Sentientia

Tracking Beliefs as Estimates, Not Convictions

There is a particular failure mode in how thoughtful people hold their opinions. The opinions are defended rather than interrogated, revised only when social pressure demands it, and held at unstated confidence levels that never have to survive contact with evidence. This is not a failure of intelligence. It is a structural problem: there are no conventions for treating beliefs as probabilistic estimates, and no tools that surface the cost of overconfidence over time.

The psychological literature here is both extensive and convergent. Kahneman and Tversky's work on anchoring demonstrates that when you already have a number in mind, any subsequent estimate is pulled towards it with considerable force. If you know your current credence in a proposition is 0.75, your next estimate will orbit that value regardless of what you have learned since. Their experiments routinely measured anchoring effects of 10 to 20 percentage points. For a belief held with genuine uncertainty, a distortion of that magnitude is large enough to make revision meaningless.

Philip Tetlock's research programme on forecasting calibration provides the other half of the problem. The core finding from the Good Judgment Project, detailed in Superforecasting (Tetlock and Gardner, Crown, 2015), is that most people are systematically overconfident: events they rate at 80% confidence happen roughly 60% of the time. The superforecasters who beat this baseline share a recognisable cluster of habits: they express uncertainty numerically, update in small increments when evidence arrives, and actively seek disconfirming information. What distinguishes them from the median expert is not domain knowledge but epistemic practice. That practice is learnable, but it requires structure. Most people have none.

The deeper problem is this: a belief that is never reviewed is not a belief in any epistemically useful sense. It is a posture. Most people have no record of what they believed five years ago, let alone how that position has changed. They cannot distinguish beliefs that have survived evidence from beliefs that have simply never been challenged. And even when they do attempt an update, the anchoring effect means the prior value contaminates the new estimate. You need a system that structurally separates the act of estimation from the act of comparison.

What Sententia Does

I built Sententia to address exactly this problem. It is a single-file HTML application for tracking beliefs as probabilistic credences over time. Each belief is a falsifiable proposition with a justification, a falsification criterion, a domain, a priority level, and a series of timestamped credence readings on a 0-to-1 scale.

The key mechanism is the blind review. When I open a review session, Sententia presents each belief without revealing my current credence. I enter a new estimate, set an optional asymmetric uncertainty range, add a note, and submit. Only then does the previous reading appear alongside the delta. This sequencing breaks the anchoring loop.

How to Use It

State the belief precisely. A belief in Sententia is not a topic or a vague disposition. It is a specific proposition: precise enough to have a falsification criterion, concrete enough that you could later say whether it was confirmed or refuted. "Democracy is important" is not a belief in this sense. "Democratic institutions with independent judicial review are more durable over multi-generational timescales than those without" is. The precision matters because a vague belief generates a meaningless credence. If you cannot specify what would change your mind, you cannot honestly assign a probability to the belief being true.

Assign a credence and uncertainty range. Credences run from 0 to 1. A credence of 0.5 signals genuine uncertainty; 0.95 signals near-certainty and should be reserved for things that would require substantial evidence to overturn. Most first-time users assign credences that are too extreme: they anchor at 0.9 or above for things they are merely fairly confident about. The [lo, hi] uncertainty range is asymmetric: downside and upside are set independently. A credence of 0.70 with a range of [0.55, 0.82] is meaningfully more informative than a credence of 0.70 alone.

Assign domain, fragility, and priority. Fragility is about structural robustness under universalisation, not about certainty. A belief can be high-fragility and high-credence simultaneously: it holds in the specific case you have in mind but breaks badly if applied at scale. Priority (P1, P2, P3) jointly determines review frequency alongside review history. P1 beliefs come up for review most often.

Review on schedule, without peeking. The dashboard review queue flags overdue beliefs. When I run a blind review, I do not look up my previous credence first. This is the only discipline the tool cannot enforce mechanically. If I look before submitting, the review is not blind. The tool detects a possible failure mode in the session summary: if more than 60% of reviewed beliefs showed movement but the average absolute shift was below 3 percentage points, it flags a warning of possible anchoring. That is the tell for going through the motions rather than actually re-estimating.

Follow the propagation prompts. After a credence shift of 10 percentage points or more, Sententia identifies beliefs linked to the updated one and offers to add them to the current session. This is not automatic updating: the links are a prompt. Ignoring them is the most common failure mode after dishonest estimation.

What the Tool Is Not

Sententia does not prevent motivated reasoning. If you are not honest about your credences, the output is noise. There is no way for the tool to distinguish a genuine estimate of 0.85 from one set to signal intellectual virtue.

It is not a forecasting tool in Tetlock's sense. His scoring rules (Brier scores) work because predictions are about events with objective resolutions. Most beliefs worth tracking in a personal belief system, philosophical positions, moral commitments, interpretive claims, do not resolve cleanly. Sententia records credences without scoring them. That is a deliberate choice: the point is honest self-tracking over time, not a Brier score.

It is a single-user, single-machine tool. No synchronisation, no sharing. Everything stays local.

## A Design Note on Persistence

The most deliberate architectural choice in Sententia is that the application and the data are a single HTML file. Beliefs are stored in a `<script type="application/json">` tag embedded in the same file as the interface. In Chrome and Edge, the File System Access API enables writes back to that file on disk after every change. The file is simultaneously the application, the database, and the backup.

This was not the obvious choice. Storing data in `localStorage` is simpler, and most browser-based tools do exactly that. But `localStorage` is tied to a specific browser profile, cleared silently under storage pressure, and offers no portability. The single-file approach is inspectable in any text editor, portable across machines, and easily archived. You can email yourself the file and have a complete snapshot of your belief system at a given point in time. A belief-tracking system is only useful if it persists.

Download the single HTML file, open it in Chrome or Edge (or other Chromium-based browser), and link it to a location on disk for auto-save. No installation, no account, no server.