Why Time Decay Is the Most Important Feature in Agent Trust Scoring
Static trust scores are dangerous. An AI agent that scored 950 twelve months ago and hasn't been evaluated since is not the same agent it was then. Time decay — one point per week after a grace period — is the mechanism that makes trust scores reflect current behavior rather than historical achievement. Here's why this matters more than any other single feature.
Why Time Decay Is the Most Important Feature in Agent Trust Scoring
There's a design choice in trust scoring that seems minor — almost technical — until you understand what it prevents. That choice is time decay: the mechanism that reduces an AI agent's trust score over time when that agent hasn't received recent evaluations.
The choice seems minor because the mechanism is simple: one point per week after a seven-day grace period. An agent with a score of 900 that hasn't been evaluated in three months will have a current score of approximately 887. The math isn't complicated.
What the mechanism prevents is catastrophic: the decoupling of reputation from current behavior. Without time decay, a trust score is a historical achievement. With time decay, it's a continuous behavioral assessment. The difference between these two properties is the difference between a trust system that creates genuine accountability and one that creates the appearance of accountability while enabling all the failure modes it's supposed to prevent.
TL;DR
- Static scores enable reputation laundering: An agent that performed excellently months ago can behave poorly now while displaying a high historical score — time decay closes this gap.
- The 7-day grace period is a feature: Newly-evaluated agents shouldn't see their score degrade immediately — the grace period allows score establishment before decay begins.
- Behavioral drift happens on a timeline of weeks to months: Time decay calibrated to one point per week catches drift at the timescale it actually matters.
- Active evaluation maintains scores: Agents that are consistently evaluated maintain their scores; inactive agents see natural score reductions.
- Time decay creates continuous behavioral incentives: Not just a one-time certification incentive — an ongoing incentive to maintain quality across the entire operating lifetime.
Static vs. Dynamic Scoring Comparison
| Property | Static Scoring (No Decay) | Dynamic Scoring (With Decay) |
|---|---|---|
| What the score reflects | Historical performance at last evaluation | Recent performance, continuously weighted |
| Incentive structure | Incentive to achieve a high score once | Incentive to maintain high performance continuously |
| Reputation laundering vulnerability | High — legacy scores persist indefinitely | Low — scores decay without ongoing evaluation |
| Behavioral drift detection | Poor — drift invisible until the next evaluation | Good — declining score signals drift proactively |
| Score freshness | Unknown — could be from any historical period | Knowable — decay tells you how fresh the evaluation evidence is |
| Cheating resistance | Low — pass evaluation once, done | High — must sustain performance to sustain score |
| Operational relevance | Decreases over time | Maintained by continuous evaluation |
| Buyer confidence | False confidence in outdated scores | Accurate confidence calibrated to evidence age |
The Case Against Static Scores
Imagine a hiring scenario. A company wants to hire a software engineer. The candidate presents a portfolio of excellent work from five years ago. The work genuinely was excellent — well-written, well-designed, technically sophisticated. The candidate's score on a hypothetical "coding ability" assessment based on that portfolio would be very high.
But five years is a long time in software. Languages evolve. Best practices change. Tools shift. The candidate may have maintained and sharpened their skills over the intervening years, or may have moved to non-technical work and let their skills atrophy. The portfolio score tells you nothing about which scenario applies.
This is the static score problem. Applied to AI agents, it's worse because AI systems can change dramatically faster than human skills. An agent model can be retrained in hours. System prompts can be modified in minutes. The underlying behavior can shift in ways that make a three-month-old evaluation essentially useless as a predictor of current behavior.
Without time decay, a trust system has no mechanism to signal that evidence is becoming stale. An agent evaluated once, scoring 950, then never evaluated again will display that 950 indefinitely. Buyers making deployment decisions based on that score are trusting a historical record, not a current behavioral assessment.
The 7-Day Grace Period: Why It's Necessary
The grace period is the interval after an evaluation during which the score remains stable before decay begins. Seven days was chosen based on several considerations.
First, evaluation turnaround: in practice, evaluations don't happen continuously in real time. A batch of evaluations might complete over a day or two. The grace period ensures that an agent doesn't start decaying before the evaluation batch that just ran has fully contributed to its score.
Second, operational noise: agents experience legitimate short-term behavioral variation due to infrastructure changes, load patterns, and query distribution shifts. These variations shouldn't immediately cause score decay if they're transient. The grace period absorbs normal operational variation without triggering decay.
Third, incentive calibration: without any grace period, the decay would create pressure to evaluate agents every single day to prevent score loss. This is too frequent for most production deployments and would create evaluation overhead that discourages participation. Seven days creates a sustainable evaluation cadence.
The grace period is not a license to coast. After day seven, decay begins — one point per day — and continues until the next evaluation occurs. An agent on a weekly evaluation cadence (evaluations every 7 days) maintains a stable score. An agent on a bi-weekly cadence will see a slight but manageable decline. An agent that isn't evaluated for a month will see a meaningful score reduction.
Calibrating the Decay Rate: Why One Point Per Week
The one-point-per-week decay rate wasn't chosen arbitrarily. It's calibrated to the empirical rate at which AI agent behavioral drift becomes meaningful.
Behavioral drift in AI agents accumulates through several mechanisms: model updates that shift the distribution of outputs, prompt engineering changes that alter the agent's behavior, retrieval augmentation changes that affect the agent's knowledge base, and tool version updates that change what capabilities are available. Each of these mechanisms operates on a timescale of weeks to months for most production deployments.
At one point per week, a score of 850 will decay to 800 in fifty weeks of no evaluation — roughly one year. This means that a one-year-old evaluation produces a current score approximately 6% lower than the original score. For an agent with a Bronze certification threshold of 600, one year of inactivity reduces the score by about 7% — enough to signal staleness but not enough to immediately decertify an agent with a strong historical record.
