Credentials That Carry Their Evidence: Credential design, dynamic tiering, and approved use when AI agents can do the work

Synthesis: Srivastava (2026) argues that when an AI agent can complete an online course on a student's behalf, a credential that says "earned" loses its meaning, and that the standard remedy — proving a human did the work through identity checks and proctoring — is fragile and targets the wrong quantity. The paper proposes that credentials retain value only if they carry evidence of the reasoning that produced them, captured while learning happens and inspectable by someone other than the issuing institution. It advances three design elements: the credential as a claim plus a verifiable receipt (an inspectable, independently verifiable record of what was observed, hashed against alteration); dynamic tiering, in which the same skill is certified as unaided, directed, or attested and can be upgraded or lapse as evidence accrues; and declarations of approved use, under which agent use is permitted and measured rather than prohibited and policed. The organizing principle is that a credential's value comes from what can be inspected, not from who signed it.

The paper opens from reporting that students in 2026 used AI agents to log into learning platforms and complete entire online courses in the background. It reframes the reaction from a cheating problem to a signaling problem: a degree is valuable partly because it signals that a person spent time acquiring hard skills (Spence's signaling), and when an agent can produce the same credential without the human learning, the credential no longer separates those who learned from those who did not — an instance of Akerlof's "lemons" condition in which quality uncertainty drives good products out of the market. AI-mediated education makes this acute at the leading edge of online assessment, but the mechanism is general.

Why identity verification cannot restore the signal

The natural response is to prove a human did the work: identity checks, proctoring, agent watermarks, verifiable agent identities. Some are useful; none restore the signal. First, they are contests the verifier tends to lose over time — a student who wants to pass identity checks while an agent works can arrange to be present. Second and more fundamentally, identity is the wrong quantity: knowing a particular human was at the keyboard tells an employer nothing about whether that human can do what the credential names. The value of the credential was never that a person was present, but that a person learned. Framed through Kane's validity argument, the credential rests on a chain of inferences from observed performance to score to domain to capability to the employer's decision; agents break the first link, so repair must target the observation itself rather than the paperwork around it.

Three properties make an observation hold up when agents are available: it is contemporaneous (captured while learning happens — predictions made, tests run, explanations given, revisions after being wrong — not inferred afterward from a finished artifact an agent could produce); inspectable (every claim traces to specific evidence, and the rule turning evidence into an estimate is published with reference cases so a verifier can reproduce it); and independently verifiable (the checking party is not the issuer nor the builder of the learning environment, mirroring the auditor/issuer separation that makes audited financial statements credible). None requires agents to be absent; they require the human's reasoning to be observed and the observation to be checkable.

A claim plus a receipt

Today a credential is a claim: "this person earned this." The proposal is that a credential become a claim plus a receipt — a structured record of what was observed, when, and under what conditions and permissions, with the evidence chain hashed so it cannot be altered after the fact, plus a verifier any third party can run. Evidence and any estimates built on it are kept as separate chains, so a learning environment can produce evidence without producing a score, and different parties can layer their own estimates on the same evidence. The credential stays short and readable; the receipt gives it weight, so an employer, accreditor, or licensing board wanting more than the summary can ask for and check the record.

The analogy is a pilot's logbook rather than a pilot's license: the license says "may fly," but the logbook records hours, aircraft, conditions, maneuvers, and sign-offs, and it is the logbook an insurer or airline actually reads. Education has issued licenses without logbooks; agents make the logbook necessary. Design consequences follow: the evidence schema must be open so receipts from different institutions and environments are readable by the same verifier; the verifier must be open too; the learner must control disclosure (the receipt exists to be presented, not harvested); and the receipt should record conditions including which tools were permitted, because a claim without its conditions cannot be interpreted.

Dynamic tiering

A pass/fail credential forces one answer to a question that now has several honest ones. The proposal is that the same skill be certifiable at three strengths stated on the credential. Unaided means reasoning evidence was captured with no external tool or agent permitted and is inspectable — the strongest claim and most expensive to produce. Directed means agents or tools were permitted and the evidence shows the human directing them (choosing the approach, checking and correcting output, explaining why); a learner who drives an agent well leaves a different trace from one who is driven. Attested means the work was completed with no reasoning evidence captured — what most online credentials are today; naming it honestly lets the market price it rather than discounting everything.

The tiering is dynamic in two directions. A credential can be upgraded: a learner holding an attested credential can sit an unaided or directed session later and add stronger evidence to the same record. And a credential can lapse or be flagged for renewal: skills not exercised decay, and a receipt recording when evidence was last observed lets an employer weigh recency rather than letting the credential pretend a years-old demonstration is current. This mirrors how professions already handle competence through currency requirements and recertification. Employers need only three words on the credential and the option to look further; institutions need only say which tier they issue.

