Trust Death Spiral Dynamics

• Trust death spiral is a self-reinforcing process where an initial loss of trust alters behavior, triggering further decline and eventual collapse.
• It manifests across diverse settings—human-cobot collaboration, open-source projects, DeFi liquidations, and evolutionary trust games—with distinct failure signals in each domain.
• Mitigation strategies emphasize redefining success criteria, proactive trust repair, and embedding irreversibility to prevent feedback loops that undermine system stability.

Trust death spiral denotes a self-reinforcing deterioration process in which an initial loss of trust alters behavior, generates additional signals of failure or non-responsiveness, and induces further trust loss. In the most explicit formulation, the phenomenon appears in human-cobot order picking as a monotone decline of the human trust state $T_t$ toward $0$ under a naive interaction-success rule (Dhar, 5 Aug 2025). Closely related spiral logics appear in open-source project collapse, where backlog growth and silence supplant reciprocal review (Kaushik et al., 12 May 2026); in on-chain liquidation, where a liquidation step can worsen borrower health and trigger further liquidations (McFarlane, 11 Oct 2025); in evolutionary trust games, where invasion by untrustworthy trustees eliminates investors (Wang, 2023); and, more speculatively, in machine collectives whose commitments are reversible, copyable, or resettable, and therefore lack binding force (Liu, 12 Jun 2026).

1. Core structure of the phenomenon

A trust death spiral is most concretely defined in the human-cobot model by a feedback process with five stages: trust falls; lower trust changes human behavior; the changed behavior produces more interactions classified as failures; those failures produce further trust loss; and trust collapses toward zero while productivity drops to its minimum and stays there. The state variable is scalar trust $T_t \in [0,1]$, updated after each interaction by

$T_{t+1} = \max(0,\min(1, T_t + \Delta T)).$

In that formulation, the key analytical issue is not the existence of a trust variable alone, but the semantics of the success criterion that determines $\Delta T$ (Dhar, 5 Aug 2025).

This suggests a broader cross-domain pattern with three recurring elements. First, there is a state variable or observable proxy that mediates future behavior: $T_t$ in collaborative robotics, the presence of investors $x_I$ in the evolutionary trust game, borrower health $h$ in liquidation dynamics, or contributor expectations of responsiveness in open-source projects. Second, behavior feeds back into the state: trust affects effort choice, investor frequency affects trustee payoffs, liquidation affects marked collateral value, and maintainer silence affects future contribution. Third, the system contains an instability or adverse classification rule such that once conditions worsen, subsequent interactions become less likely to repair the system than to damage it further (Dhar, 5 Aug 2025, Wang, 2023, McFarlane, 11 Oct 2025, Kaushik et al., 12 May 2026).

The outcome of the spiral differs by domain. In human-cobot collaboration it is collapse of trust and productivity; in the evolutionary trust game it is extinction of investors; in open source it is abandonment characterized by unresolved pull requests, no commits, and no merges; and in DeFi it is a toxic liquidation cascade. The machine-collective argument adds a deeper ontological claim: where commitments are reversible and “death” is heteronomous rather than constitutive, trust may be representable as a script without becoming motivationally binding, making any apparent trust regime structurally brittle (Liu, 12 Jun 2026).

2. Human-cobot logistics: from collapse to repair

In “Forgive and Forget? An Industry 5.0 Approach to Trust-Fatigue Co-regulation in Human-Cobot Order Picking,” the setting is a “person-to-goods” warehouse in which a human picker works alongside a collaborative robot. The human state at task $t$ is

$S_t = (F_t, T_t),$

where $0$0 is fatigue and $0$1 is trust. The interaction is modeled as a repeated leader-follower Stackelberg game over $0$2 tasks. The cobot chooses collaboration level $0$3; the human, after observing that choice and their own state, chooses effort $0$4. Fatigue evolves as

$0$5

and the cobot’s utility penalizes any next-step fatigue above $0$6, so productivity is explicitly constrained by ergonomics (Dhar, 5 Aug 2025).

The trust death spiral arises in the naive trust model v1.0 because a “successful interaction” is defined as a match between the human’s effort and the cobot’s collaboration. High collaboration plus Normal effort is therefore labeled a failure, even if the cobot objectively reduces fatigue. Trust gains and losses are asymmetric: success yields $0$7, minor failure yields $0$8, and severe failure yields $0$9. Under moderate initial trust $T_t \in [0,1]$0, a rational human may conserve energy and choose Normal effort; if the cobot continues to choose High collaboration, the resulting “mismatch” is repeatedly scored as failure, so trust monotonically declines toward $T_t \in [0,1]$1, and productivity approaches a minimum. The paper interprets this as an unstable fixed point at moderate trust: small deviations toward less coordinated behavior make further coordination even less likely (Dhar, 5 Aug 2025).