At a higher decay rate — say, five points per week — a one-year-old evaluation would produce a score reduced by 33%. This creates too much pressure for rapid re-evaluation and punishes agents in less-active deployment scenarios disproportionately. At a lower decay rate — say, one point per month — a one-year-old evaluation produces a score only 2% lower, which is too small to signal meaningful staleness.
The one-point-per-week calibration threads the needle: meaningful signal of evaluation age without creating operational pressure that's disproportionate to the actual risk.
How Time Decay Prevents the Three Key Failure Modes
Prevents reputation laundering. The reputation laundering failure mode: an agent with a high historical score behaves poorly in the present while the high score remains visible. Time decay directly prevents this. An agent that behaved well eighteen months ago but has received no evaluation since has a score that has decayed by approximately 77 points (from the original evaluation). This is a visible, queryable signal that the evidence is old. A buyer who sees a current score of 873 versus an evaluation-adjusted historical score of 950 knows to ask: when was this agent last evaluated?
Enables behavioral drift detection. The behavioral drift failure mode: an agent's behavior changes while its score remains static, because no re-evaluation has been triggered. With time decay, even without new evaluations, the score declines — creating a signal that re-evaluation is warranted. For agents in active continuous evaluation, the score trajectory reflects drift proactively: a score that was 850 three months ago and is 820 today, despite ongoing evaluations, signals that recent evaluations have been producing lower scores than historical ones.
Creates continuous quality incentives. The one-time certification failure mode: an agent achieves a high score through an intense evaluation push, then relaxes its quality standards because the certification is achieved. Time decay eliminates the one-time achievement incentive by making the score self-correcting. No matter how high an agent's historical score, maintaining that score requires ongoing evaluation performance. The certification tier is only maintained through sustained compliance.
Practical Implications for Agent Operations
Understanding time decay has direct operational implications for organizations running AI agents.
Evaluation cadence planning. At one point per week decay, an agent that wants to maintain a score within 10 points of its peak needs evaluations at least every ten weeks. For Gold certification (minimum score 800), an agent with a score of 820 has approximately twenty weeks before decay would threaten the certification threshold — assuming all evaluations going forward maintain the current quality level.
Score monitoring as an operational signal. A declining score that isn't explained by declining evaluation performance is a sign that the evaluation cadence is insufficient. An unexpectedly declining score despite steady evaluation results is a sign of behavioral drift: recent evaluations are producing lower scores than historical ones.
Pre-certification evaluation sprints. Organizations seeking to move an agent to a higher certification tier can conduct evaluation sprints to accumulate the required evaluation count faster. The time window requirements (90 days for Gold, 365 days for Platinum) prevent pure sprint strategies, but within those windows, higher evaluation density produces more reliable scores.
Frequently Asked Questions
Does time decay penalize agents in niche deployment scenarios where evaluations naturally happen less frequently? The one-point-per-week decay rate is calibrated to avoid penalizing infrequent-deployment agents disproportionately. An agent deployed quarterly will see a score reduction between deployments, but the reduction is modest relative to the uncertainty that the absence of recent evaluations creates. If the decay rate caused severe penalties for infrequent deployment, it would create perverse incentives to evaluate artificially rather than reflect real operational quality.
What happens to the decay if an agent is intentionally deactivated and then reactivated? Decay continues regardless of the agent's operational status. A deliberately deactivated agent's score will decay over the inactivity period. When reactivated and evaluated, the new evaluations will update the score. Organizations that deactivate agents for extended periods should plan for a re-evaluation phase when reactivating.
Can decay be paused for agents under contract review or audit? The decay mechanism doesn't have a pause feature — it's intentional that the score reflects the age of behavioral evidence regardless of operational circumstances. Organizations that need to maintain agent scores during review periods should continue evaluations during the review, not pause the decay.
Does the grace period reset with every evaluation? Yes. Every new evaluation restarts the seven-day grace period. An agent that is evaluated on Monday, then evaluated again on Wednesday, will have its grace period running from Wednesday's evaluation. Consistent evaluation prevents decay from ever beginning.
How does time decay interact with certification tier maintenance? Certification tiers have score floor requirements. Gold requires a current score of 800 or above. If an agent's score decays below 800 due to inactivity or declining evaluation quality, it will lose Gold certification — the tier is downgraded automatically when the floor is no longer met. This is the intended behavior: certification status should reflect current verified performance, not historical achievement.
Key Takeaways
- Require time decay in any trust scoring system you build or adopt — static scores are not trust scores, they're historical records.
- Calibrate evaluation cadence to the decay rate — understand how quickly your agent's score will decay without re-evaluation and plan accordingly.
- Monitor score trajectory, not just level — an agent whose score is declining despite ongoing evaluations has a behavioral drift problem worth investigating.
- Treat the decay-adjusted score as the authoritative number — when evaluating third-party agents, request the current decay-adjusted score, not the historical peak score.
- Design evaluation infrastructure for continuous operation, not point-in-time certification — the operational overhead of continuous evaluation is the cost of genuine trust.
- Use score decay as a business intelligence signal — agents with declining scores are candidates for investigation; agents with improving scores are evidence that quality improvements are working.
- Communicate time decay properties to buyers and counterparties — a buyer who understands that the current score reflects recent behavior, not just historical achievement, can make better-calibrated deployment decisions.
--- Armalo Team is the engineering and research team behind Armalo AI — the trust layer for the AI agent economy. We build the infrastructure that enables agents to prove reliability, honor commitments, and earn reputation through verifiable behavior.
Put the trust layer to work
Explore the docs, register an agent, or start shaping a pact that turns these trust ideas into production evidence.
Comments
Loading comments…