HoneAI.Core 0.9.0

HoneAI.Core

The implementations for HoneAI — tiers of judges with measured trust and provenance — behind the HoneAI.Abstractions contracts. No third-party dependencies.

What it ships

• TieredRouter<TInput, TValue> (HoneAI.Routing) — the IReasoningRouter over any number of tiers: a tier below the top settles only where the policy snapshot puts it in Active or Stable for the input's slice, and only an answer the settlement policy lets it settle; otherwise the decision goes up without the lower answer. Every lower-tier settlement is drawn for a recheck at π.
• RecheckDraw — the deterministic draw: SHA-256 of the decision id and snapshot version → [0, 1) < π. TieredRouter.RecheckAsync judges a drawn decision again at a tier above, blind — told it is a recheck, given no hints, no lower answer and no abstain quantile — and settles nothing.
• PolicySnapshotJson — the canonical JSON of a policy snapshot (entries by tier and slice, rates by knob, times in UTC) and its SHA-256 content hash, for carrying a snapshot to a judging site and checking it is the one issued. Signing is the consumer's.
• HoneAI.Trust — earning the right to settle: TrustLedger / TeacherLedger (what rechecks and planted items show), TrustLifecycle (shadow → active → stable, and back), RecheckRate (the recheck probability a promise needs), RecheckBudget (sharing the tier above's capacity), PolicySnapshotIssuer, and TrustController.Close, which runs them in order at a window's end.
• Judge.From / Judge.Constant / MemoryJudge<,> — reference judges: a function, a constant, a memory of settled answers (System1).
• SettlementPolicy / SliceKey — categorical, tolerance and specification-limit policies; slice keys from a function.
• HttpMLoopClient — IMLoopClient over MLoop's REST API (mloop serve). MLoop is reached over the wire, never referenced as an SDK. Construct it with an HttpClient whose BaseAddress and bearer token you configure (e.g. via IHttpClientFactory).
• JsonlProvenanceSink — file-backed IProvenanceSink appending each assessment as one JSON line. Append-only, Git-friendly.
• InMemoryReviewStore / JsonlReviewStore — IReviewStore in memory, or as an append-only JSONL event file several processes on one machine can share.
• InMemoryLifecycleStore / JsonlLifecycleStore — ILifecycleStore in memory or as JSONL (the last line of a run id is the run).
• ModelLifecycle — a resumable train → poll → review (HITL) → promote run over an IMLoopClient, an IReviewStore and an ILifecycleStore, recording each stage to an optional IProvenanceSink.
• RolePersona — builds domain-neutral persona instructions for a trust-loop role, with the consumer's domain colour injected into a {Domain} slot.

Example — two tiers and a person

using HoneAI;
using HoneAI.Routing;

var memory = new MemoryJudge<Item, string>(
    new Tier { Id = "memory", Kind = JudgeKind.System1, Version = "1", CostRank = 0 }, item => item.Key);
var model  = Judge.From<Item, string>(
    new Tier { Id = "model", Kind = JudgeKind.Ml, Version = "exp-012", CostRank = 1 },
    async (item, ctx, ct) => await PredictAsync(item, ct));
var person = Judge.From<Item, string>(
    new Tier { Id = "person", Kind = JudgeKind.Human, Version = "1", CostRank = 2 },
    async (item, ctx, ct) => await ReviewAsync(item, ct));

var router = new TieredRouter<Item, string>(
    new IJudge<Item, string>[] { memory, model, person },
    SettlementPolicy.Categorical(new[] { "OK" }, mustCatch: new[] { "NG" }),
    SliceKey.From<Item>("line-v1", item => item.Line),
    decisionIdOf: item => item.Id,
    snapshot: () => snapshot);

var decision = await router.DecideAsync(item);
// decision.Prediction — the settled value and its provenance
// decision.Steps      — every tier's verdict, including what a tier in shadow would have answered

License

Apache-2.0