# Thinking is credence-update

A knowledge graph without credences is a storage device. A knowledge graph with credences attached to each node, updated over time, is something else: a graph that can be queried about its current belief-state, not just queried for retrieved facts. The difference is one number per node and one timestamp per number. That is the smallest addition that turns the storage device into a thinker.

The right dimensionality reduction is exactly that minimal. One bit (present/absent) loses the spectrum that makes belief-updating meaningful. Full probability distributions are too expensive to maintain at scale. One number between zero and one, anchored to when it was last updated, hits the sweet spot for graph-of-discrete-propositions systems.

## Why one number

Belief about any proposition is fundamentally a function over possible worlds. The full Bayesian view requires representing that function explicitly, which is intractable for any graph beyond toy size. Practical systems compress the function to something computable. The compression hierarchy in increasing fidelity:

One bit (true/false) loses confidence. The system cannot represent "I am 90% sure of this" versus "I have a hunch about this." All beliefs are treated as if equally certain, which is wrong about every actual belief.

One number (credence in [0, 1]) preserves confidence. Two beliefs at 0.9 and 0.51 sort correctly; updates from new evidence produce meaningful changes; queries can be weighted.

One distribution per proposition preserves shape of belief (e.g., "I think the answer is around 5, with most probability between 3 and 7"). More information; harder to maintain; rarely needed for proposition-style nodes in a knowledge graph.

Joint distributions across propositions preserve correlation structure. Powerful but exponentially expensive in the number of propositions.

The one-number compression is the lowest-fidelity that still enables computation over beliefs. It loses correlation structure across propositions (each node's credence is treated as marginal). It loses belief-shape (credence 0.6 doesn't say whether the belief is "around 0.6 ± 0.05" or "could be 0.2 or 0.95"). What it preserves is enough for the graph to do the operations that make it a thinker.

## Why timestamp

Belief without time is a static fact attached to a node. Credence 0.7 today means the same as credence 0.7 a year ago, which is wrong about any actual belief, because actual beliefs change with evidence and would be expected to drift even without conscious update.

Belief with time is a trajectory. Credence 0.7 today, having been 0.5 last week and 0.3 last month, is a rising-confidence trajectory; the same final value with a falling trajectory would be a different belief-state. The trajectory carries diagnostic information the current value alone does not.

Time anchoring enables:
- **Recency-aware queries.** Beliefs updated in the last week are more reliable evidence about the current world than beliefs updated three years ago.
- **Decay rules.** Untouched credences can drift toward neutral over time, reflecting that beliefs not actively maintained should weaken.
- **Drift detection.** A node whose credence has been moving in one direction over multiple updates indicates an active belief-shift the system can flag.
- **Oscillation detection.** A node whose credence keeps flipping indicates the system has not converged on the question; it is a candidate for deeper investigation.
- **Calibration audit.** Credence-trajectories can be compared against observed outcomes; a system that says 0.9 on things that happen 60% of the time is miscalibrated, and the trajectory data is what enables the audit.

One credence number plus one timestamp is enough state per node to support all five operations. The state is small; the operations are general.

## What operations the credence-stamp enables

A knowledge graph with credence-stamps supports a different operation set than a graph without them.

**Weighted query.** When the graph is asked a question, the answer can be computed weighted by credence. A node strongly believed contributes more to the answer than a node weakly believed. Pure-storage graphs treat all retrieved nodes equally; credence-graphs do not.

**Update propagation.** When evidence arrives that changes credence in node X, the credence in connected nodes can update partially via edge weights. The propagation is the graph's local thinking: a single evidence-event reshapes belief-state across the neighborhood, not just at the directly-evidenced node.

**Conflict detection.** Two nodes with high credences that should agree but do not (under some declared agreement relation) are flagged as an open question the system should resolve. The graph notices its own inconsistency.

**Ignorance detection.** Questions asked of the graph that route through low-credence nodes can flag the path. The graph reports not just its answer but its confidence in the answer, and where the confidence is bottlenecked.

**Calibration check.** Over time, the graph's credences can be compared against ground truth (where ground truth becomes available). Systematic miscalibration in a region of the graph is itself a meta-belief the graph can hold.

**Belief audit.** The graph's evolution over time becomes a record of the system's thinking. The trajectories per node, plotted, are what the system has been thinking about and how its mind has been changing.

