Computed Evidence over Asserted Truth
This page explains why Kappa Graph models knowledge as computed evidence rather than asserted truth, and why bi-temporal "valid time" is deliberately absent from the design.
The core inversion
Most knowledge systems assert truth and then maintain it. Kappa Graph computes evidence and never asserts. That single inversion explains most of what makes it different — including a few things that look, at first glance, like missing features.
A conventional knowledge graph stores facts: "X is true." When a contradicting fact arrives, it has to decide — usually the new one wins, the old one is retired. The graph is always a snapshot of the current best-known truth.
Kappa Graph stores no truths. A concept carries a grounding score — a continuous measurement computed from the balance of supporting versus contradicting evidence. A claim backed by forty-seven sources with twelve contradictions scores differently from one with a single unchallenged mention. When sources disagree, both sides stay. Nothing is retired for being "wrong"; its standing simply reflects the evidence pulling on it.
Why there is no "valid time"
The clearest place this shows up is temporal modeling.
Bi-temporal systems (Zep/Graphiti is the well-known one) track two clocks: transaction time — when the system learned a fact — and valid time — when the fact was true in the world. The second clock is the interesting one, and it is exactly what Kappa Graph refuses.
Valid time only means something if a fact has a truth value that holds over a real-world interval and can later become false. That is a correspondence theory of truth baked into the schema: the graph is supposed to mirror reality and track when reality changed. The operational tell is how such a system resolves a contradiction — it stamps the old fact with a valid_to, declaring when it stopped being true. Contradiction is settled by supersession.
Kappa Graph keeps the transaction-time clock — that is precisely what the ingestion epoch is (created_at_epoch, last_seen_epoch): when evidence first arrived and when it was last reinforced. But it has no valid-time clock, and can't have one, because to write valid_to you must first assert that the claim was true — you must pick a winner. That is the one move the epistemic model is built to never make.
The narrower temporal axis is not a gap. It is the same decision viewed from a different angle: we do not encode truth, so we cannot encode when truth changed. What we can encode — and do — is when evidence accrued.
Decisions are views, not writes
The deeper principle underneath both points is non-destructiveness.
Every judgment in Kappa Graph is a derived view, not a stored fact. Grounding is computed, not written. Contradictions are preserved, not resolved. Ontologies are annealed — continuously reorganized by a background worker — rather than fixed at ingest. Because the evidence is never discarded, any judgment can be recomputed when the data grows, the methods improve, or the compute gets cheaper. The graph is an accumulating substrate of evidence; the "answers" are functions over it that you can re-run.
Non-destructiveness matters when a judgment turns out to be wrong. A bad invalidation or premature supersession that bakes into storage loses the overruled evidence permanently — you cannot recompute your way out of data you deleted. Kappa Graph preserves the evidence, so any past judgment can be revisited.
This is also why the system keeps the original image beside its prose description, why it preserves both sides of a disagreement, and why "don't reset the graph" is a working principle rather than a slogan: the messy, contradictory, accumulated state is the asset.
The bitter lesson, and the training wheels
Rich Sutton's "bitter lesson" is that general methods which scale with computation beat methods that bake in human knowledge — every time, eventually, by a large margin. Hand-built domain structure gives satisfying early wins and then plateaus.
Bi-temporal truth-maintenance is, in this light, a human-knowledge layer: someone has to decide what is true and when it stopped being true, and keep those intervals correct as the world contradicts itself. That is the kind of structure the bitter lesson comes for.
Kappa Graph's bet is the other direction: bake in no decisions; compute over accumulated evidence. But honesty requires naming where it isn't there yet. There are hand-tuned constants today — a merge threshold, grounding weights, a closed vocabulary of annealing actions. Those are training wheels, not the philosophy. They exist to get the system working before it can be trusted to find its own settings. The direction of travel is to let computation take them over.
There is already first evidence that it can. Left to run on its own, the annealing worker reorganized raw extracted concepts into coherent emergent ontologies — splitting and naming themes nobody specified, without a human drawing the schema. That is a small, concrete instance of the scaffolding coming off: the structure emerged from computation over evidence, not from a designer's ontology.
The loop is built to close further. Annealing already governs its promotion, demotion, split, and merge thresholds with tunable pressure curves rather than fixed cutoffs (ADR-200; the Bezier "ecological pressure" mechanism). Every annealing decision is recorded against the epoch event log (ADR-203), so the pool of decisions deepens with each epoch. The active frontier is to feed that accumulating decision history back into the curves — letting the system's own track record, not a designer's initial guess, set how aggressively it reorganizes. A constant becomes a function of what the graph has learned about itself. The accumulated epoch signal is the teacher; the curves are what it adjusts.
The honest limits
This is a stance, not a victory lap. Three things to keep clear-eyed about:
- It is not pure computation yet. The measurement apparatus — how grounding is scored, how concepts merge, what relationship types exist — is still human-designed. The bitter lesson applies to those too. The defensible claim is only that the human knowledge has moved up a level: we design the apparatus that measures evidence, not the truths themselves. That is a better place to stand, because the apparatus is general and the truths are brittle — but it is not the asymptote.
- Scaling computation scales bias. "Massively compute it" is only the right side of the bitter lesson if the thing being computed is a general objective. A brittle, hand-tuned grounding function run at scale just industrializes its own bias. The non-destructiveness is the insurance: because the evidence is preserved, a better objective can be swapped in and the answers re-derived. A system that expired its losing evidence has no such recourse.
- Some questions genuinely want valid time. "Who was CEO in 2013?" is cleanly answered by validity intervals. Kappa Graph answers it differently — the dates live in content and relationships, and grounding reflects evidence — which is the right primitive for measuring how well-supported a claim is and the wrong one for maintaining a single current truth. Different question; deliberate choice.
Why it matters
The payoff is conviviality (in Illich's sense): a tool that shows its work. Kappa Graph hands you the evidence — the supporting sources, the contradicting ones, the grounding score, the provenance — and lets you judge. The design equips judgment rather than replacing it.
"There is no truth, only computed evidence" is not relativism. It is a refusal to launder a model's confident guess into a stored fact. Keep all the evidence, compute the standing, preserve the disagreement, and let the structure emerge. The rest is engineering toward the day the training wheels come off.
Related: the annealing/self-organizing ontology design (ADR-200) and its tunable pressure curves (issue #249); the grounding and epistemic-ledger mechanisms (ADR-206); the epoch event log (ADR-203) and the accumulating-signal feedback work (issues #386, #387); the modality-embedding stance that prose is the cross-modal bridge (ADR-803). For the grounding score in detail, see Grounding and Epistemic Confidence.