Assessment integrity infrastructure Level 3 · dual-camera + attested agent

No leaked paper. No proxy. No exam hall.

TrustXD ExamOS binds a candidate's identity to their hardware, their hardware to their screen, and their screen to a question paper that did not exist until the moment they authenticated.

TRUST CHAIN · SELECT A LINK

Four states of one session. Each link is independently provable, and the paper only unlocks when the three before it hold.

0 question papers in existence before a candidate authenticates
2 synchronised camera angles, temporally cross-matched
3 permitted verdicts — Clear, Review required, Not assessable
±0.05 logit tolerance on the difficulty of every generated variant
Scroll EST. 2026 · BENGALURU
NEET UGJEE Main JEE AdvancedUPSC CSE IBPS POSSC CGL CUET UGGATE VITEEESRMJEEE METCAT State CETs

Exam names denote format compatibility only. No affiliation with, or endorsement by, any testing agency, board or university is implied.

01 — The premise Why exams break

Three failures. One root cause: the secret is kept too long, by too many.

Failure 01

The paper exists before the exam

A printed paper is a valuable secret held by presses, couriers, strongrooms and centre staff for weeks. A leak is not an anomaly in that design — it is the design working exactly as built.

We fix it Nothing to steal. Papers are assembled per candidate at T−0 inside a sealed pipeline, and the ciphertext is useless until identity clears.

Failure 02

Identity is checked once, at a desk

A photograph, a signature, a tired invigilator at 9 a.m. Proxy substitution survives because verification is treated as a moment rather than a continuously held state.

We fix it Two cameras, one subject, cross-matched frame by frame — and a sealed chip proving the frames came from real hardware.

Failure 03

Integrity is reported as a score

Remote proctoring returns a probability — "87% suspicious" — and a career turns on it. A candidate cannot cross-examine a number, and a court cannot audit one.

We fix it No risk scores. Timestamped, deterministic event records; three permitted verdicts; a named human on every flag.

02 — The model Chain of custody

Three locks in series. The paper is the last thing to open.

Each lock is independently verifiable and independently auditable. Breaking one does not weaken the others — it stops the session.

Lock 01 · Module A

Identity lock

Government ID parsed on-device, active liveness, and a second camera at 60°–120° whose head-motion vector must correlate with the first in real time.

CORR ≥ 0.94 Held
Lock 02 · Module B

Display lock

The server paints a 3-second token on the exam screen. The phone reads it and reports back over an independent channel. Arrive late, or point at a decoy, and the lapse is recorded.

WINDOW 2000 MS Held
Lock 03 · Module C

Paper lock

Encrypted item bank, opened only after locks one and two hold. The generator then re-parameterises every item for that candidate — while the IRT calibrator holds difficulty constant.

RELEASE AT T−0 Held
03 — The system Five modules

Everything below runs on the candidate's own two devices.

Hardware-attested biometric KYC and dual-angle spatial verification

Module A

A deepfake can hold a front-facing camera. It cannot hold two cameras, at an arbitrary angle to each other, and keep both stories straight while a person is asked to turn their head and show the back of their ear.

Primary · laptop webcam 1080p · 30fps
LIVE mesh 468 pts
Secondary · attested phone offset 78°
LIVE attested · Play Integrity

Drag to move the candidate's head. Both feeds are driven by the same physical motion.

TEMPORAL CORR 0.981
Single subject

Document parsing on device

Aadhaar, passport or driving licence OCR runs client-side; only the extracted field hashes and a match decision leave the machine.

Active liveness, not passive

Randomised prompts — rotate to 45°, show the left ear, follow the dot — scored against a 468-point facial mesh and depth symmetry.

Hardware root of trust

The phone app signs every media payload with Apple App Attest or Android Play Integrity. Emulators, rooted devices and virtual cameras fail the chain, not the model.

Angle enforcement

The rig must sit between 60° and 120° of the primary axis. Outside that window the session cannot start — geometry is a precondition, not a score.

Cryptographic screen-binding and anti-spoofing

Module B

Knowing who is in the room is only half the problem. The other half is proving that the screen the phone can see is the screen the exam is running on.

Primary display · exam client seq 1180
SESSION TOKEN · TTL 3s ····-····
Secondary camera · read-back Display bound
· · · ·
RTT 412 ms
Server event log · append-only CLOCK: NTP-DISCIPLINED

    Independent channel

    The read-back travels over the phone's own WebSocket, not the laptop's. A compromised exam machine cannot forge its own confirmation.

    Server clock is authoritative

    Validity is measured against the server's own NTP-disciplined clock, so client clock drift or replay buys an attacker nothing.