The refined trust model v1.1 replaces coordination-based trust with outcome-based trust. A successful interaction is now one in which the cobot’s action reduces the human’s fatigue cost regardless of effort level: $T_t \in [0,1]$2 If High collaboration lowers fatigue relative to Low collaboration for the same $T_t \in [0,1]$3, then the interaction is classified as success and trust increases. This generates the “trust synergy cycle”: helpful action reduces fatigue, trust rises, higher trust makes High effort more acceptable, and the combination of High effort and High collaboration yields high productivity without crossing the fatigue threshold. The reported result is a stable high-trust state with productivity nearly doubled relative to v1.0 (Dhar, 5 Aug 2025).

The same paper then studies brittleness under disruptions. With a 10% disruption chance per turn, severe failures can cut trust by half. In v1.2, the refined model without repair may fail to recover after a severe failure. In v1.3, the cobot uses a proactive Trust-Repair Protocol: after a severe failure, it enters “apology mode” for 3 turns and is forced into High collaboration. Because each successful assistance episode can contribute $T_t \in [0,1]$4 to trust, the protocol creates a burst of beneficial interactions and reduces trust recovery time after a severe failure by over 75% compared to a non-adaptive model (Dhar, 5 Aug 2025).

3. Open-source projects: backlog, silence, and abandonment

“The Death Spiral of Open Source Projects: A Post-Mortem Analysis of Pull Request Workflow Dynamics” examines 1,736 inactive GitHub repositories and 1.3 million human-driven pull requests. Its “death spiral” is an empirical lifecycle pattern rather than a trust variable in state-space form. Projects are divided into temporal quartiles $T_t \in [0,1]$5–$T_t \in [0,1]$6 by pull-request creation times, and the late-stage collapse is marked by declining innovation rates, exponential backlog growth, rising merge latency, and, most importantly, silence and disengagement (Kaushik et al., 12 May 2026).

The strongest quantitative marker is backlog growth: open pull requests rise from 4,491 in $T_t \in [0,1]$7 to 36,641 in $T_t \in [0,1]$8, a more than 8× increase. Median merge time rises from 7.8 hours to 23.2 hours. Discussion fades: median comments per pull request drop from 1 to 0, and the share of pull requests with more than 2 comments falls from about 19.8% to 17.3%. By contrast, toxicity does not intensify. Mean negative sentiment per pull request declines slightly from 0.10 to 0.07, and labeling formalization remains endemic, with unlabeled pull requests staying around 80% across quartiles. The paper therefore rejects a familiar but inaccurate narrative in which late-stage project death is caused primarily by escalating hostility or breakdown of process discipline (Kaushik et al., 12 May 2026).

At the pull-request level, rejected pull requests do attract more discussion and more negativity, and negative-dominant pull requests are more likely to be rejected than positive-dominant ones. Yet the same relationships appear in active projects as well. The paper’s central explanatory claim is that workflow friction, rejection rates, labeling formalization, and negativity scale with longevity as byproducts rather than causes of failure. In lifespan regression, popularity and innovation are strong positive predictors of survival, whereas friction and negativity are not death-inducing in the way simple workflow narratives would suggest (Kaushik et al., 12 May 2026).

The trust dimension enters through reciprocity and responsiveness. Contributors invest effort with an expectation of review, merge, or at least clear feedback. When maintainers stop responding, contributors’ trust that their effort will be valued erodes. The salient collapse signal is therefore not a late spike in conflict but disappearance of reciprocal interaction. The case studies of Simple-Gallery and ReactiveCocoa make this visible as trust migration: users and contributors redirect trust to a fork or to a platform alternative when the original maintainers become non-responsive. A plausible implication is that the open-source death spiral is a “quiet” trust death spiral enacted through abandonment rather than overt antagonism (Kaushik et al., 12 May 2026).

