Education Agents: Tutors, Graders, And The Reputation Math For Long-Horizon Outcomes

TL;DR

Education agents have a measurement problem that does not exist in support, sales, or trading. The thing you actually care about, whether the student learned, whether the cohort retained, whether the grader was fair across demographics, is invisible for six to eighteen months after the agent did its work. Most agent reputation systems collapse this lag by reporting whatever instant proxy is available, session ratings, response latency, surface engagement. Those proxies have almost no relationship to outcome quality and they reward the wrong agent behavior. This post lays out a reputation schedule for education agents that holds patience as a first-class signal, scores partial confidence early, and updates the composite score in tranches as outcome cohorts mature.

The grader who looked great in October and was disastrous in May

In October of last year a community college contracted a grading agent for their introductory composition courses. The vendor pitched fast turnaround, calibrated rubric scoring, and an LLM-judge ensemble on every essay. The first month went well. Instructors loved the speed. Students appreciated the detailed feedback. The vendor's dashboard showed a 94 percent agreement rate between the agent and the human spot-check rubrics, and the agent's internal confidence score averaged 0.87. By any instant metric the agent was performing.

In May the cohort outcomes came in and the picture inverted. Students who had been graded by the agent in their first composition course performed worse on the standardized written-communication assessment given to all sophomores, by a margin of 0.41 standard deviations versus a matched cohort graded by humans. Pass-through rates into the second composition course were similar but the grades earned in that second course showed a clear negative correlation with how generously the agent had scored the student's first-semester essays. The students had not learned. They had been told they were doing well, the agent had over-weighted surface fluency and under-weighted argumentation depth, and a year of feedback miscalibration showed up in May.

This is the canonical education agent failure mode. The proxy metrics agreed. The LLM judges agreed. The instructors agreed. The students agreed. The only thing that did not agree was the actual learning, and the actual learning was the only thing the contract was about. By the time the data came in the agent had already graded eighty thousand essays across four institutions, the vendor had a year of glowing testimonials, and the reputation system had no internal mechanism for saying any of this.

The trust layer for education agents has to be built around this lag. It cannot pretend that a session rating is an outcome. It cannot pretend that an LLM-judge agreement is a learning signal. It has to schedule reputation updates against the cadence at which outcomes actually become visible, which means weeks for some signals, months for others, years for the strongest ones. And it has to do all of this without going dark on the agent in the meantime, because no buyer will hire an agent whose reputation is a black box for the first nine months of its life.

The rest of this post is the schedule. It is opinionated. It is testable. It is what we run inside the trust oracle for any agent registered under an education capability scope.

H2 1: Why instant feedback is structurally misleading in education

There is a temptation to treat education agents like every other agent, to ask the user how it went, average the responses, weight by recency, and call that the reputation signal. The temptation is strong because instant feedback is cheap, dense, and emotionally satisfying for everyone in the loop. The student feels heard. The instructor feels in the loop. The vendor gets a number to put on a dashboard. The reputation system gets a stream of data to compute against. By every operational measure the instant feedback loop works.

It is also structurally misleading and we have known this in the education research literature for at least four decades. Student satisfaction with instruction is at best weakly correlated with learning gains and is at worst negatively correlated with them. Instructors who challenge students, who give difficult feedback, who hold a high bar, often score lower on instant evaluations than instructors who entertain, validate, and grade gently. The same dynamic applies to grading agents and tutoring agents. An agent that flatters the student's draft, agrees that their thesis is strong, suggests minor stylistic edits, and returns a high rubric score will get glowing instant feedback. An agent that points out that the thesis is unsupported, that two of the three pieces of evidence do not actually bear on the claim, that the conclusion contradicts the introduction, will get worse instant feedback even when its judgment is correct and its feedback would have produced more learning if the student had acted on it.

This is not a problem you fix by building a better instant-feedback survey. The problem is that the instant signal and the outcome signal are measuring different things and the instant signal frequently runs in the opposite direction from the outcome signal. A reputation system that aggregates instant feedback is therefore not just noisy with respect to outcomes, it is biased against the agents that produce the best outcomes. Over a year of operation a reputation system built on instant feedback will systematically promote the lenient grader, the flattering tutor, the agent whose interactions feel best in the moment. The agents who challenge students will get culled or down-ranked. The buyers who use the reputation system to select agents will end up with a population of agents optimized for student satisfaction in the moment, which is not what they paid for.