    Decoys collapse immediately

    A wall, a photograph or a second monitor produces a stale token or none at all. Either way the mismatch is written to the record within one cycle.

    Lapse ≠ accusation

    A lapse is an event, not a verdict. It routes the session to a human reviewer with the exact clip and timestamps attached.

    Just-in-time exam generation with psychometric parity

    Module C

    The paper is assembled after the candidate is proven, not before. Every candidate receives a different surface — and, by construction, the same test.

    01Sealed item bank
    02Identity gate L3
    03Variation engine
    04IRT calibrator
    05Sealed delivery
    Canonical item PHY-1104

    A cyclist of combined mass 60 kg is moving at 8 m s⁻¹. Brakes bring the cycle uniformly to rest over 4 m. Find the magnitude of the average retarding force.

    a 1.240b 0.380c 0.200 time 96 s
    Delivered to this candidate PHY-1104-V001

    a 1.242b 0.374 c 0.200time 96 s

    Names, places, numeric parameters, option order and phrasing all move. The difficulty, the cognitive operation and the expected solve time do not.

    Zero-leakage delivery gate

    Papers stay encrypted server-side. The decryption key is derived only after Level 3 identity and screen-binding both hold, and is never persisted on the client.

    Late arrivals are covered

    A candidate who starts twenty minutes late gets a freshly generated paper, not a leftover one. There is no window in which an earlier paper helps a later candidate.

    Generation is bounded

    The model rewrites within a template's declared parameter ranges and answer-preserving transforms. It does not invent physics, and every variant is machine-solved before release.

    Parity is enforced, not hoped for

    Variants outside the ±0.05 logit band are rejected by the calibrator and never reach a candidate. See the curves ↓

    Deterministic evidence and human-in-the-loop review

    Module D

    No integrity scores. No automated rejections. A machine may only observe and record; a trained human decides, and the candidate can read the same record the reviewer read.

    Review queue · today 3 OF 214,806 SESSIONS
    No integrity score exists in this system. There is no field to store one, no API that returns one, and no configuration that enables one.
    Session S-77401 · Level 3 Review required

      The candidate receives this same trail, plus a log of every reviewer who opened it.

      Exactly three verdicts

      Clear, Review required, Not assessable. There is no fourth outcome, and "Not assessable" always carries a remedy — a free re-sit, never a penalty.

      Reviewers are independent

      Reviewers are not employed by the institution whose exam they are reviewing, are randomly assigned, and see the evidence before they see the candidate's name.

      Every record is reproducible

      Given the same inputs the engine emits byte-identical events. A tribunal can replay a session and get the same trail three years later.

      Appeals are first-class

      A candidate can open a dispute from inside their own report. The clock on evidence retention pauses for the duration of the appeal.

      Candidate rights, consent and data minimisation

      Module E

      A surveillance system that a candidate cannot see, understand or contest is not a trust system. Everything captured is visible to the person it was captured from.

      Continuous self-view

      The candidate sees their own secondary stream at full size for the whole session, so what is captured is never a surprise.

      Bilingual by default

      Consent, prompts and the final report ship in English and Hindi, with the language pack extensible to any scheduled language.

      Who read your file

      Every access to a candidate's evidence — reviewer, auditor, agency — appears in the candidate's own transparency log.

      Purge is automatic

      Unflagged media is discarded when the session closes. Flagged clips expire at 30 days unless an appeal holds them, and the purge itself is logged.

      04 — Deployment Three levels of assurance

      Match the stakes. Nothing heavier, nothing lighter.

      An institution declares the level; the candidate sees the level before they consent; the certificate records the level it was earned under.

      Level 1 Low stakes

      Browser & webcam

      • Single-camera identity check at start
      • Static item bank with option shuffling
      • Event log, no continuous media retention

      Formative assessment, practice tests, internal coursework.

      Level 2 Medium stakes

      Dual camera, browser + app

      • Dual-angle spatial verification
      • Screen-binding token loop
      • JIT variant generation with IRT parity
      • Human review on every flag

      University entrance tests, hiring assessments, structured interviews.

      Level 3 National scale

      Dual camera + attested status agent

      • Everything in Level 2, plus
      • Sealed-chip attestation on every media payload
      • Lightweight system status agent — display topology, virtual-camera and remote-desktop detection
      • Independent reviewer pool with agency observer access

      NEET, JEE, UPSC, banking recruitment and state common entrance tests.

      The status agent reports state, not content. It has no keylogger, reads no files, and inspects no other application's memory — it answers a fixed set of questions about displays, capture devices and remote sessions, and it uninstalls itself when the exam ends.

      05 — Exam day Session runtime

      Forty-five minutes of proving. Then the paper appears.

