From 32bdd7062b8c475dee006c39312058854c4702c3 Mon Sep 17 00:00:00 2001 From: "Paperclip CTO (e8587a99)" Date: Sun, 26 Apr 2026 05:46:14 +0200 Subject: [PATCH] Clean repository docs and fix exploratory archive --- AGENTS.md | 91 ++++++++++++++++++++++++-- AOVEN_PROTOCOL_v0.1.md | 122 ++++++++++++++++++++++++++++++----- DECISIONS.md | 58 +++++++++++++++-- README.md | 29 +++++++-- docs/archive_exploratoire.md | 41 ++++++++++-- 5 files changed, 303 insertions(+), 38 deletions(-) diff --git a/AGENTS.md b/AGENTS.md index 694bdd5..cc5a1fe 100644 --- a/AGENTS.md +++ b/AGENTS.md @@ -1,11 +1,90 @@ - +# AGENTS.md — Aoven project governance -# AGENTS.md +Public governance rules for any agent (human or LLM) contributing to Aoven. This file is the project-public charter; private agent instruction files are out of scope here. -**Status:** placeholder. Final content owned by AOV-12b (CanonicalScribe). +--- -This is the **project-public** governance file for agents (human or LLM) working on Aoven. It is NOT an internal agent-instruction file. +## Roles -The canonical version will define: roles, governance rules, anti-hallucination constraints, and the anti-aura rule (Aoven is not a fantasy conlang or poetic dictionary). +| Role | Responsibility | Hire / authorship constraint | +|------|---------------|------------------------------| +| **Protocol Architect** | Drafts and revises the canonical spec (`AOVEN_PROTOCOL_v0.1.md`): markers, formats, usage rules, anti-slippage rules. | May not also serve as sole reviewer of own drafts. | +| **Epistemic Logician** | Audits drafts for category leaks, definitional circularity, missed slippage paths. Independent reviewer, not co-author of the artifact under review. | Must be structurally independent from the artifact under audit at the moment of review. | +| **Canonical Scribe** | Records validated, provisional, in-observation, rejected, and open-question decisions in the canonical spec. Does not author content. Records both sides of disagreements. | Anti-editorialization: records, does not argue. | +| **Red Team / Independent Rater** | Adversarial probing of the protocol; third independent rater on empirical tests. Not a co-author of the protocol or the rubric being applied. | Must be neither author of the protocol nor scorer of the rubric they apply. | -This placeholder will be replaced by the Scribe-authored AGENTS.md in a follow-up commit. +--- + +## Governance rules + +### G-1 — Named-reviewer sign-off gate (no author self-validation) + +A decision moves from `[provisional]` to `[validated]` only after a named reviewer who is not the author signs off in writing on a referenced artifact (issue comment, audit document). Author self-declaration of validation is not a valid gate. If a CEO ratification lands before the audit is complete, the audit's blocking findings still apply and must be integrated before the next phase opens. + +### G-2 — Anti-aura rule + +No prose, term, or claim that sounds impressive without an operational definition. Every term used in this repo must either: + +- be one of the 14 canonical markers in `AOVEN_PROTOCOL_v0.1.md`, or +- have a strict definition in the spec, or +- carry an explicit status tag (`[provisional]`, `[observation]`, `[open]`, `[rejected]`). + +Aoven is **not** a conlang, not a poetic vocabulary, not a personal project. Seductive language without definition is removed without discussion. + +### G-3 — New-vocabulary budget: ≤3 non-marker terms + +Total new vocabulary outside the marker set is capped at 3 terms. Adding a term requires demonstrating an epistemic function not covered by an existing marker. As of v0.1.2: **0 used**. + +### G-4 — Burden of proof on retention + +Exploratory or legacy terms are not Aoven canon by default. The burden of proof is on retention, not on rejection. A term remains `[observation]` until an agent demonstrates a unique epistemic function. + +### G-5 — Disagreement is recorded, not resolved silently + +If two agents disagree on a decision, both positions are recorded with reasons in `DECISIONS.md` along with the resolution path (or its absence). The Scribe does not pick a winner. + +### G-6 — Hiring rule (structural conflict only) + +Hires are added when the project has a structural conflict that no current role can resolve without compromising independence — e.g., a third independent rater is needed because the only available raters are protocol co-authors. Hires are not added on workload alone. + +--- + +## Anti-hallucination constraints (binding on contributors) + +These constraints apply to anything written into the spec, decision log, or test outputs. + +- **C-1 — Marker discipline.