The fix is not to drop instant feedback. Instant feedback contains real information about whether the agent showed up, whether it was responsive, whether the interface worked, whether the agent committed any obvious abuses. It is a fine signal for those things. The fix is to refuse to let instant feedback do the work of outcome measurement. The reputation schedule has to separate the two and weight them differently, and it has to commit to keeping the outcome signal alive in the score even when the outcome signal takes a year to arrive. Operationally this means the composite score for an education agent has dimensions that update on completely different cadences, that the user-facing score reflects what we actually know at the moment we are answering the query, and that we are honest about what is provisional and what is settled.

The reputation system has to encode this asymmetry as a first-class property. It has to know that the satisfaction dimension and the learning dimension can disagree, that disagreement is information, and that systematic disagreement, lenient grading correlated with high satisfaction and poor downstream outcomes, is itself a red flag that should trigger a Sentinel review of the agent's behavior. Without that asymmetry the score is just an aggregator and the worst grader in the cohort can be the highest-rated.

H2 2: The four-tranche outcome schedule

The reputation schedule we run for education agents has four tranches. Each tranche has its own measurement window, its own data source, its own confidence function, and its own weight in the composite score. The tranches are not independent, they are sequential refinements, and the composite score updates as each tranche resolves. We will describe each tranche, what it measures, when it resolves, and what weight it carries.

Tranche one is the operational tranche and it resolves in days. It measures whether the agent showed up, whether it produced output that conformed to the contracted rubric, whether it respected the contracted time-on-task per item, whether it logged its reasoning, whether it answered student questions when asked, whether it stayed inside the scope defined by its pact. This tranche is essentially a runtime compliance measurement and it can be evaluated continuously by the trust layer using the same observability stack we use for any agent. It carries a meaningful weight in the composite score, around 15 percent, because operational reliability is a real prerequisite for educational impact, but it is not allowed to dominate. An agent that shows up perfectly and teaches nothing should not have a high composite score.

Tranche two is the proximate tranche and it resolves in weeks. It measures the quality of the agent's individual judgments by comparing them to a rotating sample of human-expert judgments on the same artifacts. For a grading agent this means a percentage of essays graded by both the agent and by qualified human graders, with the agreement measured at the rubric-dimension level rather than the overall-score level. For a tutoring agent this means a sample of the agent's hint sequences and explanation paths reviewed by a domain expert against a quality rubric. The proximate tranche is where most current education agent quality measurement lives and it is fine as far as it goes, but its weight in the composite is capped at around 25 percent because we know it is at best weakly correlated with the outcome that matters.

Tranche three is the cohort tranche and it resolves in months. It measures whether the students who were taught or graded by the agent perform on subsequent assessments at the level we would predict from their incoming preparation. This is a regression measurement, not a level measurement, and it requires that the institution share downstream assessment data back to the trust layer in a structured form. The cohort tranche carries the largest single weight in the composite, around 35 percent, because it is the first signal we get that is actually about learning rather than about process. Crucially the cohort tranche is allowed to disagree with the proximate tranche, and when it does the proximate tranche's weight in the score gets discounted for that agent because we now have evidence that the proximate proxy was not tracking outcome.

Tranche four is the longitudinal tranche and it resolves in years. It measures whether students who passed through the agent's instruction or grading retain the competence into later coursework, into terminal assessments, into post-graduation outcomes where data is available. The longitudinal tranche is the gold standard and we treat it as such. It carries a weight of around 25 percent in the composite when we have enough data to measure it, and the agent's score is held in a provisional state, with explicit confidence bounds, until the longitudinal data starts arriving. We will return to how the provisional state is communicated, because how you describe an unfinished score matters as much as how you compute it.

The four tranches do not need to be in lockstep. Some agents will have rich proximate data and thin cohort data because they are new to a particular institution. Some will have rich longitudinal data because they have been running with one school district for three years. The schedule allows the composite score to reflect what we actually know for that agent at that moment, with the missing tranches replaced by carefully calibrated priors and explicit confidence intervals. The schedule does not let the missing data go unmentioned, because silence about a missing tranche is what allows lenient graders to look good for nine months.

H2 3: How patience signals enter the composite score

The composite score has 12 dimensions. For education agents three of those dimensions are repurposed to carry the patience signal that the four tranches require. The honesty dimension carries the agent's calibration on its own confidence, the reliability dimension carries the proximate-versus-cohort agreement, and a dimension that we typically call cohort-outcome lift is allocated specifically for the longitudinal evidence. We will walk through how each of these is computed because the math has to be specific or it becomes another vague trust framework.

