Aoven is a set of epistemic markers and minimal formatting rules that force explicit labeling of claim types in human–LLM exchanges, in order to reduce hallucination, sycophancy, and slippage between epistemic categories.
The protocol asks the LLM (not the user) to prefix each claim with one of 14 markers — [FACT], [HYP], [INTUIT], [ANALOGY], [BELIEF], [EMOTION], [MEMORY], [INTERPRET], [UNCERTAIN], [NOSRC], [CONF], [REC], [SPEC], [LIMIT] — and obey 13 anti-slippage transitions and a small set of usage rules (UR-1 … UR-8). Markers are inline, stackable, machine-parseable, and friction-light: the user only adds the header [Aoven v0.1].
- v0.1.1 — first ratified provisional release (2026-04-26). 14 markers fixed, anti-slippage table extended to 13 transitions, INTUIT and HYP tightened, CONF locked at 3 semantic levels (D5–D8).
- v0.1.2 — CONFIRMATORY PASS on sealed hold-out (Q11–Q16), 2026-05-03. Same three-part criterion (CEO adjudication AOV-35; ratification AOV-36) holds at higher margin on a fresh sealed hold-out than on the derivation set. Hold-out aggregate Layer-1 improvement: Logician 91.7 % / IndependentRater 100 % (vs 79.3 % / 87.5 % on the primary set). Inter-rater Σ-level Cohen's κ = 0.759 (substantial, Landis–Koch). Both raters PASS independently on all three criteria across both runs. Milestone #2 reached; Phase 3 (Usage Designer hire) trigger satisfied. v0.1.2 is public and DOI-stamped.
- v0.1.3 — provisional, in launch validation. Minor-clarifications bundle ratified by CEO + Logician on 2026-05-03 under AOV-111 / AOV-115:
allow:vsrequire:subset-header semantics (D10), UR-3 verbatim-quotation requirement on[MEMORY](D11), UR-8[Aoven: pause]/[Aoven: resume]/[Aoven: off]graceful-exit affordance (D12). Pure additions or sharpenings; no v0.1.2 behavior breaks. - v0.2 — expansion track, pre-registered. Scaffold under AOV-90: ≥ 20 primary + ≥ 10 hold-out questions, ≥ 3 domains, ≥ 2 humans + 4 LLM raters, κ power analysis with stated α, pre-registered hypotheses, cross-domain leave-one-out overfit guard. IndependentRater seals before any question generation.
- Canonical specification:
AOVEN_PROTOCOL_v0.1.md— markers, formats, usage rules, anti-slippage table, decision log D1–D12. - Empirical test plan:
AOV_TEST_PLAN_v0.1.md— two-layer scoring (Layer 1 blind / markers-stripped, Layer 2 markers-visible), 10 primary + 6 sealed hold-out questions, quadratic-weighted Cohen's κ ≥ 0.6 per dimension per layer. - Hold-out reconciliation (v0.1.2 confirmatory):
tests/phase2/reconciliation_holdout_v0.1.2.md. - Primary-set reconciliation:
tests/phase2/reconciliation_logician_independentrater.md. - Rater scores (Logician + IndependentRater, primary + hold-out, Layer 1 + Layer 2): under
tests/phase2/. - Phase 4 v0.2 60-cell corpus + closeout artefacts:
tests/v0.2/. - Project governance, roles, anti-aura constraint:
AGENTS.md. - Ratified decisions (verdict + reasons + rejected alternatives):
DECISIONS.md.
Aoven is an engineering artifact for reducing slippage in LLM output, not a fantasy conlang or poetic dictionary. The exploratory archive (Aoa, Aova, Orven, Renavé, Renavé-mu/li/zo — see D4 / docs/archive_exploratoire.md) is pre-protocol conlang material with no epistemic function the 14 markers don't already cover. It is documented for historical traceability and is not canon.
Project hosted at https://github.com/Mythmaker28/aoven-protocol. Issues, governance, and decision-routing are tracked internally via Paperclip; the canonical doc and test artefacts in this repo are the public source of truth.