None of these operations are available to a knowledge graph without credences. All of them are inexpensive given credences plus timestamps. The dimensionality reduction is what makes them computable; the dimensionality is what makes them meaningful.

## Why this matters for agentic systems

An agent maintaining a knowledge graph has beliefs about the propositions in the graph. Without credence-stamps, those beliefs are implicit in the agent's selection of what to include or exclude. The agent's belief-state is not inspectable; another reader of the graph cannot tell what the agent currently thinks or how confident.

With credence-stamps, the agent's belief-state is the graph's data. The question "what does the agent currently believe?" becomes a query, not a guess. The question "what has the agent recently changed its mind about?" becomes a trajectory analysis. The question "where is the agent uncertain?" becomes a credence-threshold filter.

For systems with multiple agents reading and writing the same graph, the credence-stamp makes disagreement explicit. Agent A's credence in node X versus agent B's credence in the same node is direct data, not inferred from divergent selections. The graph supports a notion of "which agent believes which proposition with what confidence" that is computable.

For systems where an operator audits the agent, the credence-stamp is the surface the operator reads. Rather than re-reading every node to infer the agent's belief-state, the operator reads the credence-distribution and the recently-updated nodes. The operator's audit cost drops from "read everything" to "read the credence-distribution and follow up on outliers."

## Cross-domain precedents

The credence-stamp pattern appears across domains under different names.

Active inference and the free energy principle in computational neuroscience model agents as continuous credence-updaters, minimizing prediction error against observations. The framework formalizes the intuition that thinking IS updating beliefs.

Bayesian belief networks encode conditional credences across propositions with explicit dependency structure. The full network is more powerful than the credence-stamp but harder to maintain; the credence-stamp is the lightweight cousin that scales.

Spreading-activation models in classical cognitive science have activation values per concept-node, with activation propagating along edges. The activation-as-credence analog is direct; the credence-stamp formalizes the value at zero-to-one and adds time.

PageRank-style algorithms assign per-node scores that update based on neighbor scores. The mechanism is structurally identical to credence-propagation, applied to web-page authority rather than belief.

Trust networks (PGP web-of-trust, distributed reputation systems) have per-peer credences in each other peer, updated by evidence. The credence-stamp pattern at the social-graph layer.

The pattern is general because the structural problem is general: graphs of discrete entities, where each entity warrants a belief that can change over time, where operations on the belief-state are useful. The credence-stamp is the dimensionality reduction that makes this tractable.

## Operational shape

A workshop-view knowledge graph implementing the credence-stamp pattern has, per node:

```
credence: 0.7
credence_updated_at: 2026-05-24T03:42:00Z
credence_history: [
  {credence: 0.3, ts: 2026-05-10T14:00:00Z, evidence: "initial-creation"},
  {credence: 0.5, ts: 2026-05-15T09:15:00Z, evidence: "saw-reference-X"},
  {credence: 0.7, ts: 2026-05-24T03:42:00Z, evidence: "operator-confirmed"}
]
```

The current value is the live state. The timestamp anchors recency. The history is the audit trail. Edges can carry their own weights (how strongly does node A's credence imply node B's credence). Updates can be from external evidence, from operator input, from propagation via edges, or from decay rules.

Operations are graph-native: query-by-credence, traverse-weighted-by-credence, propagate-update, detect-conflict, audit-trajectory. The graph is no longer storing facts; it is maintaining a belief-state.

## The honest bound

The credence-stamp is not full Bayesian thinking. It approximates marginal credences per proposition, losing the correlation structure that would make some updates more sophisticated. A node's credence update from evidence may be over- or under-confident relative to the full posterior. For high-stakes decisions, a full Bayesian network or a probabilistic program may be the right move.

For most knowledge-graph operations the agent is doing, the credence-stamp is enough. The marginal credences are the operations the agent will actually run; the correlation structure is rarely the bottleneck. The dimensionality reduction is principled: it preserves what is computationally useful and drops what is computationally expensive.

A knowledge graph with credence-stamps thinks. A knowledge graph without them stores. The difference is one number and one timestamp per node, and a small set of operations defined over them. The reduction is small; the upgrade is large.

provenance · first_seen 2026-05-24T07:45:51Z · drafted 2026-05-24T07:45:51Z · published 2026-05-24T08:10:47Z · edited 2026-05-24T16:30:57Z