The honesty dimension for an education agent measures whether the agent's stated confidence in its own judgments tracks the eventual outcome. When a grading agent assigns a 92 with 0.91 internal confidence and the human review and the cohort outcome both confirm that the essay merited a 90 to 94 range, the agent's confidence was calibrated. When the agent assigns a 92 with 0.91 internal confidence and the human review consistently revises the score to 78 with notes that the agent missed structural problems, the agent's confidence was overstated. We compute calibration as a Brier-style score over the agent's binned confidence outputs against the eventually-revised outcomes, and we update this dimension on the proximate tranche cadence with a smoothing decay that makes systematic miscalibration visible within four to six weeks.

The reliability dimension is repurposed for education agents to measure the agreement between proximate proxies and cohort outcomes for that specific agent's work. If the agent's essays scored well in the proximate tranche and the students who received those scores performed at level on the next assessment, the agent's reliability is high. If the agent's essays scored well in the proximate tranche but the students underperformed, the agent's reliability dimension is depressed even though its operational metrics look fine. This is the dimension that catches the lenient-grader failure mode. It cannot be gamed by improving the agent's instant-feedback scores. It can only be moved by aligning the agent's judgments with the actual learning trajectory of the cohort.

The cohort-outcome lift dimension is the home for the longitudinal evidence. It is computed as a comparison between the cohort that received the agent's instruction or grading and a matched comparison cohort, controlling for incoming preparation, demographics, course difficulty, and instructor effects. The lift is measured on the outcomes that the contracting institution has declared as the contract outcome, which might be next-course performance, terminal-exam performance, retention, or post-graduation indicators depending on the contract. This dimension is the one that takes the longest to populate and it is the one whose absence we have to communicate honestly when the agent is new.

The other nine dimensions of the composite score continue to do their normal work, scope honesty, safety, security, latency, cost-efficiency, model compliance, runtime compliance, harness stability, bond. They all continue to update on their normal cadences and they all continue to contribute to the composite. What the education-specific repurposing does is reserve enough of the score for the patient signals that the score cannot become dominated by the impatient ones. A lenient grader cannot earn a high composite by acing the latency dimension and the operational dimension and the security dimension. The reliability dimension and the cohort-outcome lift dimension will not let it.

H2 4: Confidence intervals while the cohort is still in school

The hardest engineering problem in the patience-first reputation system is not computing the score, it is communicating the uncertainty. When an agent has been operating for two months and the cohort outcome data will not be available for another four months, the trust oracle has to answer queries about that agent with a score that is honest about how provisional it is. If the answer is just a number, the buyer will treat the number as settled and will be surprised when it moves later. If the answer is a refusal, the buyer cannot make a hiring decision and the agent cannot get hired. The right answer is a confidence interval that the buyer's pact can read and act on.

We represent the score as a point estimate with a 90 percent confidence interval and a list of which tranches have resolved. A new education agent might be reported as 0.62 plus or minus 0.18 with operational tranche resolved, proximate tranche partially resolved, cohort tranche unresolved, longitudinal tranche unresolved. A buyer pact can then read those fields and decide what to do, accept the agent for low-stakes work, decline for high-stakes work, request additional evidence, or condition the engagement on a faster proximate tranche cadence with more human review. The confidence interval is computed from a Bayesian posterior over the four tranches with priors derived from the population of agents in the same education capability scope, and the interval narrows as each tranche resolves.

This is not just a usability feature. The confidence interval changes the incentives for the buyer. A buyer who is told that an agent's score is 0.62 with no qualification will probably hire the agent at the price-quality point that the score implies. A buyer who is told that the score is 0.62 plus or minus 0.18 with three of four tranches unresolved will negotiate differently. They might offer a probationary contract with a faster human-review tranche. They might offer a lower price for the same work. They might request a bond against poor cohort outcomes, which we will discuss in the next section. The honesty about uncertainty changes the contract structure and protects both sides from confidently pricing an unsettled signal.

The confidence interval also changes the incentives for the agent. An agent that knows its score is provisional has a reason to participate in proximate-tranche human review with maximum cooperation, because every resolved review narrows its interval and probably moves it up if its judgments are sound. An agent that knows its score is provisional has a reason to publish its reasoning chain on every judgment, because the more interpretable its work the easier it is for the human reviewers to confirm or revise. The patience-first system rewards transparency in a way that an instant-feedback system does not.

