The phrase "cheat-proof" creates a false certainty that no assessment instrument can satisfy. What we actually mean when we call a scoring approach integrity-conscious is that it makes certain categories of response contamination detectable at the scoring model level — not invisible to the system. The Response Profile Integrity Score (RPIS) that LearnVyx computes is a probabilistic signal, not a conviction. This distinction matters enormously for how institutions should use it.
What RPIS measures
RPIS is a session-level score ranging from 0 to 1. It is computed by combining four behavioral signal streams through a calibrated Bayesian network:
- Response latency distribution: Item-level time-to-answer compared against the expected distribution for that item's difficulty and the learner's current theta estimate. A very fast correct answer on a high-difficulty item (b = 1.8) is a statistically improbable event under genuine ability-based performance. Not impossible — some learners are fast and able — but improbable at the rate it appears in contaminated response profiles.
- Answer confidence correlation: For items where learners self-report confidence, the calibration between stated confidence and actual accuracy. A learner who systematically reports high confidence and is correct on items above their estimated ability level is exhibiting a pattern that differs from both genuine high-ability performance (where calibration is consistent) and random guessing (where calibration is near zero).
- Item revision direction: Whether answer changes go from correct-to-wrong or wrong-to-correct, and on which items. External assistance tends to shift the distribution of revisions toward wrong-to-correct patterns on high-difficulty items — a second-order signal that compound with latency data.
- Person-fit statistics: Infit and outfit mean-square values computed against the 3PL model's expected response pattern. An infit mean-square above 1.5 or below 0.5 indicates that the pattern of correct and incorrect responses is inconsistent with any single ability level — a signature of non-ability-based responding.
These four streams are weighted and combined. The resulting RPIS score is not a simple average — the weights are item-specific and calibrated from a training set of known-integrity response profiles. An RPIS above 0.85 is consistent with genuine ability-based performance. An RPIS below 0.65 indicates a statistically improbable session profile that warrants human review.
What RPIS does not tell you
RPIS does not tell you that a learner cheated. It tells you that their response profile is statistically inconsistent with the expected distribution for genuine ability-based performance. These are different claims. Three scenarios produce low RPIS that have nothing to do with cheating:
- Technical interference: A learner with slow internet receives latency artifacts that inflate or deflate their response time measurements. A 200ms network lag added to every item response can shift the latency signal in ways that mimic rapid-answer patterns at some difficulty levels.
- Anxiety-induced responding: Test-anxious learners sometimes change answers repeatedly (inflating revision signals) or freeze on easy items (inflating latency signals for low-difficulty content). This pattern can depress RPIS without any external assistance.
- Genuine expertise with rapid retrieval: An expert in the domain being assessed may legitimately respond to high-difficulty items quickly — because the items, relative to their ability, are not actually difficult. If the item bank was calibrated on a lower-ability population, the b-parameter estimates may not accurately reflect the item's difficulty for this examinee, generating spurious latency flags.
This is why LearnVyx's integrity architecture is designed to surface cases for human review, not automate penalty decisions. An RPIS score is an input to a human process, not a substitute for one. We've been explicit about this in our product documentation since launch. Institutions that want to use RPIS as a trigger for consequence (grade withheld, retest required) need a human review step between the signal and the outcome — and that review process needs to be documented in academic integrity policy before deployment, not after the first disputed case.
The right use cases
RPIS is most defensible in three contexts. First, as a population-level audit tool: if a specific item bank or testing administration is generating unusually low RPIS scores at scale (say, 15% of sessions below 0.65 when typical rates are 3–4%), that is a signal that something systematic is happening — possibly item exposure, possibly organized external assistance, possibly an administration environment problem. The signal is most interpretable at aggregate levels.
Second, as a flag for instructor attention in contexts where consequence is proportional to evidence quality. A session with RPIS = 0.48 in a low-stakes formative assessment is worth noting; it doesn't justify an academic integrity referral. A session with RPIS = 0.41 alongside an outlier latency profile and multiple answer reversals on hard items in a high-stakes certification exam is worth a structured review conversation. The context of the decision scales the appropriate response to the signal.
Third, as evidence of measurement quality for credentialing. Every LearnVyx credential includes the session's RPIS in its metadata. An employer or accreditor who questions the validity of a particular credential can verify not just that the score meets the threshold, but that the response profile was consistent with genuine performance. This is a form of audit transparency that static test scores don't provide.
Spectrum, not binary
The framing that matters is that integrity is a spectrum of evidence quality, not a pass/fail gate. A session with RPIS = 0.92 is not "definitely clean" — it is consistent with clean performance. A session with RPIS = 0.51 is not "definitely contaminated" — it is statistically unusual enough to require attention. The honest institutional posture is to treat RPIS as one signal in a multi-signal integrity picture, alongside instructor judgment, item bank security practices, and proctoring conditions where those are relevant and proportionate.
What we reject is the pretense that any technical system can resolve the integrity question by itself. Camera-based proctoring claims to see everything and misses plenty; RPIS claims to detect improbable patterns and acknowledges it cannot observe the test environment. Neither is a complete solution. The complete solution is a combination of good assessment design (adaptive difficulty that limits strategic guessing), model-level integrity analysis (RPIS), defensible scoring transparency, and a human review process for edge cases. LearnVyx handles the first three. The fourth is the institution's responsibility, and we're explicit about that boundary in every implementation conversation we have.