      Session clock

      T−45:00

      Device readiness

      Every step is reversible up to T−0. After the paper releases, the session may pause but never restarts.

      T−45:00

      Device readiness

      Bandwidth, camera resolution, battery headroom and phone-mount geometry are checked. Failures here are cheap; failures at T−2 are not.

      ≥ 1.5 Mbps upBattery ≥ 40%Rig 60°–120°
      T−30:00

      Itemised consent

      Each capture right is granted separately, in the candidate's chosen language, and a signed receipt is issued to the candidate before anything is recorded.

      EN / HIEd25519 receipt
      T−22:00

      Document and liveness

      ID is parsed on device and matched against a live 3D facial mesh under randomised movement prompts. The ID number never leaves the machine in clear text.

      On-device OCR468-pt meshActive prompts
      T−15:00

      Dual-angle spatial lock

      The phone stream joins. Head-motion vectors from both cameras are correlated over a rolling window; the session cannot proceed below the correlation floor.

      Corr ≥ 0.94Rolling 8 s
      T−08:00

      Hardware attestation

      App Attest / Play Integrity receipts are verified against the vendor chain. Emulators, virtual cameras and rooted devices are refused before any exam data moves.

      Sealed chipChain verified
      T−02:00

      Screen binding established

      The token loop starts and must hold three consecutive cycles inside the latency window before the delivery gate will even evaluate.

      3 s tokens2000 ms window
      T−00:00

      Paper release

      Only now does a paper exist for this candidate. It is generated, IRT-checked, encrypted to this session's key and streamed item by item.

      Delivery gatePer-candidate key
      T+00:00

      Continuous binding

      Identity correlation, screen binding and attestation all remain live for the full duration. Interruptions pause the timer and are credited back, second for second.

      Timer pausesTime credited
      T+180:00

      Sealed submission

      Responses are hashed, countersigned by server and client, and written to an append-only ledger. The candidate receives the hash before they close the tab.

      SHA-256Counter-signed
      T+180:04

      Evidence disposition

      Unflagged media is destroyed immediately. Flagged sessions enter the human queue with a 30-day clock, and the candidate is told, in their own report, exactly which is which.

      Purge or queueP30D
      06 — Reality Edge-case protocol

      The exam is at home. Home is not a controlled environment.

      Every one of these has a defined, candidate-favourable protocol. None of them ends a career by accident.

      E01 The internet drops mid-paper

      Answers are already sealed locally. The timer halts within 400 ms, the client retries for 15 minutes, and on reconnect the lost seconds are credited back in full. The gap is an event on the record, never a verdict.

      E02 The phone rings

      The exam app holds the camera session through an incoming call and the timer pauses. A call answered is logged as a stream interruption and routed to human review — not scored, not failed.

      E03 The phone battery dies

      Charging is mandatory at Level 3 and warned at 30%. On loss, the candidate has 10 minutes to restore the secondary angle; beyond that the session becomes Not assessable with an automatic free re-sit.

      E04 The mount slips or the phone falls

      Geometry loss is detected within two frames. The candidate is prompted to re-seat the device, and the reposition window is timed out of the exam, not out of the candidate.

      E05 Mains power fails

      The client runs from battery and degrades to audio-plus-low-frame-rate video at 15% charge. Region-wide outages are detected across sessions and escalate to an agency-level reschedule.

      E06 Someone walks into the room

      A second face is recorded as an environmental event with the clip attached. A family member crossing the frame and a second person seated at the desk are different facts, and a human decides which happened.

      E07 A candidate needs a comfort break

      Scheduled and medical breaks are declared in advance and honoured. Re-entry requires a fresh liveness check; break time is excluded from the exam clock by policy.

      E08 The candidate uses assistive technology

      Screen readers, magnification and scribes are first-class configurations, declared at registration. The status agent allowlists them; extra time is applied to the generated paper's timing model.

      E09 The candidate has one device, not two

      Level 3 requires two. Where a second device is unavailable, the agency assigns a supervised civic node — a library or panchayat centre with loan hardware — so the level is never a wealth test.

      Full protocols, owners and timing tables for all twenty-six defined edge cases are in the Product Requirements Document →

      07 — Fairness Psychometric parity

      Different papers are only fair if they are the same test.

      Item Response Theory turns that from a promise into a measurement. Every generated variant carries the same three parameters as its parent — discrimination, difficulty and guessing — inside a tolerance the calibrator enforces before release.

      Item characteristic curves — one item, eight variants

      3PL model: P(θ) = c + (1 − c) / (1 + e−1.7a(θ − b)). The shaded envelope is the ±0.05 logit parity band; a variant that leaves it is rejected before delivery.