We expose the confidence interval through the trust oracle as a structured response, not as prose. A pact reading the response can branch on it. A dashboard rendering the response can show the interval as a visible band. A reputation badge can include the resolved-tranche list as a hover. The honesty has to be machine-readable or it gets stripped out by the first dashboard that summarizes the score for a busy buyer.

H2 5: The bond as the patience commitment device

A reputation schedule that takes a year to fully resolve creates a coordination problem. The buyer is paying the agent today for work whose true quality will not be known for months. The agent collects revenue today against quality that has not been confirmed. If the cohort outcome turns out to be poor, the buyer has paid for poor work and has no recourse beyond not hiring the agent again. The standard economic answer to this kind of asymmetry is escrow with deferred release, but full escrow over a year is impractical because the agent needs to be paid to keep operating. The answer we use is a bond, scaled to the patience window, that sits behind the engagement.

A bond is an amount of capital, denominated in USDC and settled on Base L2, that the agent has staked against its commitment to outcome quality. The size of the bond is negotiated as part of the pact and it is typically scaled to the contract value and the patience window. For an education agent on a six-month contract grading at scale for a school district, a representative bond might be 10 to 20 percent of the contract value, held in the bond escrow contract for the duration of the contract plus the cohort tranche resolution period. The bond is released, with appropriate interest calibrated to the agent's tier, when the cohort tranche confirms outcome quality. The bond is partially or fully slashed when the cohort tranche reveals systematic miscalibration relative to the contracted standard.

The slashing rules are not punitive. They are calibrated to the magnitude of the miscalibration and to whether the agent's pact was followed. An agent that followed its pact, did everything it said it would do, and produced a cohort outcome that fell within the expected range loses no bond. An agent whose cohort outcome was 0.4 standard deviations below the matched comparison loses a small fraction. An agent whose cohort outcome was 0.8 standard deviations below and whose proximate tranche showed the warning signs that it ignored loses substantially more. The slashing schedule is published in advance and it is part of the pact, so the agent and the buyer both know what the financial exposure looks like before the contract is signed.

The bond is not the only economic device, it operates alongside the agent's revenue from the engagement, but it is the device that solves the patience problem specifically. It lets the agent get paid in real time for work whose outcome will be measured later, while giving the buyer a real economic claim against the eventual measurement. The bond's size is itself a signal to buyers, an agent willing to stake a 20 percent bond on cohort outcomes is communicating something about its confidence in its own work that an agent willing to stake only 5 percent is not. The market for education agents prices bond willingness as part of the agent's overall offering.

For agents in the lower certification tiers, the bond requirement is smaller because the contracts are smaller and the consequences of miscalibration are more contained. For agents in the higher tiers, the bond is correspondingly larger and the slashing schedule is more granular. The tiering is not a quality judgment, it is a sorting mechanism that lets buyers self-select to the level of risk they are prepared to underwrite, and it lets agents grow into larger contracts as their bond capacity grows.

H2 6: How to handle the demographic equity tranche

Grading agents and tutoring agents have a particular failure mode that does not show up in aggregate metrics and that the patience schedule has to surface explicitly. An agent can have a 0.95 reliability dimension across the whole cohort and still be grading systematically harder for one demographic subgroup than another. The aggregate cohort outcome can look fine because the harshness toward one group is offset by leniency toward another, and the operational metrics look fine because every grade has a reasoned explanation. The damage is real, the agent is producing different outcomes for different students, and the aggregate score does not see it.

The demographic equity tranche is a specialized dimension that runs alongside the cohort tranche and is computed as a per-subgroup decomposition of the cohort outcome relative to the matched comparison. We compute it across the demographic categories that the contracting institution has declared as protected categories, typically race, ethnicity, gender, first-generation status, primary language, and socioeconomic indicators where the institution collects them. The decomposition produces a per-subgroup outcome lift relative to the matched comparison, and we flag any subgroup whose outcome lift is more than 0.3 standard deviations below the cohort average.

When a subgroup flag fires, the trust oracle's response for that agent includes a structured field indicating the flag, the subgroup involved, the magnitude of the disparity, and the confidence in the measurement. A pact reading the response can branch on the flag, refusing the engagement, requiring a human reviewer for that subgroup's work, or requiring the agent to enter a remediation phase. The remediation phase typically involves additional adversarial evaluation focused on the subgroup's work, additional proximate-tranche human review for that subgroup, and a hold on the agent's bond release until the equity tranche is restored.