Approved use

The instinct to prohibit agents is understandable but will fail, recreating the identity contest under another name. The alternative is to declare, per assessment, which uses are permitted and to make the human's contribution inside that permission the thing that is measured. Approved use is a specification of conditions, not a loophole. Under it, the receipt records the policy that applied as a consent-and-attestation block in the record itself, along with evidence of the human's work within it, so a verifier can check not only that reasoning happened but that it happened under the declared conditions — did the learner set the goal or accept the agent's, test the output or submit it, notice and correct errors? These observable behaviors distinguish the driver from the passenger more reliably than any attempt to detect whether an agent was present. Where safety is at stake (a licensing examination), the policy says so, the environment enforces it, and the credential is issued at that tier; where a task is better done with tools, the policy and credential say what was measured. The effect is to replace "Was an agent used?" — unanswerable at scale and beside the point — with "What did the human demonstrably do?", which is measurable, disclosable, and worth paying for.

The infrastructure this requires is not exotic: learning environments that emit structured evidence of reasoning as it happens; an open evidence schema and verifier so receipts are portable and checks reproducible; institutional separation between environment, issuer, and verifier; and measurement engines that have themselves been audited against known truth, with audits published whether or not they flatter the engine. The paper also scopes what it does not solve: compromised environments can still defeat evidence capture; contemporaneous evidence is easier in Simulation and interactive learning than in extended written work; observation carries equity and Privacy costs (learner-controlled, minimal-by-default disclosure is necessary); tiering can be gamed by weak certification; and standards bodies, not vendors, should own the schema. The AI Governance and Ethics of such systems, along with the validity arguments of Messick, Kane, and others, apply with particular force — including warnings against circular validation in which one model certifies another.

Key Findings

1. A credential's value is a signaling quantity, not a presence quantity. When an agent can reproduce a credential without the human learning, the credential can no longer separate those who learned from those who did not, invoking Akerlof's lemons condition. Identity checks and proctoring fail on both counts: they are losing contests and they measure the wrong thing (that a person was present rather than that they learned).
2. Evidence must be contemporaneous, inspectable, and independently verifiable. Repair targets the observation of learning itself, not the paperwork around it — capturing predictions, test runs, explanations, and revisions while they happen, with published rules and a verifier separate from both issuer and environment builder.
3. A credential becomes a claim plus a receipt. The receipt is a hashed, structured record of what was observed, when, and under what conditions, kept as a separate chain from any score so that different parties can layer their own estimates on shared evidence. Analogy: the pilot's logbook, not the pilot's license.
4. Dynamic tiering prices three honest states. Unaided, directed, and attested certify increasingly weaker evidence of human reasoning, and tiers can upgrade or lapse as evidence accrues or decays, mirroring recertification in the professions.
5. Approved use beats prohibition. Declaring, per assessment, which agent uses are permitted turns "Was an agent used?" (unanswerable) into "What did the human demonstrably do?" (measurable and worth paying for), recorded in the receipt's consent-and-attestation block.

Implications for AI in Education

• Assessment design shifts from policing to evidence capture. Rather than trying to detect or prohibit agents, institutions should design assessments whose reasoning traces are recorded as they happen, so the credential's signal survives agentic learning environments.
• Open, portable evidence schemas become a public good. The paper argues standards bodies — not vendors — should own the schema, echoing the knowledge base's AI Governance concern about who controls the infrastructure of assessment in the age of generative AI.
• Learner-controlled, minimal-by-default disclosure is a prerequisite. Evidence-rich credentials carry real Privacy and equity costs; the paper insists the learner controls disclosure and the receipt exists to be presented, not harvested.
• Tiering lets the market price honesty. By naming attested credentials for what they are, dynamic tiering avoids a blanket discount on all credentials and gives employers, accreditors, and licensing boards an inspectable middle ground between full audit and unverified claim — a concrete alternative to integrity policing.
• Validity and circular-validation warnings apply directly. The credential rests on a chain of inferences (Kane, Messick); agents break the first link, and systems must avoid validating one model with another. This connects to the knowledge base's broader measurement concerns about Educational Measurement under agentic AI.

Connected Concepts

• Assessment
• Academic Integrity
• Higher Education
• Edtech Platform
• Generative AI
• Agentic AI
• AI Governance
• AI in Education
• Automated Assessment
• Ethics

Connected Articles

• Knowledge, Skills, Attitudes, Production: Competency-Based Education After Generative AI — Knowledge, Skills, Attitudes, Production: Competency-Based Education After Generative AI
• Policy Fragmentation or Institutional Alignment? Institutional Governance of AI in Universities and Business Schools — Policy Fragmentation or Institutional Alignment? Institutional Governance of AI in Universities and Business Schools
• Once a Response, Always a Response: Detecting LLM-generated Text via Latent Prompt Restoration — Once a Response, Always a Response: Detecting LLM-generated Text via Latent Prompt Restoration
• Towards Valid Student Simulation with Large Language Models — Toward Valid Student Simulation with Large Language Models
• Large Scale AI Grading of Handwritten Physics Assessments: Score Agreement and Olympiad Team Selection Outcomes — Large Scale AI Grading of Handwritten Physics Assessments

Citation

Srivastava, S. (2026). Credentials That Carry Their Evidence: Credential design, dynamic tiering, and approved use when AI agents can do the work. EdArXiv preprint.