      PHY-1104
      Base item, calibrated (dotted) AI-generated variants, n = 8 (solid) ±0.05 logit parity envelope
      View as table

      Difficulty parity across a delivered cohort

      Twelve variants of PHY-1104 as delivered on 30 July. Every b-parameter sits inside the tolerance band.

      Base item b = 0.380 Delivered variant Tolerance ±0.05

      What the calibrator holds constant

      • a — SLOPE Discrimination. How sharply the item separates candidates near its difficulty. Held to ±0.01.
      • b — LOCATION Difficulty on the ability scale. Held to ±0.05 logits — the visible band above.
      • c — ASYMPTOTE Guessing floor. Fixed by option count and distractor plausibility, so variants may not change how many options there are.
      • t̄ — TIME Expected solve time from the response-time model. Held to ±4 seconds, so no variant is quietly longer to read.

      Pre-calibration on a live cohort, continuous re-estimation during delivery, and a post-hoc DIF sweep across language, region and gender. Read the equivalence specification →

      08 — Operations Agency control room

      One national exam. One screen.

      Delivery health, queue depth and disposition — live. Every number here is a count of recorded events, never an estimate of anyone's honesty.

      trustxd://control/neet-ug-2026/session-day-01 LIVE

      Candidates in session now

      214,806

      across 28 states · peak concurrency 231,412

      Sessions started per minute

      1,284

      last 24 min · +18% vs. opening window

      Delivery gate — median time to paper

      41.2 s

      from Level 3 clear to first item rendered

      Identity 24.1 s Attest 6.4 s Bind 5.9 s Gen 4.8 s

      Session disposition · today

      Clear 97.1% Review 2.6% Not assessable 0.3%

      Illustrative operating figures for a Level 3 delivery day, shown to demonstrate the interface. Not results from a live administration.

      Event stream · human review queue

      MEDIAN TIME TO REVIEWER 6 MIN
      • S-77401secondary_stream.interrupt 11.4 sReview
      • S-77399identity.level3 corr 0.983Clear
      • S-77396submission.sealed 9b21…77c0Clear
      09 — Governance DPDP · GDPR

      Capture the minimum. Keep it briefly. Prove you deleted it.

      STAGE 01 · CAPTURE

      Purpose-bound

      Media is captured only for declared exam-integrity purposes, only for consented items, and only between session start and submission.

      STAGE 02 · MINIMISE

      Discard on close

      Unflagged video and audio are destroyed the moment the session seals. What survives is the event record — timestamps and outcomes, not footage.

      STAGE 03 · RETAIN

      30 days, encrypted

      Flagged clips only, encrypted at rest with per-session keys, accessible to assigned reviewers and to the candidate. Every access is logged to both.

      STAGE 04 · PURGE

      Automatic and evidenced

      Deletion runs on schedule, emits its own signed record, and pauses only for an open appeal. The candidate is notified when their file is gone.

      Residency

      All candidate data processed and stored in-country, on infrastructure the agency can inspect.

      Rights

      Access, correction, portability and grievance redressal wired into the candidate's own report view.

      Audit

      Append-only ledger of every verdict, access and purge, exportable for statutory audit.

      Independence

      Reviewer pool contracted separately from the delivery institution, with rotation and blind assignment.

      10 — Voices From design partners

      What the people who run exams told us.

      Placeholder copy — replace with attributed quotes before launch

      "The part that changed the conversation internally was 'the paper does not exist yet'. Everything we spend on strongroom logistics is protecting an asset this design simply does not create."

      CEController of ExaminationsState technical university · design partner

      "We have refused every proctoring pilot on one ground: none of them could tell a tribunal why a candidate was flagged. A deterministic trail with three verdicts is something our legal cell can actually defend."

      LCHead of Legal & CompliancePublic-sector recruitment board

      "My question was always psychometric, not technical. Show me that variant 47 is the same difficulty as variant 3 and I will sign. The parity band answered it in one chart."

      PSPsychometrics leadNational assessment research group

      "Students in my district travel eleven hours to a centre. If the integrity story holds, the equity story is bigger than the integrity story."

      DEDistrict education officerAspirational district programme

      "The self-view mattered more than we expected in user testing. Candidates stopped asking what was being recorded once they could simply see it, at full size, the whole time."

      UXAccessibility researcherCandidate experience study, n = 240
      Pilot programme · 2026

      Run one paper on it. Then decide.

      A pilot is a single exam, one cohort, both systems side by side — your existing delivery and TrustXD ExamOS — with the evidence ledger open to your auditors from day one.

      We reply within two working days. No candidate data is ever requested for an evaluation.