Shipped Feature Inventory

1. Writing workspace

• Sanctioned plain-text composition surface with themes; prompt ↔ citation-library tab switch in the assignment bar; all student tooling on one screen.
• Proctor-gated: the surface unlocks only while capture is live; idle sessions pause capture within 5 seconds and resume on the next keystroke.
• Session start/resume nudge: a dismissible reminder to close AI/chat windows before writing begins — false-flag prevention as a courtesy, not a trap.
• Past-due lockout, honesty-pledge + signature submission flow, and a per-assignment rubric with provenance-integrity weighting.

2. Provenance chain

• Every composition event — keystroke batches, pastes (with source classification), AI-collaborator prompts, citation placements/removals — becomes a SHA-256 running-hash chain entry.
• Progressive witness ledger: the client commits the running chain root to the server every few events during composition; a chain fabricated after the fact cannot reproduce the witnessed roots.
• Server-side verification at submission: independent full-chain replay, summary re-derivation, paragraph auto-tagging, and a certifiable verdict the client cannot assert for itself.
• Submissions carry a citation snapshot, so the record stays stable even if the student's library later changes.

3. Citation system

• Per-student persistent library — a simple personal EndNote that follows the student across assignments, grouped by where each source was used.
• Four input modes: URL (server fetch + meta-tag extraction: og:*, article:*, citation_*, dc.*), pasted source block (heuristic author/title/year/DOI/patent parse), manual entry (grouped picker), and RIS/BibTeX file import.
• 30-type source registry — books, journal articles, websites, newspapers, patents, theses, videos, podcasts, encyclopedias, government documents, and more — auto-detected from ~45 hostname patterns and content signals, each type with its own field spec and formatter. Deterministic: no AI call anywhere in the citation path.
• Four citation styles (MLA, APA, Chicago, Simple) rendered natively at read time; live preview while entering; per-student style preference.
• Place mode: click-to-place [N] markers at the caret; auto-renumber keeps labels in text order across place, move, and remove; searchable in-workspace library with project filters.
• Teacher-side context check: for each placed marker, a meaningful-word-overlap test between the surrounding ~300 characters and the citation's title/author — green "matches on …" or amber "worth reviewing" with the quoted snippet. Heuristic and advisory.

4. Proctoring stack

• One lifecycle: webcam (13+ only) + whole-screen capture start together on explicit user action; window/tab shares are rejected — whole monitor or nothing. Capture only while writing is active.
• Privacy pipeline: background blur at capture → AES-GCM encryption in memory under a non-extractable session key → SHA-256 hash-sealing into a tamper-evident ledger → delete-on-pass. Permanent record is a frame-free, time-localized event timeline.
• Deterministic pass (free, always on): absence-while-writing detection via person-coverage masking; gaze awareness via a two-score model (pupil score from face blendshapes + head score from the head matrix), thresholds relative to a per-session rest baseline learned in the opening seconds — camera geometry and natural head pose are calibrated out, which is what makes the same build Chromebook-ready.
• Escalation gate: the AI deep-pass runs only on the time windows the deterministic pass flagged, on frames decrypted from exactly those windows. Clean sessions cost zero inference. Deterministic flags are computed upstream of the AI call and survive any AI failure.
• Screen analysis, two stages: stage 1 reports four targets only (AI assistant visible, answer source visible, unrelated app foreground, writing space not visible) under strict privacy constraints; stage 2, only on an AI-window flag, judges whether the assistant was actually used (typed query + rendered response vs. idle tab) and whether the visible content is about this assignment (topical comparison against the prompt).
• Keystroke rhythm: cadence, timing variation, and correction-rate profile per session — the low-self-correction flag catches uncorrected superhuman typing.

5. Teacher dashboard

• Per-assignment roster with submission status; per-student view with integrity badge, chain summary, paragraph provenance tags, and pasted-content counts.
• References section with the per-citation context check; proctor timeline with flags localized to the second, written in neutral, advisory language — evidence for a human decision, never a verdict.
• Student messaging, checkpoint history, and proctor-seal verification.

6. Platform

• Next.js (App Router) + React, browser-native, Chromebook-first; marketing site and product ship as one deployment (landing at /, live demo at /demo).
• Provider-agnostic AI brain (Anthropic / OpenAI / Google adapters, all vision-capable); every AI task is defined by a constitutional intelligence.md, so the proctor reviewer, the depth judge, and the future writing collaborator are the same machinery with different constitutions. Non-Claude providers get the constitution filter prepended automatically.
• Dev store: JSON-file-backed with atomic mutation; SQLite migration queued for production. Demo reset script included.

7. Roadmap (honest gaps)

• Multi-monitor: detection of a second, uncaptured display with the flag report honestly saying so; managed-Chromebook policy can auto-grant per-screen capture.
• Stylometric backstop (Tier 3): three-model comparison of a submission against the student's own corpus of proctored work — the irreducible answer to human-cadence drip tools. Buildable now, meaningful once pilot corpora exist.
• Database migration, progressive full-text checkpointing (drafts currently survive integrity-intact but not content-intact through a mid-essay browser wipe), CSL-strict citation styles, server-side style preference, SSO/rostering.