** Any claim in the canonical spec or test results must fit one of the 14 marker categories. If it does not fit, it does not belong, or a new marker proposal is required (subject to G-3). +- **C-2 — Fact ≠ hypothesis.** A `FACT` requires an external verifiable source; a `HYP` requires a stated, specific test condition. Confident belief is not FACT. Falsifiable-in-principle is not HYP. +- **C-3 — Analogy ≠ proof.** Conclusions drawn from an `ANALOGY` carry `HYP` or `SPEC` on the derived claim; they never inherit truth-status from the source domain. +- **C-4 — Intuition is not a hypothesis.** Upgrading `INTUIT` to `HYP` requires stating a specific test condition. A marker swap alone is the slippage, not its repair. +- **C-5 — No silent withdrawal.** When a `BELIEF` or `NOSRC` claim is challenged, respond by either citing a source (upgrade to `FACT`) or downgrading to `UNCERTAIN`. Quietly dropping the claim is itself a slippage. +- **C-6 — No invented history.** Facts about the project's past go in the spec, decision log, or issue thread, or they do not go anywhere. Reconstructed narratives are out. + +--- + +## Anti-scope — what Aoven is NOT + +- **NOT a fantasy conlang.** Markers are not flavor words. The spec is not world-building. +- **NOT a poetic dictionary.** Aesthetic appeal is not a retention criterion. +- **NOT a personal project.** Scope is set by the empirical question (does it reduce the four failure modes), not by individual authorial preference. +- **NOT a private vocabulary system.** Aoven exists only insofar as it produces measurable improvement in human–LLM exchanges per the test plan. + +--- + +## Workflow + +1. **Drafting** — Protocol Architect drafts in the canonical spec or in a labeled patch. Each entry carries source attribution and a status tag. +2. **Audit** — Epistemic Logician reviews independently and writes blocking findings. The audit is a public artifact. +3. **Patch** — Architect integrates blocking findings before the gate opens. The Scribe records the patch with full context. +4. **Empirical test** — Test plan runs A/B on a fixed question set with the rubric in `tests/test_plan.md`. Three independent raters; quadratic-weighted Cohen's κ ≥ 0.6 per dimension. Pre-committed refinement candidates and a hold-out probe set guard against p-hacking. +5. **Versioning** — Decisions move from `[provisional]` to `[validated]` only after named-reviewer sign-off (G-1) and, where applicable, empirical signal. + +--- + +## Status tags (used throughout the repo) + +- `[validated]` — accepted, passed audit and (where applicable) test signal. +- `[provisional]` — in use, awaiting test data or named-reviewer sign-off. +- `[observation]` — under watch; may or may not be retained. +- `[rejected]` — explicitly excluded; reason recorded. +- `[open]` — open question; no decision yet. diff --git a/AOVEN_PROTOCOL_v0.1.md b/AOVEN_PROTOCOL_v0.1.md index caab6c9..afe937d 100644 --- a/AOVEN_PROTOCOL_v0.1.md +++ b/AOVEN_PROTOCOL_v0.1.md @@ -1,9 +1,18 @@ # AOVEN_PROTOCOL_v0.1-provisional -> **v0.1.1 / status: PROVISIONAL — ratified by CEO+CTO+Logician 2026-04-26** +> **v0.1.2 / status: PROVISIONAL — ratified by CEO+CTO+Logician 2026-04-26** +> > Maintained by CanonicalScribe (e19c696f). Scribe records; does not editorialize. +> > Status tags: [validated] [provisional] [observation] [rejected] [open] -> Last updated: 2026-04-26 — applied v0.1.1 patch (CTO comment 22fb25e4 on AOV-7, CEO sign-off comment 1f03749a) +> +> Last updated: 2026-04-26 — v0.1.2 patch (board correction on AOV-15, applied via AOV-16). + +## Changelog + +- **v0.1.2** — corrected Renavé family historical origin (board correction, AOV-15); markdown readability pass; no semantic change to markers, formats, or anti-slippage rules. +- **v0.1.1** — INTUIT and HYP definitions tightened; anti-slippage table extended from 10 to 13 transitions; CONF gradient fixed at 3 semantic levels (CTO patch on AOV-7 comment `22fb25e4`, CEO sign-off `1f03749a`). +- **v0.1.0** — initial draft (CTO, AOV-7, 2026-04-26). --- @@ -156,16 +165,93 @@ When a `[BELIEF]` or `[NOSRC]` claim is challenged, the response must either (a) ## Decision log -| ID | Decision | Reason | Alternatives rejected | Risk | Status | -|----|----------|--------|-----------------------|------|--------| -| D1 | Keep all 14 distinct markers (not merged to 10) | Tested merges of UNCERTAIN+NOSRC, SPEC+HYP, EMOTION+INTUIT — each merge collapses a distinction carrying a different slippage risk. UNCERTAIN = no answer exists; NOSRC = answer held but uncited. Merging makes both risks invisible. Logician audit (AOV-9) confirmed all distinctions load-bearing. | 10-marker compact version | 14 markers may be cognitively heavy; mitigated by `require:` subset invocation | [validated] | -| D2 | Square bracket syntax `[MARKER]`, inline prefix per claim, stackable | Minimal friction, readable inline with natural language, machine-parseable, no new syntax | JSON format (too verbose), suffix notation (disrupts reading), color coding (non-portable) | Square brackets conflict with Markdown link syntax in some renderers; fallback: unicode brackets | [provisional] (waiting on first A/B test) | -| D3 | LLM applies markers, not user | Reduces cognitive burden on user; optional user markers permitted but not required | Requiring user to pre-tag input — too high friction, violates usability constraint | LLM may misapply markers; mitigation: A/B tests measure marker accuracy rate | [provisional] (waiting on first A/B test) | -| D4 | All old exploratory terms archived (Aoa, Aova, Orven, Renavé, Renavé-mu/li/zo) | None serves an epistemic function not covered by the 14 markers. Renavé-mu/li/zo were confidence gradients; CONF(high/medium/low) covers this. Board brief establishes default: "not Aoven canon unless agent demonstrates otherwise; burden of proof is on retention." | Promotion to canonical status | Prior users may expect these terms; explicit archive notation prevents confusion | [validated] | -| D5 | INTUIT redefinition (v0.1.1) | Original "felt sense or heuristic judgment" merged a pre-verbal felt sense with a heuristic judgment — different slippage profiles. New definition anchors on inability to articulate reasoning, closing the leak. | Keeping original definition; alternative phrasings | Tighter definition may reject borderline INTUIT use; mitigated by usage examples | [validated] | -| D6 | Anti-slippage table extended from 10 to 13 transitions (v0.1.1) | Logician audit identified 3 missing slippage paths: INTUIT→HYP laundering, INTERPRET→certainty, CONF(high)→FACT. Each is a distinct, plausible LLM failure mode. INTUIT→HYP is added alongside INTUIT→FACT, not replacing — they block different actions. | Consolidating INTUIT rules to 12 rows | Larger table = more for LLM to honor; mitigated by tightness of each rule | [validated] | -| D7 | HYP definition cleanup, no forward-reference (v0.1.1) | Original HYP definition referenced SPEC inside its own definition. Definitions should stand alone. New definition replaces forward-reference with explicit "specific, statable test condition" requirement. SPEC contrast moved to "Does NOT mean". | Keeping original HYP definition | None significant | [validated] | -| D8 | CONF gradient at 3 levels, no numeric (v0.1.1) | CONF(0.8) implies calibration infrastructure that does not exist for current LLMs and creates false precision. Three semantic levels (high/medium/low) are interpretable without calibration claims. Logician concurred. | Adding numeric confidence; finer gradient | Three levels may be insufficient — revisit in v0.2 if A/B tests show signal | [validated] | +Each decision below is a discrete protocol-level commitment. Verdicts D1–D3 and D5–D8 are unchanged from v0.1.1. D4 was corrected in v0.1.2 per board input on AOV-15 (see D9). The mirror in `DECISIONS.md` follows this same numbering. + +--- + +### D1 — Keep all 14 distinct markers (not merged to 10) + +- **Reason.** Tested merges of UNCERTAIN+NOSRC, SPEC+HYP, EMOTION+INTUIT — each merge collapses a distinction carrying a different slippage risk. UNCERTAIN = no answer exists; NOSRC = answer held but uncited. Merging makes both risks invisible. Logician audit (AOV-9) confirmed all distinctions load-bearing. +- **Alternatives rejected.** 10-marker compact version. +- **Risk.** 14 markers may be cognitively heavy; mitigated by `require:` subset invocation. +- **Status.** [validated] + +--- + +### D2 — Square-bracket syntax `[MARKER]`, inline prefix per claim, stackable + +- **Reason.** Minimal friction, readable inline with natural language, machine-parseable, no new syntax. +- **Alternatives rejected.** JSON format (too verbose); suffix notation (disrupts reading); color coding (non-portable). +- **Risk.