Skip to main content

Analysis is algebra, not conversation — the deterministic data plane (v2.63.0)

Ask an agent framework "how many leads per region?" and watch what actually happens: the rows are pasted into a prompt (or worse, described to the model), and a language model narrates a number. The narration is unfalsifiable. Nobody scanned the rows; nobody can re-derive the figure; the "analysis" is a claim in prose. That is the assertion-laundering failure (v2.53.0/T916) wearing a spreadsheet costume.

axon's data plane refuses the costume. v2.63.0 makes the five data-plane verbs relational algebra over a deterministic columnar engine:

The store — dataspace (axon-T928)​

dataspace Sales {
    column region: Text
    column amount: Float
}

A dataspace IS its schema: the closed type catalog (Text, Int, Float, Bool, Timestamp, Json) maps 1:1 to physical buffer layouts (validity bitmaps — nulls are structural, never sentinels; fixed-width buffers; offset buffers for variable width). An empty schema, a duplicate column, or an unknown type is a compile-time refusal. At deploy, the declaration is INSTANTIATED in the engine. Batches are immutable and append-only — the analytical dual of axonstore's transactional rows.

The load — ingest (axon-T929, v2.54.0, v2.52.0)​

ingest raw_csv into Sales { format: csv, limits { max_bytes: 1048576, max_rows: 100000 } }
  • Bounds BEFORE parse (v2.54.0): the byte bound is checked against the raw stream before a single byte is interpreted.
  • Refusal, not coercion: "not_a_number" in a Float column refuses the WHOLE batch, naming row and column. A missing value is a structural null — the only flexibility.
  • Born Untrusted (v2.52.0): every batch is stamped source + sha256 + ingested_at + Untrusted at construction. No unstamped batch can exist.
  • An ingest is a declared WRITE: a method: QUERY endpoint whose flow ingests is refused (axon-T927) — a safe method cannot append server state.

The algebra — focus / aggregate / associate / explore (axon-T930)​

focus Sales     { where: "amount >= 100.0", select: [region], as: big }
aggregate Sales { group_by: [region], compute: [count, sum(amount)], as: by_region }
associate Sales Accounts using region
explore Sales   { as: shape }

σ∘π, γ over the closed aggregate catalog, hash equi-join (NULL keys never join; a keyless join is a cartesian product and is refused), and a profile that describes shape, never content (Text zone boundaries are row content — an email, a name — so they are suppressed).

The where: clause is the ONE data-plane filter grammar the product already ships for retrieve/navigate (v1.30.0): closed, whitelisted operators, ${name} bindings resolved inside tokenized literals, injection-safe by construction. SQL semantics hold: AND binds tighter than OR; = NULL is is-null; an ordering against NULL matches nothing.

Why it is fast AND honest — the same property​

Each batch carries per-column zone maps ([min, max] + null count). A batch is skipped ONLY when the predicate is provably false over its zones — sound by construction (a skipped batch contains no matching row; a maybe batch is scanned). The pruning is observable: every result is a deterministic envelope

{ "rows": […], "taint": "untrusted", "stats": { "batches_total": 8, "batches_pruned": 5, "rows_scanned": 1200, "rows_matched": 37 } }

and the taint is the epistemic meet over the store's batches — an aggregate over untrusted data is born untrusted; no operation in the algebra can raise trust (raising trust is v2.52.0/v2.56.0's governed territory).

The fail-closed floor​

Without the engine port, all five verbs refuse: MissingDependency { name: "dataspace_engine" } — the mint/rotate posture, verbatim: an LLM fallthrough would HALLUCINATE a bearer token, a rotation, or — here — your data. axon does not ask a language model to pretend it loaded your rows.

The proof at deploy​

The PCC class DataspaceSchemaSoundness re-derives axon-T928/axon-T930 from the stored IR (recursing into if/for/par/warden — a nested violation is still a violation). A hand-edited artifact that queries a ghost column, smuggles an unknown column type, or aggregates outside the closed catalog is refuted BEFORE deploy.

See also​

  • a_safe_method_is_proven_safe (v2.62.0) — why an ingest behind QUERY refuses.
  • delivery_is_assertion_egress (v2.60.0) — the egress dual: what leaves the lattice.
  • effects_are_linear / dispatch_vs_cognition — the effect discipline this plane rests on.