended
Semantic application operation was extended through the existing governance.
Why it mattered. General application control grew without removing its controls.
- Body architecture
All 44 post-LM A-rows have integrated bodies
A reconciliation matrix showed every A-row with integrated code and cited evidence paths: 36 IMPLEMENTED, 7 IMPLEMENTED_PRODUCTION_REACHABLE, 1 IMPLEMENTED_INTEGRATED.
Why it mattered. The body was implementation-complete but not release-qualified. The matrix says so explicitly.
- LM training
CP-7 contract freeze and final corpus; final epoch validation FAILS
The model-visible contract (CP-7) and a reviewed final corpus were frozen, and a bounded local trainer was integrated. A candidate trained and reloaded, but its validation failed.
Why it mattered. Training was held to contracts frozen in advance, and failure was recorded instead of re-scoped.
- Windows operation
Bounded native application re-proof
Bounded native application operation evidence was recorded again.
Why it mattered. It is fresh evidence for real installed-app operation, bounded in scope.
- LM training
Recovery campaign: reviewed corpora editions 004 to 007 and recovery runs 002 to 004
Watched recovery training runs were performed against reviewed, frozen corpora. Package qualification is blocked whenever semantic review is missing. The audible reflex qualification was bound to the C013 voice.
Why it mattered. Every run was custody-tracked and none was promoted. Recovery004 trained and reloaded but was not development-qualified.
- LM training
Final architecture candidate trained; development gates fail
A fresh 14B base with q/v rank-16 adapters ran on corpus edition008 for 125 optimizer updates, and checkpoint 96 was selected by a pre-declared rule. Validation passed 8 of 26 cases, adversarial passed 5 of 25, and development scored 111 of 180.
Why it mattered. The best-engineered candidate still failed the unchanged gates. The sealed battery stayed unopened.
- LM training
Negative bounded-feasibility decision
Several bounded probes were run: more exposure, adding MLP down-projection capacity, a precision comparison, and an alternative 14B base with no adapter. None met the frozen behavioural viability requirements. The decision was not to start another full training run.
Why it mattered. It was a reviewed, evidence-bound 'no'. Production cognition stays hosted, and the decision explicitly does not claim every local model is incapable.
Era 07 / 22 Sept 2026
Advanced Systems
Spatial engineering workspace, CAD operations, frames and shadow roots, capability discovery.
9 dated milestones
- September 2026 , continued
- Advanced Systems / capability discovery
Advanced Systems begins: temporal world blackboard and bounded capability discovery
Six parallel lanes (MAIN/SYSTEMS, COGNITION, OPERATOR and three executors) branched from 893913a9c. The first commits added a temporal world blackboard, a lossless requirements census, versioned body reads and CapabilityDiscovery, which has no execution or grant method.
Why it mattered. JARVIS can now ask what it can do right now through a read-only body interface.
- Memory
Obsidian projection, project notebook, Failure Atlas
A governed Obsidian projection and project notebook were added, together with Failure Atlas diagnosis history and later cross-trajectory failure patterns. Separately, the COGNITION lane added an eight-layer memory hierarchy with atomic supersession.
Why it mattered. Memory became human-readable without a second authority, and past failures became searchable patterns that are never upgraded to proven causes.
- Spatial scene / engineering workspace
Revision-bound spatial scene with STEP, STL, OBJ and 3MF
A persistent, revision-bound scene store was added, with import of STL, OBJ and 3MF and display of STEP. Spin has an E-stop lease.
Why it mattered. The engineering workspace got a real 3D model surface. 'Load STEP designs into the scene' is the only COMPLETE row in the Advanced Systems matrix.
- Agents
Specialist checkpoints, contracts and a bounded local worker
Specialist checkpoints were added to MissionStore, along with readiness, conflict and freshness handling. Specialist contracts cover 11 roles, and a local specialist worker has no autonomous model dispatch.
Why it mattered. Missions can split work among roles, but 'scope is a ceiling, never execution authority'.
- CAD / fabrication
FreeCAD STEP operations and STEP-to-slicer pipeline
Governed FreeCAD STEP operations were added: booleans, split, components, transform, mirror, edges, fillet, and later plane split, offset, faces, face extension and sealed hollow. A STEP to verified-slicing-mesh path was added, with fabrication.slice_asset and slicer estimates feeding engineering runs.
Why it mattered. CAD edits became journaled calculations with receipts, and print preparation stays separate from physical printing.
- Engineering workspace
Recorded trade spaces and resumable sweeps
Engineering trade-space comparison was added, with constraints, archives, HUD plots and resumable scale sweeps.
Why it mattered. Design comparison became auditable. A nominal frontier is explicitly not feasibility or approval.
- Windows operation / browser
Operator lane: frames, shadow roots, UIA DPI geometry, a 545-test acceptance run
Browser DOM work added frame and shadow-root relationships, with refusal of stale targets that moved, were replaced, covered or changed. UIA geometry became DPI-correct: a disposable window was moved across four displays with physical-pixel readback, and E-stop refused a later move. The Operator lane passed 545 tests across 43 files.
Why it mattered. It hardened computer operation where it usually fails, on stale targets and scaled displays.
- Research / cognition
Reflective tool-use controller; experiment predictions bound to receipts
A lane-only controller checks whether tool use actually verified and revises plans from receipts. The Operator lane binds experiment predictions to observation receipts.
Why it mattered. It is a first step toward Predict-Before-Act. 'A dispatch that did not verify stays unverified.'
- Fabrication / qualification
Print readiness checks; measurement hold
A lane added READY / NOT READY / WHY print readiness across nine checks, plus a governed handoff that asks for approval and does not print. A primary-LM measurement hold began at 19:50Z, after which most commits are PENDING_PRIMARY_RELEASE (untested).
Why it mattered. 'READY IS SOFTWARE READINESS AND NOTHING MORE.' The day ends as a running report, not a completion declaration.
Honest edges
What the timeline does not show
- No release has been qualified and nothing is formally verified. The release phase was still open when this data was exported (22 September 2026).
- There is no evidence that JARVIS has ever started a physical print. The print path deliberately refuses the final send.
- Several planned engineering systems (simulation, design of experiments, a digital twin) exist only as briefs without phase commits.
- Owner-present acceptance on the real microphone, camera and hands is still owed for several journeys.
- The 2024 prototype’s source code was lost; its behaviour is reconstructed from compiled files and timestamps.
Era labels recorded on the milestones: Original concept, First voice, Early architecture, Cognition, HUD, Memory, Agents, Qualification, Local model work, Windows operation, Early architecture (production), Voice, Spatial scene, Engineering workspace, Fabrication, LM training, C013 voice, Body architecture, Advanced Systems, CAD, Research.