Flagship · pluricellular runtime

HECATE

Several local AIs that behave as one organism: cells that serve on their own, a tissue that coordinates them, and a growth rule that changes the shape when hardware arrives.

  • Version 3.4.0
  • Licence MIT
  • Crates 10 under crates/core/
  • Language Rust · Python
Six outer cells and six inner cells joined into one tissue.
Illustration: a hexagonal cell drawn in glowing cyan line work on black, with a solid lime green hexagon at its centre.

What it is, in the repository's own words

ADR 0026 was accepted on 2026-10-31 and it changes the identity of the project, not its numbers: HECATE is a pluricellular organism of local AIs. The router is an organ, not the product.

The ADR exists because every ticket of the previous round was measured with a routing benchmark: regret, quality per cost. When the only metric is the router's, any router ends up looking like a router. That is the finding worth reading, and it is why this page separates what the design declares from what the code already contains.

  • Cell a node that serves, decides and adapts alone
  • Tissue coordination that does not dissolve the cells
  • Organ a combination, never a single structure
  • Emergence measured, or it does not exist

The pieces: declared against present

Two columns that must not be confused. Present means a symbol exists in the tree and the tests that exercise it ship in the archive. Declared means ADR 0026 names it as the thing that would make the hypothesis testable, and it is not written yet. Publishing the second column as if it were the first is exactly what the ADR forbids.

Status at commit b2a04ba, taken from the source tree
Piece Present What it does, or what it would do
plan_swarm yes Deploys n copies of a model split by layers. Builds topologies; does not pick between models.
redundancy_map yes Assigns layers and replicates heads so a pipeline survives losing one holder.
growth_decision yes FSENNL chooses Replicate, ExtendDepth or AddExpert when a node arrives.
ODAR yes Orders candidates by what has actually been observed per dimension, with an A/B against the old prior.
ExpertCatalog yes Says who is more capable at what; the estimator above is useless without it.
PeerReliability yes Keeps a per-peer record so the tissue can route around a cell that keeps failing.
HostResourceGuard yes Pauses a cell when the host it borrows is busy: the organism gives the machine back.
telomerase · merge_payloads yes Merges LoRA from several nodes with a per-task report, so the set produces something none of them did.
NcmAdaptiveHarness yes A cell picks its own depth from the budget it has.
TissueTopology declared The tissue as a first-class object rather than an emergent side effect of the scheduler.
ConformalCascade yes Escalates to the expensive cell only when the cheap one is not sure, with a coverage guarantee instead of a hand-picked threshold.
CellHealth declared Homeostasis: a cell that knows it is degrading and acts before the tissue notices.
EmergenceMetrics declared The comparison set against best cell. Without it the ADR cannot be closed either way.
RouterSLM declared A small local model choosing the route. The router that exists today is deterministic.

Ten of the fourteen rows are in the tree. The four in the second column are the work the ADR authorises, and the reason this page can show a status table instead of a feature list. The count moved from nine to ten while this page was being written: the conformal cascade landed the same day as the ADR, and a portfolio that cannot say that is a brochure.

The three measurements that exist

Every figure below is reproducible from the archive with the command printed beside it. Two of the three are negative results, which is the point.

  • Selection beats chance: 100 against 50.

    Over 100 tasks and five categories, the judge picks the better of two candidates 100 times; random choice would pick 50. The breakdown is published because a judge that aced maths and failed reasoning would hide behind an average. What this does not show is that the swarm beats one node. See below.

    cargo test -p hecate_node benchmark_quality

  • The swarm is wider, not smarter.

    The requested three-nodes-versus-one comparison cannot be built today: the mock backend has no seed, so three nodes return three identical strings, and picking between identical candidates is a tie by construction. The measurement was refused and the three missing preconditions were written down instead of being replaced by a conclusion about the scaffolding.

    docs/benchmarks/swarm-quality.md

  • Ninety per cent of the context can be dropped.

    A cell does not receive only the question; it inherits the context another cell accumulated. On this repository's own corpus — 130 documents, 102 671 terms when it was measured — a tenth of the context keeps full recall, for a normal query and for a paraphrased one. Chance needs all of it to reach the same place.

    docs/benchmarks/context-compression.md

What is not proven

The central claim of ADR 0026 is that the organism is more than the sum of its cells. That claim is not measured, and the document that would measure it does not exist.

docs/benchmarks/emergence.md is named inside the ADR as its proof, not as an annex. It is not in the tree. By the ADR's own rule, the ontology therefore sits in observation: the design is accepted, the property is not demonstrated, and no page on this site will state that more nodes make the system smarter. The earlier measurement points the other way (wider, not smarter), and it was taken without the organs the ADR authorises, which is precisely why the question is open rather than settled.

The most recent measurement sharpens the point, and it is the one worth reading twice. The conformal cascade works as a mechanism: it escalates on 20 % of tasks and lands at 1 985 ms between the 1 455 ms of the cheap cell and the 4 108 ms of the expensive one. But it scores 44 % against the 46 % of simply always using the cheap cell, so it does not beat its best competitor. The reason is that in this benchmark the expensive cell is not better than the cheap one (36 % against 46 %), which makes the original acceptance criterion — "95 % of the strong cell" — degenerate: it passes trivially against a weak competitor. The document closes it as a partial result, and so does this page.

The same discipline applies to the cell's own awareness. The ticket that measured whether a local SLM can tell when to answer and when to hand over found that it mostly hands over: on 90 decisions, 85 were delegations, and the nine-point edge over always-delegating is not significant at that sample size. The honest reading is in the benchmark, not in a slogan.

Read it or run it

Each archive is the full tree at the last commit, cut during the build. Extract one and check any number on this page: the tests that count them, the scripts that measure them and the notes that publish them are inside. Verify before you extract: SHA256SUMS and SIZES are served beside them.

hecate.tar.gz The fifteen documents

The GitHub mirror is temporarily restricted, so the documents are also served directly under /docs/, each in the language it is written in.