Hidden Reservoirs of Dissatisfaction: Testing a Store-and-Release Model of Self-Report Suppression

People often do not say what they think when they think it. They hold back complaints to be polite and voice them later, sometimes all at once. Most models of self-report bias treat each report as a noisy, shifted reading of the person's current state. On that view, what someone says at time $t$ depends only on how they feel at time $t$, plus a stable personal tendency to over- or under-report. This proposal tests a different structure, borrowed from engine physics. Unvoiced dissatisfaction is stored. Part of it fades. The rest is released abruptly when a trigger pushes the stored amount past a threshold. If that is right, an outward complaint should be predicted by how much unexpressed dissatisfaction has built up over earlier turns, not just by the current one. Once a person complains, the store should empty, and the chance of a further complaint should drop.

The physical template is liquid fuel wetting a cold engine wall. When the wall is cold, injected fuel does not evaporate. It sticks to the surface as a film and goes unseen by the bulk signal, the cylinder pressure. Later, compression heats the chamber, and the stored film vaporizes and ignites at once. In a heavy-duty diesel simulation, lowering the initial charge temperature from 380 K to 307 K increased wall-adhering fuel by 73%. The share of that film that later evaporated rose from 20.7% to 70.7%, and peak pressure rose from 8.1 to 11.6 MPa at 328 K (Thermal Load Analysis of Piston Damaged by Wall-Wetting Combustion in a Heavy-Duty Diesel Engine). Liquid ammonia shows a related pattern. It absorbs so much heat while evaporating that, without a nearby pilot flame, it delays or quenches ignition (Study on spray combustion characteristics of liquid ammonia/dimethyl ether dual fuel based on different injection strategies). We do not claim that human minds obey thermodynamics. We take one falsifiable structural idea from this literature, "hidden storage, leak, threshold-triggered release." We then ask whether it describes how self-reports diverge from private states better than memoryless models or stable-trait models do.

Core Claim

The claim is specific and testable on existing data. ThoughtTrace collected private pre-prompt and post-response thoughts alongside outward messages in human–chatbot conversations. It found 2.2 times more dissatisfaction in private thoughts than in the messages themselves (ThoughtTrace: Understanding User Thoughts in Real-World LLM Interactions). That paper measured the size of the gap. It did not model how the gap evolves over a conversation.

We propose that the gap behaves like a reservoir. Let $d_t \ge 0$ be the private dissatisfaction a user logs after the model's response at turn $t$. Let $S_t$ be the stored, unexpressed amount. Let $y_t \in \{0,1\}$ indicate whether the user's next outward message voices dissatisfaction. The model is

$$S_t = (1-\lambda)\,S_{t-1}(1-\rho\,y_{t-1}) + d_t, \qquad P(y_t=1) = \sigma\big(\beta\,(S_t - \tau_u)\big),$$

The symbols mean the following. $\lambda \in [0,1]$ is the fraction of stored dissatisfaction that fades each turn, the analogue of film evaporation. $\rho \in [0,1]$ is the fraction released when the user complains outwardly. $\tau_u$ is a per-user expression threshold, the analogue of wall temperature: a "colder," more polite user stores more before releasing. $\beta$ sets how sharp the release is, and $\sigma$ is the logistic function.

The model makes three predictions that a memoryless model does not. First, conditional on the current $d_t$, earlier unexpressed dissatisfaction raises $P(y_t=1)$. Second, after an outward complaint, expression probability drops below what the user's average rate would predict. Third, users with high thresholds complain less often, but their complaints are more intense. This mirrors the cold-wall result that more stored film produces a larger pressure spike.

Research Program

The first stage builds a turn-level corpus from ThoughtTrace's 2,206 conversations, 17,792 turns, and 10,556 thought annotations. Private post-response reactions are converted into a graded dissatisfaction score $d_t$. Outward messages are labeled for expressed dissatisfaction, $y_t$, and its intensity. Both labels come from an LLM judge, audited against a few hundred human labels. We also record conversation abandonment, meaning the user ends the chat right after a turn. Abandonment acts as the "flash release" endpoint.

The second stage fits the reservoir model by maximum likelihood, with hierarchical priors on $\tau_u$. It is then compared with the alternatives described below on held-out users. Splitting by user, not by turn, keeps memorization of a person's style from passing for dynamics.

The third stage handles the deployment problem. In practice, private thoughts are not available, and ThoughtTrace notes that asking for them changes how people behave. We therefore train an encoder that infers $S_t$ from outward text and timing alone, with ThoughtTrace thoughts as supervision during training. We test whether this inferred reservoir predicts outward complaints and abandonment better than outward-only features. This is the practical payoff. A chatbot could detect pent-up dissatisfaction before a user complains or leaves.

The fourth stage is a smaller replication in a domain where hidden and voiced states are measured differently. The Intern Health Study micro-randomized trial pairs daily self-reported mood with Fitbit sleep and step counts in 1,565 interns (Assessing Real-Time Moderation for Developing Adaptive Mobile Health Interventions for Medical Interns: Micro-Randomized Trial). There, we treat accumulated passive sleep deficit as the hidden store. We ask whether it predicts abrupt drops in reported mood beyond the effect of the previous night's sleep. The trial also found that prompts helped low-mood interns and hurt high-mood ones. We test whether the estimated reservoir explains that reversal better than the trial's fixed cut-offs: mood below 7, steps below 10,614, sleep below 6.9 hours. This replication is secondary. Fitbit data are often missing, which the trial itself flags, and missingness will limit the analysis.

