nescio 0.7.0

nescio (Latin: "I do not know") is a probabilistic database for data that is uncertain, contradictory, and decaying — an embedded Rust library and HTTP/JSON server that stores ignorance as a first-class object. A field without evidence is not NULL — it is a region of maximal entropy. Evidence narrows regions, time widens them again, and the database can tell you which evidence to acquire next.

Instead of values, you store claims: who said what, when, and how reliable they are. Everything else — credible regions, entropy in bits, answers — is derived at query time. Every source has a half-life; old claims lose their grip on the data by physics, not by TTL.

The verbs

Use cases

Built for data that is inherently uncertain, contradictory, and decaying:

• Lead & company data — enrichment from sources of mixed reliability; freshness is physics, not an updated_at column
• OSINT & investigations — contradictory claims with provenance, and resolve tells you what to verify next
• Sensor fusion — noisy measurements as intervals, fused across sources with per-source reliability
• Real-estate intelligence — price regions instead of point guesses; comparable search as an uncertain join
• Entity resolution & deduplication — join --op same scores candidate duplicates with a probability and a certainty

Quick start

cargo install nescio

nescio init mydb --template real-estate

nescio ingest mydb --entity villa_1 --slot price --interval 900000..1000000 \
       --source broker --at 2026-06-25

nescio bound mydb --entity villa_1 --slot price --at 2026-07-03

BOUND villa_1.price as of 2026-07-03
  region (95%): [570000, 1210000]
  entropy: 4.20 of 7.64 bits (knowledge 45%)
  MAP estimate: 905000

Ask again a year later — same command, --at 2027-07-03 — and the region has widened on its own. Erase a source with nescio forget-source, and every derived region widens correctly: there is no aggregate that could forget to forget.

Joins compare uncertain regions, so each match carries a probability and a certainty:

nescio join mydb --op approx --left price --right price --tol 50000   # comparable properties
nescio join mydb --op gt --left price --right price --certain          # A certainly dearer than B
nescio join mydb --op same --left city --right city                    # entity resolution: candidate duplicates

As a server (HTTP/JSON API)

nescio serve mydb --port 7777

curl 'localhost:7777/bound?entity=villa_1&slot=price&at=2026-07-03'

All verbs over HTTP/JSON, usable from any language. One process owns the database; reads run in parallel, writes are exclusive.

Typed clients for TypeScript and Java wrap every verb — both zero-dependency.

As a Rust library

use nescio::prelude::*;
use nescio::time::now_unix;
use std::path::Path;

let db = Db::open(Path::new("mydb"))?;
let q = Query::new(&db, now_unix());

let bound = q.bound("villa_1", "price", 0.95)?;
println!("{:.2} bits", bound.entropy_bits);

How it compares

• vs. NULL in SQL — NULL says only "no value". nescio says how much is unknown (entropy in bits), what is still credible (the region), and which evidence would shrink it. A classical relational database is the special case where every claim is an axiom and every region is a point.
• vs. probabilistic databases (MayBMS, Trio, BayesDB) — those attach probabilities to tuples in otherwise clean tables. nescio models the evidence itself — source reliability, decay, contradiction — and derives distributions at query time, so forgetting a source is exact, not approximate.
• vs. TTL and cache expiry — a TTL deletes at a cliff. Half-life decay widens uncertainty continuously: old data degrades gracefully instead of vanishing, and the database can report how stale is too stale for a given decision.

Performance

Measured on an M-series MacBook, 200,000 entities / 400,000 evidence records (cargo run --release --example bench):

ingest (group commit, one fsync)   ~1.1M records/s
open / log replay                  ~1.2M records/s
bound                              4.5 µs  (8.6 µs with couplings)
resolve                            < 1 ms

The server runs reads in parallel; every write is durable before it is acknowledged.

Storage

A database is a directory. Config is human-readable JSON; the evidence log is a compact, append-only binary format (~2.6× smaller than JSONL, no parse cost on replay). nescio export reconstructs readable JSONL any time, and nescio import goes the other way.

mydb/
  schema.json     slots and couplings
  sources.json    reliability, half-life, axiomatic
  priors.json     shared priors
  log.bin         the evidence log (append-only binary)

License

MIT