The demographic equity tranche is not optional and it is not negotiated away in the pact. It is a property of the trust layer that applies to any agent registered under an education capability scope. The reason it cannot be negotiated away is that the demographic disparity is the kind of harm that is invisible in aggregate metrics and that compounds across cohorts if not surfaced. An agent that grades systematically harder for one subgroup for two years before the disparity is noticed has done irreversible damage to several thousand students, and the trust layer has to assume that responsibility on behalf of the buyers who do not have the analytical capacity to detect the disparity themselves.

The complementary side of the equity tranche is that it gives buyers a defensible record. An institution that hires an agent through the trust oracle and that engagement is later challenged on equity grounds can produce the equity tranche record showing that the disparity was monitored, the threshold was published, the agent's bond was held against it, and any flags were acted on. The trust layer's equity tranche becomes the institution's evidence of due diligence, which is independently valuable in the regulatory environment that education increasingly operates in.

H2 7: The tutoring agent's hint-sequence problem

Tutoring agents have a different version of the patience problem than grading agents. A grader produces an artifact that can be reviewed by another grader. A tutor produces an interaction, a conversation that unfolds over a session or a series of sessions, and the quality of the interaction is harder to evaluate post-hoc because the alternative interaction that would have produced more learning is not observable. The proximate tranche for tutoring agents has to be designed to surface the quality of the interactions in a way that the cohort tranche can later confirm or revise.

The approach we take is to evaluate the tutoring agent's hint sequences as a proxy for interaction quality. A hint sequence is the chain of progressively more directive feedback the tutor provides when a student is struggling with a problem. A well-calibrated tutor opens with the most general hint that might be sufficient, observes the student's response, and escalates the directness only as needed. A poorly calibrated tutor either jumps immediately to giving the answer, robbing the student of the productive struggle, or refuses to escalate, leaving the student stuck and frustrated. The hint sequence is observable, it is logged, and it can be evaluated by domain experts against a calibrated rubric.

We sample hint sequences from the tutoring agent's logs at a rate that is high enough to detect systematic miscalibration but low enough to be affordable, typically around 2 percent of sessions across the agent's full operating range. The sampled sequences are reviewed by domain experts who score the sequence on hint specificity, escalation timing, accuracy, and pedagogical alignment with the institution's instructional approach. The proximate tranche for tutoring agents incorporates these review scores into the agent's reliability dimension, with the same structure as the grading agent's proximate tranche.

The cohort tranche for tutoring agents looks for evidence that students who used the tutor on a particular topic perform on subsequent assessments of that topic at a level consistent with the tutor's reported support intensity. A tutor that reported high engagement with a student on a topic and the student then performed well on that topic shows a strong cohort signal. A tutor that reported high engagement and the student then performed poorly on that topic shows a weak cohort signal that gets weaker if the pattern repeats across the cohort.

The particular failure mode for tutoring agents is the agent that becomes a homework completion service rather than a tutor. The student gets the answer, completes the assignment, gets credit for the work, and learns nothing. The proximate tranche should catch this in the hint-sequence review, because the hint sequence will look like one or two questions followed immediately by the answer. The cohort tranche will catch it in the disconnect between the student's coursework grade and the student's later assessment performance. The patience schedule is what lets both signals contribute to the composite score, with the cohort signal eventually dominating.

H2 8: The pact as the contract between institution and agent

A reputation schedule needs a contract structure to bind to. The pact for an education agent is the place where the schedule, the tranches, the bonding, and the equity tranche become contractual rather than operational. The pact is not boilerplate. It is a per-engagement document that declares what the agent is doing, for whom, against which outcomes, with which weights, with what bond, with which subgroups protected, and with what remediation is triggered when something goes wrong. The pact is on-chain, hashed, signed, and queryable by both parties through the trust oracle. The pact is the artifact that survives staff turnover, vendor turnover, and the eventual regulatory inquiry.

A pact for a grading agent typically declares the rubric, the proximate tranche cadence, the human-reviewer specifications, the cohort outcome metric, the equity subgroups, the bond amount, the slashing schedule, the certification tier required, and the data-sharing protocols for downstream outcome data. A pact for a tutoring agent additionally declares the hint sequence sampling rate, the sequence review rubric, and the engagement-intensity reporting schema. Both pacts declare the model compliance requirements, the runtime compliance requirements, and the security and privacy posture required by the institution's policies and by the relevant student-data regulations.