Baselines

The naive reference point uses outward text alone. A classifier predicts the next outward complaint from the visible conversation, the way a deployed system without thought data would. Every other method must beat it on held-out users for the reservoir idea to matter at all.

The memoryless baseline is a logistic regression of $y_t$ on the current $d_t$ plus a per-user random intercept. This is the standard psychometric picture: a current state plus a stable response style. It is the model the core claim most directly contradicts. It receives the same inputs and the same per-user effects as the reservoir model.

The strongest existing method is ThoughtTrace's own approach. A language model is given the conversation together with the full history of private thoughts at inference time, and it predicts the next message. ThoughtTrace reports a 41.7% relative improvement in next-message prediction from adding thoughts. That result is for full next-message prediction, not binary complaint prediction. For predicting outward dissatisfaction or abandonment from thought histories, we found no reported result on Lacuna. We will therefore re-run this baseline ourselves. We use the same base model family as in our stage-three encoder, the same user split, and a matched prompt-tuning budget. We also run a GRU sequence model on the thought sequence, a recurrent network that learns its own memory, with hyperparameter search equal in compute to the reservoir fit.

The ablations target the key ingredient. Setting $\lambda=0$ gives a pure cumulative sum with no fading. Setting $\rho=0$ removes the release step. Fixing $\tau_u$ across users removes the threshold heterogeneity. The reservoir model "wins" only under two conditions. It must beat the memoryless model by a significant margin in held-out log-likelihood. It must also match the GRU and the LLM-with-thoughts baseline to within their seed variance, while using a handful of interpretable parameters. Its distinctive predictions, the release dip and the intensity–threshold link, must also hold up.

Why not just…?

Why not just use a generic sequence model and skip the physics? This alternative is a real threat. A GRU or an LLM can in principle learn fading and release on its own, and it may predict as well or better. The contribution is therefore not raw accuracy. It is a claim about structure that a black box cannot test: past suppression matters, and release resets it. We settle this in two ways. First, we check whether adding the fitted $S_t$ as a feature improves the GRU. Second, we probe whether the GRU's hidden state tracks $S_t$ and dips after complaints. If the GRU wins clearly and shows no reservoir-like dynamics, we report that the store-and-release hypothesis is unsupported.

Why not just model stable response styles, as item response theory or structural equation models do? Cross-sectional survey models, such as the chain-mediation analysis of nurses' self-rated performance (The effects of work readiness, organizational justice and professional identity on the work performance of new nurses: a cross-sectional survey), capture who over-reports. They cannot capture when people release what they have held back. The two views make different within-user predictions. A trait model predicts no dip after a complaint, while the reservoir predicts one. The release-dip test discriminates between them directly.

Why not just collect private thoughts or passive sensor data and skip inference? ThoughtTrace notes that prompting people to log thoughts changes their behavior. The intern trial notes that wearable data go missing when people stop wearing devices. Passive traces alone also leave much of the variance unexplained. Music-tag features classified depression risk at 66.4% accuracy (Tag2Risk: Harnessing Social Music Tags for Characterizing Depression Risk). Stage three tests whether a reservoir inferred from outward signals alone recovers most of what thought logging provides. We measure this as the share of the gap between the outward-only baseline and the thoughts-at-inference model that the inferred reservoir closes.

Why not import the full transport equations, such as Stanton-number analogies or activity coefficients, as the research file's candidate gaps suggest? Those equations describe molecular mass transfer. Their quantities have no meaningful human counterpart, and the wall-film modeling work (Development and implementation of multicomponent liquid wall film evaporation model for internal combustion engine applications) validates them only for physical films. We deliberately borrow only the qualitative storage–leak–threshold structure, which is simple enough to be falsified.

Proposed Experiments

The first decisive check is the history effect. On held-out ThoughtTrace users, we test whether the coefficient on earlier unexpressed dissatisfaction is positive and significant after controlling for current $d_t$ and per-user intercepts. A null result here falsifies the core claim.

The second check is the release dip. We test whether expression probability on the turn after an outward complaint falls below the user's predicted baseline. A dip supports the reservoir model; a trait model predicts no dip.

The third check is the cold-wall analogue. Across users, we test whether a higher fitted threshold $\tau_u$ goes with fewer but more intense complaints and with a higher rate of abrupt abandonment.

The fourth check compares fit against the GRU and the LLM-with-thoughts baseline, as specified under Baselines. The fifth reports how much predictive power survives when the thought-supervised encoder must infer $S_t$ from outward text alone.

Finally, the Intern Health Study replication tests whether an accumulated sleep-deficit reservoir predicts sharp drops in mood reports. It also tests whether the reservoir explains why the same prompts helped some interns and hurt others, beyond the trial's fixed thresholds. We report the replication with explicit handling of missing wearable data, through multiple imputation and a complete-case sensitivity analysis.