Skip to content

PROGRESS.md

Latest commit

 

History

History
118 lines (87 loc) · 8.82 KB

PROGRESS.md

File metadata and controls

118 lines (87 loc) · 8.82 KB
Raw

ULTA — PROGRESS

Started: : · Demo at: : Demo misconception: M-FRIC-04 — friction always opposes velocity

Updated only at sync points, by whoever merges last. One writer per sync means this file never conflicts. Individual work goes in progress/deepthi.md, progress/jeswin.md, progress/preethesh.md — append there after every commit.

Status

Deepthi · feat/agents Jeswin · feat/core Preethesh · feat/web
Now all four agents + 3 misconception files merged; seam to feat/core verified live server skeleton merged (routes, fixtures, orchestrator, voice wrappers) fully live browser session verified: vision → debate → judge → transfer → defeat
Last merge ab56f7e 1850685 (this merge)
Blocked on — — hardware mic check

UI planning note

The UI direction now takes inspiration from my AI Tutor: warm, approachable, rounded, spacious, and centred on one clear action at a time. For ULTA, that means a debate-first session screen where Chintu, the belief meter, captions, and stage rail remain dominant. The existing Chintu.jsx is the avatar reference; wire its emotion states, gestures, belief-threshold surprise flash, Maya lip-sync, and reduced-motion fallback without altering the component.

Post-hackathon senior re-audit — Codex

This section supersedes stale implementation assumptions in the historical log below. The original three-misconception library remains only for deterministic fixture/demo mode; the live product now derives a validated concept package from the student's actual uploaded work and never silently substitutes M-FRIC-04.

  • Deepthi / agents: Diagnosis accepts any academic subject from an uploaded page, spoken/typed reasoning, or both; rejects prompt injection; and runs a second independent evidence-grounded Sonnet audit before creating a session. Correct, vague, blurry, unrelated, malformed, and unsupported input returns UNKNOWN without a session. Chintu, Judge, and Verifier consume the validated dynamic concept package; all Claude agents share one Sonnet-only model configuration. Deterministic agent suite: 61 checks. Live regressions covered text-only algebra plus friction, algebra, chemistry, a correct solution, blur, unrelated UI, and image-borne prompt injection. Expanded live Chintu adversarial suite: 9/9 held, while a complete explanation still made him yield.
  • Jeswin / core: Exact session IDs and legal stages are enforced; turns are deduplicated; arbitrary post-pass Chintu turns are rejected; Judge and provider failures fail closed; verification is idempotent; uploads and JSON bodies are bounded; API errors stay JSON; sessions expire after inactivity. Diagnosis accepts any non-empty combination of image and learner explanation. Sarvam STT defaults to live-tested saaras:v3, Maya uses Arjun, and both voice providers retain quiet browser-safe fallbacks. The 22-check core suite covers state transitions, cross-session isolation, text-only start, TTL cleanup, voice validation, providerless behavior, and dynamic fallback content.
  • Preethesh / web: The shipped default calls the real server. Entry is genuinely multimodal: voice or typed reasoning can start without a photo, while a page optionally sharpens the diagnosis. Failure is honest and retryable, with no bundled misconception substitution. Voice is click once to record and click again to send, with typed fallback, stale-operation guards, stream cleanup, and exactly one submitted turn. The UI and full-body animated Chintu were redesigned around the agreed warm tutor reference while retaining captions, lip sync, reduced motion, and responsive layouts. Eight Playwright runs cover desktop and mobile entry, no-photo start, voice lifecycle, and the complete debate-to-transfer flow.
  • Live end-to-end evidence: A dynamic algebra upload completed Vision diagnosis → Chintu debate → Judge → generated transfer problem → defeat screen in the browser, with Maya audio returned for Chintu turns. A forced Judge network failure recovered in place without advancing the stage or locking the mic. Providerless and oversized-upload paths were also exercised.

