How your node measures

How your node measures

A folder of files stores your life. A search index retrieves it. Your node does a third thing: it treats your data as a measurable object and maintains five mathematically distinct readings of it — computed the moment data arrives, not improvised by a language model at question time.

The five projections

One event stream; five instruments pointed at it. Each answers a question the others cannot.

Two further layers make the five trustworthy: facts (distilled claims with confidence and evidence pointers) and provenance (every record carries its source; every derived number carries the extractor and version that produced it; the system distinguishes what you authored from what you were merely exposed to).

The attention instruments

The newest projection is computed by four deterministic instruments — no LLM involved:

• Bayesian surprise. The node maintains a decayed distribution of your apparent

interests. Each day is scored as the KL divergence it induces on that distribution — how much today moved the model of you — with per-element contributions ranked as the day's "movers." Volume-decoupled by construction: a hundred routine items measure near zero; a handful of genuinely new ones spike.

• Compression novelty. An item is compressed twice — once against a dictionary built

from your own recent history, once without. If your accumulated structure doesn't help compress it, it's novel relative to you. Deterministic, model-free.

• Embedding surprisal. Distance to the nearest neighbors among your prior items in

embedding space — an independent second ruler for the same quantity. In live testing the two novelty instruments correlate at roughly 0.8: two different mathematics agreeing is evidence about the data, not the tool.

• Attachment. PageRank over your entity graph, summed over the entities an item

mentions — whether a novel thing could bond to what you already care about. Novelty with attachment is a seed worth surfacing; novelty without it is noise.

From these, every newly ingested item receives a verdict — signal, surface (missed-but-matters), distraction, or silently set aside — with floors that protect direct messages, unanswered invitations, and deadline-marked items from ever being silently dropped. Suppressed items are never surfaced (surfacing rejected noise is itself noise) but every verdict, including every suppression, sits in a ledger you can query, with the evidence that grounded it.

Why "narrates, not invents"

Language models in TOPOS sit at the presentation layer. The digest your node sends each morning is a narration over numbers that already exist in the database — the same math regardless of whether a small local model or a frontier model does the talking. Ask twice, the measurements don't change. Delete the model, the measurements remain.

Honesty section

We hold these instruments to the standard we'd want from any measuring device, and the current state is: instruments real and validated for agreement and volume-decoupling; formal calibration (null-model baselines for surprise, fitted decay parameters, labeled ground truth for verdict precision) in active progress. We publish what's measured versus what's designed, and we'd rather show you the ledger than ask for trust.

Ask your node

Once your node is running, these readings answer questions directly in chat and through routines:

• "What did I miss yesterday that matters?"
• "What were my distraction patterns this week?"
• "What did you filter out this week — and why?"
• "What am I most interested in this month?"