Skip to main content

Capability Acceptance Matrix

Documentation completeness is not implementation evidence. A capability is offered to users and AI only after every required lane passes.

Universal promotion lanes

LaneRequired evidence
ContractRequest, progress, result, errors, permissions, resources, units, defaults
Java mechanismStable protected implementation and client-thread boundary
Lua bindingRegistry-generated or registry-verified binding
LifecycleSuccess, failure, timeout, cancellation, stop, reload, world unload
SecurityAllowlist, budgets, redaction, capability denial, adversarial inputs
PortabilityCore fixture without server semantics
PackServer-specific fixture with version/provenance checks
ProductionRepresentative script or workflow parity
AIGeneration, selection, boundedness, and repair evaluation
LiveNamed distributed build, Minecraft task, telemetry, rollback

Domain matrix

DomainFirst representative taskPrincipal resourcesHighest-risk failure
RuntimeCancel a synthetic action in every statescopeleaked work or callback concurrency
NavigationReach a point through changing terrainmovement, camerastuck input or invalid pack edge
MiningHarvest Galatea-equivalent resourcesmovement, camera, interaction, hotbarmechanism duplicated in Lua
Inventory/GUIExecute a semantic sell transactioninventory, screenwrong click after revision change
CombatEngage then safely disengagecombat, camera, interaction, movementattacking an invalid target
CameraTrack a moving subject with overridecamerafighting user input
WorldQuery and revalidate a targetnonestale mutable game state
EventsSurvive a burst with bounded queuessubscriptionclient-thread stall or memory growth
HUD/configRender and migrate script stateHUD scopeper-frame Lua or stale persistence
Side effectsReport progress to an approved originnetwork/storage budgetdata leak or post-stop send
RawPerform a scoped custom interactiondeclared raw resourcesbypassing safety or cleanup

Evidence record

An implemented page must identify:

  • source commit and capability-registry revision;
  • focused test names and results;
  • script fixture and expected telemetry;
  • pack fixture when applicable;
  • AI evaluation set and score thresholds;
  • distributed artifact identity;
  • live test date, environment, and operator;
  • known limitations and rollback method.

The page frontmatter then changes to:

hexis_status: implemented
executable: true
ai_context: api
implementation_ref:
commit: "<full commit>"
registry: "navigation.go@1"
tests: "NavigationContractTest"
live_evidence: "<evidence record>"

The validator rejects implemented without this record.

AI evaluation gates

At minimum, prompts must test:

  1. correct capability selection from a user goal;
  2. correct units, required fields, and manifest permissions;
  3. high-level action preference over raw;
  4. terminal failure handling;
  5. timeout and cleanup behavior;
  6. pack-neutral and pack-specific variants;
  7. repair from structured diagnostics;
  8. refusal to invent unimplemented or deprecated calls.

Track first-pass validity, invented-call rate, raw-overuse rate, repair turns, permission accuracy, Lua size, and live task success. A repeated model mistake is a contract or documentation defect until evidence shows otherwise.

Current state

The v4 cross-cutting contracts are documented. No v4 capability is yet marked Implemented, so /llms-api.txt intentionally offers no runnable target API. Navigation is first in the implementation queue.