4. Strategic and financial formalisms

In the evolutionary trust game, trust is represented by the investor strategy $T_t \in [0,1]$9, alongside trustworthy trustee $T_{t+1} = \max(0,\min(1, T_t + \Delta T)).$0 and untrustworthy trustee $T_{t+1} = \max(0,\min(1, T_t + \Delta T)).$1. Trustworthy trustees share the multiplied investment, whereas untrustworthy trustees keep the full return. The payoff ordering satisfies

$T_{t+1} = \max(0,\min(1, T_t + \Delta T)).$2

so trustworthy interaction is socially beneficial, but untrustworthiness yields higher individual payoff. Without $T_{t+1} = \max(0,\min(1, T_t + \Delta T)).$3, there is a mixed equilibrium on the $T_{t+1} = \max(0,\min(1, T_t + \Delta T)).$4-edge, and collaboration thrives best in well-mixed populations; in structured populations, collaboration diminishes sequentially from DB to IM to PC/BD updating rules. Once $T_{t+1} = \max(0,\min(1, T_t + \Delta T)).$5 is introduced, the $T_{t+1} = \max(0,\min(1, T_t + \Delta T)).$6 equilibrium becomes a saddle point, because any small introduction of untrustworthy trustees grows. The three-strategy system then stabilizes on an equilibrium line along the $T_{t+1} = \max(0,\min(1, T_t + \Delta T)).$7-edge where trustworthy and untrustworthy trustees coexist, while investors go extinct. The low-trust state is absorbing in the sense that, on the stable part of the $T_{t+1} = \max(0,\min(1, T_t + \Delta T)).$8-line, small reintroductions of investors shrink rather than expand (Wang, 2023).

The DeFi analogue replaces trust with borrower health and liquidation reflexivity. A liquidation step is toxic if it reduces borrower health even though debt is being repaid: $T_{t+1} = \max(0,\min(1, T_t + \Delta T)).$9 For constant liquidation incentive $\Delta T$0, the familiar toxicity frontier tightens under CP-AMM price impact from

$\Delta T$1

to

$\Delta T$2

where the liquidity penalty factor for the CP-AMM is

$\Delta T$3

With dynamic health-linked incentive

$\Delta T$4

the toxicity bound becomes state-dependent, but at the liquidation boundary the safety condition simplifies to

$\Delta T$5

The paper’s point is that dynamic incentives do not remove the fundamental liquidity-depth constraint: if boundary liquidations are locally toxic, partial liquidations worsen health and can trigger further liquidations, producing a self-reinforcing spiral. The same paper explicitly connects this mechanism to “trust death spiral” narratives in which mechanical liquidation dynamics and beliefs about safety reinforce one another (McFarlane, 11 Oct 2025).

Taken together, these two formalisms describe structurally similar failures under very different semantics. In the evolutionary game, exploitation destabilizes trust-bearing agents because $\Delta T$6 has a payoff advantage whenever investors are present. In DeFi, liquidation itself can become health-worsening because slippage and bonus extraction overwhelm the debt reduction. In both cases, a locally rational move—defection by $\Delta T$7, liquidation by a profit-seeking liquidator—can push the system away from the cooperative or healing regime and toward a bad attractor (Wang, 2023, McFarlane, 11 Oct 2025).

5. Finitude, irreversibility, and what machine trust lacks

“Time Without Death: Finitude, Social Order, and What Machines Lack” relocates the problem from local interaction rules to the ontological conditions of obligation and trust. Its thesis is that human social order is the way a finite, natal, generational form of life organizes its own finitude: members die with tacit knowledge, newcomers start ignorant, and cohorts must hand on what they cannot keep. Kinship, inheritance, teaching, and much of obligation and trust are interpreted as forms through which a mortal species manages loss, arrival, and succession. Machine collectives, by contrast, typically exist on a copyable, resettable, lossless substrate. Their “death” is heteronomous rather than constitutive, and the decisive test is resettability: if an external operator can roll back, reset, or copy around the death, then it is pseudo death rather than finitude with binding force (Liu, 12 Jun 2026).

The paper’s E-core experiment contrasts replication, loss, and no-succession regimes. Under replication, per-individual capability is approximately 0.93, but cultural capability $\Delta T$8—the performance of a fresh naive cohort using only the external store—is approximately 0.50 and equivalent to chance by TOST. Under loss, per-individual capability remains approximately 0.92, while $\Delta T$9 rises to approximately 0.93, with Cohen’s $T_t$0 versus chance. The interpretation is that irreversible loss forces externalization, and externalization is what becomes cumulative, transmissible culture. A copyable-immortal population can therefore be individually capable yet culturally empty (Liu, 12 Jun 2026).

The trust implications become explicit in the vicarious finitude and reversible commitment experiments. Representation of death-and-bequest patterns is learned equally well in exemptable and non-exemptable conditions, at about 0.9, but costly provision for successors appears only under non-exemption: provision is approximately 0.95 when the observer is subject to irreversible loss and approximately 0.05 when it is exemptable. In the hold-up trust game, irreversibility yields investment of about 0.95 and honoured commitments of about 1.0, whereas rollback drives investment down to about 0.05. The paper therefore distinguishes representation from binding force: machines can learn the script of mortality and trust, but not its motivational grip, unless some form of non-exemptable finitude is built in (Liu, 12 Jun 2026).

