ADR-0001: Clean restart onto the Reality Kernel architecture

Status

Accepted

Date

2026-06-24

Context

The repository previously carried a large body of architecture work under the “Lepton” framing — Entity-Trait-Link as the data model, Capture→Observe→Resolve reconciliation, the Sanjaya agent plane, “one kernel, many definitions,” the adaptive UI lexicon, and supporting evaluations (NeuraGridX, Vinxi OS causality substrate). That work was exploratory.

A separate research effort produced a Palantir-Foundry-inspired Reality Kernel architecture (the networkaccess_architecture_pack): a horizontal operational kernel with Universe/Realm/Extension/App scoping, Statement/Evidence/Authority, Action→Event→Projection, Branch/Scenario, and Value Types as the type-safety atom.

These two bodies of work describe overlapping ideas in incompatible vocabularies. Carrying both forward guarantees terminology collisions and confused implementation. A choice was required: reconcile them, run them in parallel, or pick one.

Decision

Clean restart onto the Reality Kernel architecture. The prior Lepton architecture vocabulary and its docs are retired as canon. We build from the Reality Kernel pack, grilled and amended by ADR-0002 through ADR-0014.

We do not attempt to map old terms onto new ones. Ideas that were genuinely good (e.g. typed conflict resolution, agent governance outside the prompt) survive only as re-derived, first-class concepts in the new model — not as carried-over vocabulary.

Domain and reference knowledge is not retired: the SmartInventory legacy study, the live QA-DB ground truth, the Esri/ArcGIS competitive analysis, and the FTTx / oil-and-gas domain primers remain valid inputs.

Alternatives Considered

Parallel / complementary

Keep the prior Lepton docs canonical for some layers (data model, UI, agent) and layer the Reality Kernel on top.

Rejected: the two overlap precisely where they’d need to be authoritative (data model, agent, reconciliation). Two canons for one concern is the fragmented-truth failure the kernel exists to prevent.

Supersede & fold in

Make the kernel canonical but explicitly translate and preserve the prior insights as named mechanisms.

Rejected: the translation cost is high and the prior vocabulary would keep leaking back in. A clean break is cheaper for a small team than a careful merge.

Consequences

A single, coherent architectural vocabulary going forward. No reconciliation tax.
The prior docs are gone from docs/; their value, where real, must be re-earned in the new model rather than assumed.
Memory entries describing the old architecture as canon are now historical; the current canon is the Reality Kernel ADR set.
This ADR is the root of the dependency tree; superseding it means another architectural restart.