SPHERE Virtual Lab: Operate a Metrology Lab in Your Browser

Changed

• The realistic 3D lab (Three.js) and its audio and technician-voice code now download only after the 2D operator dashboard has painted and the browser is idle, so the dashboard is no longer competing with the 3D assets for the initial load.
• Academic release metadata (Zenodo and ORCID) is included in package.json and CITATION.cff.

Added

• Dr. Wire's everyday conversation is about ten times larger: 134 topics, nearly 400 differently worded answers and over 600 example questions, all answered offline and in character.
• He keeps a conversation going: he sometimes asks something back ("What about you?"), reacts to your answer (dogs, kids, coffee, sports), and then steers gently back to the lab.
• "When do you get off work" now has time-aware, natural answers (nearly done, plenty of day left, already ended).
• One consistent character (personaBible.ts) shared by the offline answers and the AI persona, so improvised answers cannot contradict the scripted ones.

Changed

• Everything he says now passes through a fluency layer: contractions, "Cause / Effect / Check / Fix" labels turned into spoken sentences, and a lead-in or closing question on longer explanations.
• The AI persona prompt now asks for connected, flowing speech instead of lists of facts.

Fixed

• High quality is no longer washed out: the bloom pass that smeared every bright metal highlight into a white haze is removed, the exposure is lower, the chamber lamp and ceiling light panels are dimmer so the polished chamber liner no longer flares, and the walls and floor are a little darker so there is contrast to see.

Added

• A right-hand wall and many more lab objects: fume hood with a sash and exhaust duct, a bench with a sink, tap, towel dispenser and reagent shelf, an eyewash and drench shower, a tall chemical storage cabinet, a rolling tool chest, waste and recycling bins, stools, a gas cylinder rack, a coat rack with a spare lab coat, a wet-floor sign, an anti-fatigue mat, overhead ductwork, a cable tray, sprinkler heads, exit signs, a first-aid kit, a fire alarm pull station, an extinguisher and a hazards poster.
• Dr. Wire looks more lifelike: skin with pores and colour variation, flushed cheeks and a five o'clock shadow, sheen on skin and hair, strand-textured hair with a fringe and fuller back, limbal rings and catchlights in the eyes, nostrils, shaped lips, fingernails, woven-cloth bump on the lab coat, shirt and trousers, shirt buttons and a belt buckle.

Added

• Ask Dr. Wire for any lab action on the phone, by voice or typing.
• Wall clock that shows Eastern time (America/New_York, EST or EDT as it applies) whatever the visitor's own time zone, and a wall calendar that shows the current Eastern date and redraws when the date changes.
• Realistic lab audio: layered footsteps (heel thud, contact click, sole scuff, occasional rubber squeak, alternating feet, per-step variation) with room reverb and stereo position and loudness that follow the camera; an ambient bed with HVAC, light hum, rack fans, a roughing pump and a turbo whine that follow the real instrument state; a quiet wall-clock tick; better latch and relay sounds.

Changed

• The call panel is much smaller (about 264 px wide, 11 px text), translucent, shows the last three lines, and keeps the audio and microphone controls behind an Options button.
• The two computer monitors are larger again (0.96 m by 0.56 m, up from 0.74 m by 0.43 m).

Added

• Natural "studio" voice for Dr. Wire: a neural text-to-speech service (/api/tts, ElevenLabs or OpenAI) plays instead of the browser voice when a key is set on the server (ELEVENLABS_API_KEY or OPENAI_API_KEY).
• The studio endpoint has the same guards as the other endpoints: same-origin only, a per-client rate limit, a daily spend cap (TECHNICIAN_TTS_DAILY_CAP_USD, default 2 USD), short bodies, and text and audio are never logged.
• Performance test (tests/e2e/perf.e2e.py) that checks the draw-call saving and that the merged scene looks identical, including the moving parts.

Changed

• Much smoother 3D lab: the static parts of the room are fused into a few meshes per material (about 830 down to about 290 draw calls per frame, half the triangles), the shadow map refreshes every second or third frame instead of every frame, the bloom runs at half resolution, SMAA is skipped on high-DPI screens, and phones and tablets skip 4x MSAA.
• The speech bubble no longer forces a page layout on every frame.

Added

• Dr. Wire asks who he is speaking with when he picks up the phone and uses the first name now and then (not in every sentence).
• Many different ways of not knowing: ten "I don't know" phrasings, eight "let me look it up" lead-ins, "I'm not sure, but..." hedges for weak matches, and the language model is told to vary its wording too.
• Search almost any topic: Brave Search (if BRAVE_SEARCH_API_KEY is set) or Claude web search, then Claude's general knowledge when search is unavailable (clearly labelled "not a live search"), then Wikipedia straight from the browser when the service is not set up.
• A more natural voice: lab abbreviations and units are pronounced properly (V N A, kilohertz, 1e-3), speech is cut into phrases with their own pace, pitch and pauses (questions rise, "Hmm," is slower), the best male English voice the browser offers is chosen automatically, and a voice picker appears in the call bar.
• A realistic Dr. Wire: a solid, shaded figure with proportioned head, face, glasses, hair, a blue shirt, trousers, shoes and an open white lab coat.
• Screensaver on his left monitor after 90 seconds of no use (a turning copper sphere and a 230 kHz trace).
• Three security cameras mounted in the lab (blinking red lights) and a Cameras control: Cam 1 (overview), Cam 2 (workstation), Cam 3 (chamber and rack) as fixed views with a time stamp and REC light, plus an All cams grid.

Changed

• The greeting is "Lab, Dr. Wire speaking.