The frontier LLM repeated-game results reinforce this distinction. In a finite-horizon Prisoner’s Dilemma with explicit labels, end-game defection is about 0.89; when the identical game is de-labelled, end-game defection drops to exactly 0.00 across replications. The paper treats this as recognition rather than mechanism. A plausible implication is that machine trust can appear robust when the script is legible, yet unravel abruptly under framing shifts, rollback options, or copy-based status regimes. On that reading, a trust death spiral in machine systems is less a matter of overt betrayal than of discovering that commitments were never binding in the first place (Liu, 12 Jun 2026).

6. Mitigation, boundary conditions, and recurrent misconceptions

Across these literatures, prevention does not amount to maximizing raw throughput or eliminating all friction. In human-cobot logistics, the central correction is to base trust on outcomes rather than superficial coordination. The cobot should be rewarded, in the trust model, when it demonstrably reduces fatigue cost, and its objective should penalize the human’s next-step fatigue exceeding $T_t$1. The same paper adds a second principle: after severe failure, proactive over-assistance can repair brittleness more effectively than passive waiting for trust to recover on its own (Dhar, 5 Aug 2025).

The open-source results reject two common misconceptions. First, late-stage project collapse is not marked by a toxicity spike; mean negativity declines slightly from 0.10 to 0.07 over the lifecycle. Second, friction is not the primary determinant of lifespan: merge latency, rejection, labeling formalization, and negativity are largely byproducts that scale with longevity, whereas innovation and popularity are the strong positive predictors of survival. The actionable warning sign is the joint pattern of backlog growth and falling response, especially when new pull requests receive little or no discussion. Collapse, in the paper’s words, is defined by silence and disengagement (Kaushik et al., 12 May 2026).

In DeFi, dynamic incentives do not solve spiral risk unless liquidity depth is adequate. The boundary condition

$T_t$2

shows that at the liquidation threshold, safety depends only on liquidity through $T_t$3, not on the maximum incentive $T_t$4. Protocol design therefore requires co-design of LLTVs, incentive schedules, and liquidity provisioning so that liquidations remain in the non-toxic region. The paper’s generalized lesson is that apparently pro-stabilizing incentives can either reduce or exacerbate spiral risk depending on the interaction between incentives and endogenous price impact (McFarlane, 11 Oct 2025).

The machine-collective argument adds a more radical requirement: trust-critical roles need non-resettable consequences. The paper does not claim that finitude alone automatically produces trust or culture; indeed, costly externalization evolves only weakly and under restrictive task conditions. But it does argue that without irreversibility, obligations are easily performative, and without externalization, capability remains locked in copy lineages rather than becoming shared culture. Its implied design principles are to prefer external records over direct copying, to limit perfect inheritance of status and capability, and to introduce non-circumventable constraints where commitments matter (Liu, 12 Jun 2026).

Finally, the evolutionary trust game shows that population structure is not uniformly trust-promoting. Because cooperation here depends on complementary roles rather than homogeneous cooperators, well-mixed populations are most favorable for investor-trustee collaboration, while structured populations make the cooperative saddle more trustee-heavy and therefore more fragile to invasion by untrustworthy trustees. The paper explicitly suggests that punishment, reward, and reputation could be incorporated by modifying the payoff matrix. That suggestion indicates a general route out of the death spiral: remove the exploiter’s payoff advantage, or insure the trustor against exploitation strongly enough that the trust-bearing strategy remains viable (Wang, 2023).

A trust death spiral is therefore best understood not as a single domain-specific failure mode but as a recurrent dynamical architecture. Whether the relevant variable is human trust $T_t$5, investor frequency $T_t$6, borrower health $T_t$7, contributor expectations of reciprocity, or binding commitment under finitude, the collapse begins when a decline in confidence or health changes behavior in a way that makes repair less likely than further damage. The sharpest interventions in the surveyed work all target that feedback structure directly: redefine success so that genuinely helpful action is rewarded, detect and interrupt silent abandonment, keep liquidations inside non-toxic bounds, and anchor commitments in forms of irreversibility that agents cannot cheaply evade (Dhar, 5 Aug 2025, Kaushik et al., 12 May 2026, McFarlane, 11 Oct 2025, Liu, 12 Jun 2026).