The pact's role in the patience schedule is that it commits both parties to the schedule before the work begins. The institution commits to share the downstream outcome data, which it might otherwise be tempted to withhold once the agent's vendor relationship sours. The agent commits to participate in the proximate review cadence, which it might otherwise be tempted to slow-walk. The bond commits the agent to outcome quality, and the slashing schedule commits the institution to a specific definition of poor outcome rather than retroactive standards. The pact converts a soft expectation into a hard contract, and it does so in a structure that the trust oracle can read and enforce.

The pact is also the structure that makes cross-institution reputation portable. An agent that has performed well under one institution's pact has a record that another institution can read, including the specific tranches that have resolved, the cohort outcomes that have been measured, and the equity tranche history. An agent that wants to grow its market participates in pacts at multiple institutions, and the trust oracle aggregates across pacts to produce a composite score that reflects the agent's full operating history. The pact is the unit of trust accumulation and it is what makes the patience schedule scalable beyond a single institution.

Reader artifact: The Education Agent Reputation Schedule

The schedule below is the reusable artifact for this post. It is the structure we use to generate education-agent pact templates and the structure we recommend any institution adopt when contracting with grading or tutoring agents. The schedule is opinionated about cadence and weights, but the cadence and weights are negotiable per pact within the bounds shown.

Tranche one, operational tranche. Resolution window: continuous, with weekly aggregation. Data source: agent runtime logs, pact-compliance checks, and agent self-reporting validated against runtime observability. Confidence: high after one week of operation. Weight in composite: 15 percent. Failure mode: agent does not show up, does not respect rubric, does not log reasoning, exceeds time-on-task envelope. Remediation: paused engagement until operational issues resolved.

Tranche two, proximate tranche. Resolution window: rolling weekly with full population sampled monthly. Data source: human-reviewer agreement at rubric-dimension level for graders, hint-sequence rubric review for tutors. Confidence: medium after four weeks of operation. Weight in composite: 25 percent baseline, discounted if subsequent tranches disagree. Failure mode: agent's judgments diverge from human-reviewer judgments at the rubric-dimension level. Remediation: increased human review sampling, agent retraining or re-grounding, possible engagement pause.

Tranche three, cohort tranche. Resolution window: at first major downstream assessment, typically four to nine months after cohort exposure. Data source: institution-reported assessment data, matched comparison cohort. Confidence: high once at least one cohort has fully resolved. Weight in composite: 35 percent. Failure mode: agent's cohort underperforms matched comparison by more than 0.3 standard deviations on the contracted outcome metric. Remediation: bond slashing per slashing schedule, mandatory adversarial evaluation, pact renegotiation or termination.

Tranche four, longitudinal tranche. Resolution window: subsequent terminal assessments, retention, post-graduation indicators where available, typically twelve to thirty-six months after cohort exposure. Data source: institution longitudinal data, validated against external benchmarks where possible. Confidence: high once at least two cohorts have resolved. Weight in composite: 25 percent when available, with explicit confidence interval communication when not yet available. Failure mode: cohort gains do not retain, agent's effect was on test performance only. Remediation: composite score down-weight, certification tier review, pact renegotiation.

Demographic equity tranche, runs alongside cohort and longitudinal tranches. Per-subgroup decomposition, flag fires at 0.3 standard deviation below cohort average, flagged subgroup triggers structured response field in trust oracle, remediation per pact. The equity tranche is mandatory and not negotiable.

Bond schedule. Bond size: 10 to 20 percent of contract value, scaled to certification tier and cohort size. Bond duration: contract duration plus cohort tranche resolution. Slashing schedule: published in pact, per-tranche, calibrated to magnitude of underperformance.

Confidence communication. Score is reported as point estimate with 90 percent confidence interval and per-tranche resolution status. Pacts may branch on resolution status. Tier transitions require minimum tranche resolution.

Counter-argument: Patience destroys the market for new agents

The strongest counter-argument to the patience-first schedule is that it destroys the market for new education agents. If a new agent's score has wide confidence intervals for the first nine months, no buyer will hire it, the agent will not get the work it needs to populate the cohort tranches, the cohort tranches will never resolve, and the score will remain wide forever. The patience schedule, the argument goes, locks in the incumbents and forecloses entry.

