Trust Elasticity
Static governance creates the failures it tries to prevent
The problem with always-on
137 of 138 turns converged, and zero governance interventions fired. The safety net existed, but it did not need to speak.
That result came from a system where governance intensity was not fixed. Most governance frameworks treat every interaction as if it carries the same risk: a harmless lookup gets the same ritual as a consequential decision, and a capable agent on a clean task pays the same overhead as a struggling agent on a failing one. The system does not ask whether intervention is warranted. It assumes it always is.
That assumption has a cost, and the Governance Paradox names it: governance overhead competes with reasoning capacity. If that competition is real, then maximum oversight at all times becomes a standing tax on every task, regardless of whether the task needs supervision at all.
The question is whether governance intensity should be fixed or adaptive, and what it should respond to.
What Trust Elasticity claims
Trust Elasticity is a design principle: governance intensity should scale inversely with demonstrated competence.
When an agent converges quickly, stays coherent, and recovers cleanly from ambiguity, the system loosens oversight. When the agent starts repeating itself, hedging, or failing to close, the system tightens. Here, “trust” is a control variable: the amount of autonomy a system has earned through its recent behaviour. That is Trust Elasticity (classified O-4 in the AI Adoption Science taxonomy — an Observed Principle).
The principle sounds obvious. Most governance is not built that way. Static governance is easier to explain to review boards. “We apply the same controls everywhere” is a comfortable sentence. “We apply less oversight when the system proves it is behaving well” requires you to define what good behaviour looks like and to trust the detection mechanism. That is harder to defend in a steering committee.
The N-Pattern
The simplest operational version of Trust Elasticity is a heuristic we call the N-Pattern (O-6 in the taxonomy — an Observed Heuristic).
N = 1 — PASS. First attempt; trust defaults high.
N ≥ 2 — WARN. Second attempt on the same problem; check stability.
N ≥ 3 — HALT. Third attempt on the same pattern; stop and recover.
The heuristic matters because it separates iteration from looping. Iteration means the system is learning and moving forward. Looping means the system is stuck and disguising repetition as effort. The N-Pattern gives operators a cheap way to notice the difference before they spend another ten minutes watching an agent dig deeper into the same hole.
Iteration count is useful because it is cheap. It is also blunt. A third attempt can be perfectly legitimate on a hard problem. A first attempt can already be a loop if the agent is just rephrasing the same failed approach with different variable names. Later work therefore added multi-signal escalation: semantic repetition detection, confidence trend analysis, and iteration count as a fallback backstop when richer signals are unavailable. The governing rule is to respond to evidence of failure rather than the mere passage of turns; iteration count was the first workable proxy, and later work moved beyond it.
Behavioural trust
A lot of AI trust discourse is still built around what people say they trust. In live review sessions the split is easy to spot: operators describe themselves as cautious, then let a clean run continue untouched until the first visible wobble. Surveys, sentiment scores, stated comfort levels measure attitude, not behaviour, and the two diverge more often than the literature admits.
What matters operationally is different: how much autonomy does the operator actually extend? When does the system ask for help? How quickly does a failing interaction get interrupted? How often does the agent converge without intervention? Those are behavioural signals, and they tell you more about the real trust relationship than any questionnaire.
Trust Elasticity converts that observation into a design decision. High-performing behaviour earns autonomy. Divergent behaviour triggers critique. This is an engineering decision about when to intervene and when to stay out of the way. Scharowski et al. (2022) make a similar distinction between attitudinal and behavioural trust in explainable AI.
Behavioural trust
A lot of AI trust discourse is still built around what people say they trust. In live review sessions the split is easy to spot: operators describe themselves as cautious, then let a clean run continue untouched until the first visible wobble. Surveys, sentiment scores, stated comfort levels measure attitude, not behaviour, and the two diverge more often than the literature admits.
What matters operationally is different: how much autonomy does the operator actually extend? When does the system ask for help? How quickly does a failing interaction get interrupted? How often does the agent converge without intervention? Those are behavioural signals, and they tell you more about the real trust relationship than any questionnaire.
Trust Elasticity converts that observation into a design decision. High-performing behaviour earns autonomy. Divergent behaviour triggers critique. This is an engineering decision about when to intervene and when to stay out of the way. Scharowski et al. (2022) make a similar distinction between attitudinal and behavioural trust in explainable AI.
Why this matters for practitioners and buyers
Trust Elasticity matters because good governance has to satisfy two constraints at once: it must reduce risk, and it must preserve useful work. Static governance usually over-delivers on visible process and under-delivers on that balance.
For teams building AI workflows, the practical question is whether your governance layer knows the difference between a system that needs intervention and one that does not. Reliability looks different: low drag when healthy, escalation when drift appears.
For procurement teams evaluating AI tools, the question is whether the oversight layer improves outcomes or adds ritual. A vendor that can explain when their system tightens supervision — and when it loosens — is telling you something about how seriously they have thought about the tradeoff. A vendor that says “we apply maximum safety at all times” may simply not have measured the cost of doing so.
Trust Elasticity rests on repeat observation, not controlled experiment. Voice Protocol Alpha recorded 138 turns with 137 convergences and zero governance interventions fired. The session stayed governed because the interaction stayed healthy, not because the safety net kept interrupting it.
The instruction sensitivity study reinforced the pattern from the other direction: heavier governance degraded all four agents tested. What we do not yet have is controlled replication across different agents, task types, and risk categories. That gap is real. It is also typical of early-stage design principles: the mechanism is clear enough to act on, not yet mature enough to call settled.
If the Governance Paradox names the problem, Trust Elasticity is the operating principle for dealing with it: governance that responds to what is actually happening, and stays quiet when the system does not need it to speak.
References
Scharowski, N., Perrig, S. A. C., von Felten, N., & Bruhlmann, F. (2022). Trust and Reliance in XAI — Distinguishing Between Attitudinal and Behavioral Measures. arXiv. https://arxiv.org/abs/2203.12318
Zhang, Y., Liao, Q. V., & Bellamy, R. K. E. (2020). Effect of Confidence and Explanation on Accuracy and Trust Calibration in AI-Assisted Decision Making. arXiv. https://arxiv.org/abs/2001.02114
Tavakoli, L., & Zamani, H. (2025). Reliable Annotations with Less Effort: Evaluating LLM-Human Collaboration in Search Clarifications. arXiv. https://arxiv.org/abs/2507.00543
Alex Perry is the founder of Janus Labs and creator of the Janus Protocol. With 15 years leading product and technology delivery, including executive-level oversight of regulatory programs, privacy governance, and multi-squad agile transformation, he specialises in governance infrastructure for human-AI co-reasoning in regulated environments. This article is part of the AI Adoption Science series.
Janus Labs is an independent AI governance research initiative. Learn more at januslabs.com.au.