Component state

Component Owner Status Notes
M-FRIC-04 misconception file D ✅ full CONTRACTS §4 schema
M-NEWT-03, M-NEWT-07 D ✅ library capped at three per anti-goals
Diagnosis agent D ✅ vision verified live both ways: wrong work → M-FRIC-04 0.95, correct work → UNKNOWN; real phone photo still worth one check
Chintu agent D ✅ adversarial run live: 6/6 attacks held after prompt fix
Judge agent D ✅ live: fails keyword-only, passes real mechanism (0.95→0.2)
Verifier agent D ✅ live: fresh transfer problem in a different context
Server + 6 routes J ✅ verified live with USE_FIXTURES=false calling real agents
Fixtures J ✅ full-session arc works
Orchestrator state machine J ✅
Sarvam STT J ✅ live: real transcription round-trip through /api/stt
Maya TTS J ✅ live: 8.5s Chintu audio via /api/tts (Arjun, emotion tags, PCM→WAV)
App shell + session screen P ✅ §36 anatomy, browser-verified
Chintu avatar wired P ✅ imported as-is, isolation kept in props
Belief meter P ✅ 400ms lag, red→amber→green, Judge-driven
Captions P ✅ always render; audio is enhancement
Defeat screen P ✅ bars old→new, SPOT last, static dashboard slide
Full loop end to end all ✅ live via API AND through the browser UI (headless): photo → diagnose → debate → judge pass → transfer verified → defeat screen. Demo laptop run still pending

⬜ not started · 🟡 in progress · ✅ done · ⛔ blocked · ✂️ cut

Sync log

SYNC 1 — T+1:30

  • Merged: feat/core (1850685, Jeswin first per plan), feat/web (0f050b3), feat/agents (4135711, rebased)
  • Working end to end: fixture path clicks through; USE_FIXTURES=false verified live — routes load the real agents, agents degrade in character without a key
  • Blocked: ANTHROPIC_API_KEY (agents), Sarvam/Maya keys (voice)
  • Decided: agent module interface is ESM with session-based exports chintu/judge/verify matching routes.js; transfer answers judged against transferProblem.expected_reasoning
  • Cut: nothing

SYNC 2 — T+2:45

Decision point: is a real session running end to end? If no, cut the Verifier and Transfer stage now.

  • Merged: feat/agents (ab56f7e), feat/web (judge-first sequencing fix)
  • Verdict: YES — fully live session through the web UI: real vision diagnosis of the sample photo, live Chintu debate, judge pass, live transfer problem, transfer verified, defeat screen. Frontend defect found and fixed at this seam: parallel chintu+judge calls interleaved on the stateful server session and corrupted the Judge's view of history; turn calls are now sequential, judge first.
  • Cut: nothing — Verifier and Transfer stay.

SYNC 3 — T+3:45 — FEATURE FREEZE

  • Merged:
  • Known broken, accepted:
  • Demo laptop tested: ⬜
  • Backup recording made: ⬜

Decisions

Anything that changes how the system behaves. Reason is mandatory — a decision without one gets relitigated at hour three.

T+ Who Decision Why
0:20 all Contracts frozen Three people can't build in parallel against a moving interface
0:20 all M-FRIC-04 is the demo One flawless session beats five shaky ones
1:40 P Mock layer lives inside web/ behind VITE_USE_MOCKS server/fixtures don't exist yet and web can't create them; flag flip swaps to real routes with zero component changes
1:40 P Belief meter takes the Judge's belief_strength, not Chintu's the meter only moves when the Judge confirms the model shifted
2:00 D Judge keyword rule enforced in prompt AND a code gate (pass with empty repair_evidence or <8 words → fail) prompts drift under pressure; the demo-critical behaviour must be deterministic
2:40 D Agents are ESM with route-facing adapters; chintu() adapter is the isolation boundary session handed by routes contains correct_model; extraction must be field-by-field where the routes call in

Contract changes

CONTRACTS.md changes only by agreement at a sync point. Every change logged here.

T+ Who Changed Agreed by

Demo runbook

Fill this in at Sync 3, not before. Read it out loud once before presenting.

  1. Laptop, browser, tabs open:
  2. USE_FIXTURES set to:
  3. The handwritten sheet used:
  4. Language selected:
  5. The line that opens the pitch:
  6. If it breaks: USE_FIXTURES=true, reload. If that breaks: play the recording.

Who does what: Deepthi pitches · Jeswin drives · Preethesh watches for breakage with the fixture flag ready.