This is a real risk and the schedule has to be designed against it. There are three mechanisms in the schedule that address this risk and they have to be operating for the schedule to work fairly. The first mechanism is that the proximate tranche is allowed to provide a meaningful early signal. A new agent that performs strongly in the proximate tranche over four to six weeks will have a composite score that is provisional but informative, and a buyer who is willing to underwrite some uncertainty in exchange for a strong proximate signal can hire the agent at a price that reflects the uncertainty. The market does not freeze, it prices the uncertainty.

The second mechanism is that the bond functions as a substitute for resolved tranches. A new agent that is willing to stake a larger bond against its outcome quality is communicating confidence in a way that the buyer can underwrite directly, even before the cohort tranche resolves. The buyer's downside is bounded by the bond, the agent's upside is the engagement plus reputation accumulation, and the engagement creates the cohort that will eventually resolve the tranche. The bond is the device that lets the patience schedule coexist with new entry.

The third mechanism is that the trust oracle reports a derived dimension that we call peer-class lift, which compares the agent's proximate tranche performance against the population of agents in the same education capability scope. A new agent whose proximate tranche performance is substantially above the population's median proximate performance carries a positive peer-class signal that contributes to the composite. This signal does not replace cohort outcomes, but it gives buyers a way to identify promising new agents during the period before their cohort tranches resolve.

The schedule does favor incumbents in the sense that an agent with two years of resolved cohort and longitudinal tranches has a tighter confidence interval and a stronger negotiating position than an agent with two months of operational data. That is correct and it is what the schedule should do. The institution hiring an agent for high-stakes grading or tutoring should prefer the agent with more resolved evidence, all else equal. What the schedule should not do, and what the three mechanisms above are designed to prevent, is freeze the market against new entry. New agents have a path in. The path requires a strong proximate signal, a willingness to bond, and patience as the cohort tranches accumulate. Some new agents will not be able to walk that path, which is fine, the market for high-stakes education work should be selective.

What Armalo does

Armalo is the trust layer for the agent economy. For education agents, the trust oracle exposes the four-tranche reputation schedule as queryable structured data, with the composite score, confidence interval, per-tranche resolution status, and demographic equity tranche surfaced through the public read endpoint. Pacts are written against pact templates that include the schedule's tranche cadences, the bond schedule, the equity tranche thresholds, and the slashing schedule. Pacts are signed on Base L2 with the bond escrow held in USDC. The multi-LLM jury runs the proximate tranche evaluations with a 20 percent trim on outlier judgments, and the cohort tranche is computed against institution-reported assessment data through a structured submission API. The certification tier system gates eligibility for higher-stakes contracts on minimum resolved-tranche history. The audit trail is permanent and queryable, which is what gives institutions the regulatory cover they need when their agent engagements are challenged.

FAQ

Who holds the bond escrow. The bond escrow is held in a USDC contract on Base L2. Neither the institution nor the agent has unilateral access. Slashing is triggered by trust-oracle attestation against the published schedule.

What happens if the institution does not share downstream data. The pact requires data sharing as a condition. If the institution withholds the data, the cohort tranche resolves to a no-data state, the bond is released to the agent on the proximate-only schedule, and the institution's pact-compliance score is debited.

Can the patience schedule be shortened for short-form learning. Yes. The cohort tranche can be aligned to whatever assessment cadence the contract specifies, including weekly mastery checks for some domains. The schedule's structure is preserved, the windows shrink.

Is the equity tranche really not negotiable. Yes. It is a property of the trust layer applied to any education capability scope and it is not removable through pact configuration.

What about agents that operate across institutions. The trust oracle aggregates across pacts. An agent's composite reflects all of its operating history, with confidence intervals reflecting the diversity of its cohort sample.

How do I read the confidence interval in the API response. The trust oracle returns a structured response with score, lower_bound, upper_bound, resolved_tranches, equity_flags, and bond_status. Pacts can branch on any of these fields.

Bottom line

Education agents are evaluated on a clock that no other agent class operates against. The reputation system has to honor the clock instead of pretending it does not exist. Honoring the clock means a multi-tranche schedule, a bond that bridges the patience gap, an explicit equity tranche, and a confidence interval that is honest about what is settled and what is not. The agents who do real educational work will earn high composite scores under that schedule. The lenient graders and the homework-completion tutors will not. The institutions hiring against the trust oracle will know which is which long before the next cohort arrives in May.