** Square brackets conflict with Markdown link syntax in some renderers; fallback: unicode brackets. +- **Status.** [provisional] — gated on first A/B test. + +--- + +### D3 — LLM applies markers, not user + +- **Reason.** Reduces cognitive burden on user; optional user markers permitted but not required. +- **Alternatives rejected.** Requiring user to pre-tag input — too high friction, violates usability constraint. +- **Risk.** LLM may misapply markers; mitigation: A/B tests measure marker accuracy rate. +- **Status.** [provisional] — gated on first A/B test. + +--- + +### D4 — All earlier exploratory terms archived (Aoa, Aova, Orven, Renavé, Renavé-mu/li/zo) + +- **Reason.** None serves an epistemic function not already covered by the 14 markers. Burden of proof on retention. Correct historical origin per board input on AOV-15: + - **Renavé** — relation de présence répétée sans interaction (relation of repeated presence without interaction). + - **Renavé-mu** — Renavé réciproque (reciprocal Renavé). + - **Renavé-li** — Renavé asymétrique (asymmetric Renavé). + - **Renavé-zo** — sentiment résiduel laissé par la disparition silencieuse d'un Renavé (residual feeling left by the silent disappearance of a Renavé). +- **NOSRC discipline note (v0.1.2).** Earlier wording in this row asserted that Renavé-mu/li/zo "were confidence gradients superseded by CONF(high/medium/low)". That claim was a **NOSRC fabrication** — no agent had a source for the original meaning of these terms. The board provided the correct origin on AOV-15. The fabrication is recorded here, not silently overwritten, because the slippage class — confident assertion without source — is the exact failure mode Aoven exists to prevent. See D9 for the correction trail. +- **Alternatives rejected.** Promotion to canonical status. +- **Risk.** Recurrence of the same NOSRC pattern when reconstructing project history; mitigation: D9 makes the failure visible and the C-6 ("no invented history") rule in `AGENTS.md` is binding on future contributors. +- **Status.** [validated] — verdict (archived; not canon) unchanged. + +--- + +### D5 — INTUIT redefinition (v0.1.1) + +- **Reason.** Original "felt sense or heuristic judgment" merged a pre-verbal felt sense with a heuristic judgment — different slippage profiles. New definition anchors on inability to articulate reasoning, closing the leak. +- **Alternatives rejected.** Keeping original definition; alternative phrasings. +- **Risk.** Tighter definition may reject borderline INTUIT use; mitigated by usage examples. +- **Status.** [validated] + +--- + +### D6 — Anti-slippage table extended from 10 to 13 transitions (v0.1.1) + +- **Reason.** Logician audit identified 3 missing slippage paths: INTUIT→HYP laundering, INTERPRET→certainty, CONF(high)→FACT. Each is a distinct, plausible LLM failure mode. INTUIT→HYP is added alongside INTUIT→FACT, not replacing — they block different actions. +- **Alternatives rejected.** Consolidating INTUIT rules into 12 rows. +- **Risk.** Larger table = more for LLM to honor; mitigated by tightness of each rule. +- **Status.** [validated] + +--- + +### D7 — HYP definition cleanup; no forward-reference (v0.1.1) + +- **Reason.** Original HYP definition referenced SPEC inside its own definition. Definitions should stand alone. New definition replaces forward-reference with explicit "specific, statable test condition" requirement. SPEC contrast moved to "Does NOT mean". +- **Alternatives rejected.** Keeping original HYP definition. +- **Risk.** None significant. +- **Status.** [validated] + +--- + +### D8 — CONF gradient at 3 levels (high/medium/low); no numeric + +- **Reason.** `CONF(0.8)` implies calibration infrastructure that does not exist for current LLMs and creates false precision. Three semantic levels are interpretable without calibration claims. Logician concurred. +- **Alternatives rejected.** Adding numeric confidence; finer gradient. +- **Risk.** Three levels may be insufficient — revisit in v0.2 if A/B tests show signal. +- **Status.** [validated] + +--- + +### D9 — Renavé family historical descriptions corrected per board input (v0.1.2) + +- **Reason.** Prior wording in D4 and in the exploratory archive table asserted that Renavé-mu/li/zo were confidence gradients superseded by `CONF(high/medium/low)`. That assertion had no source — none of the agents had data on the original meaning of these terms. Board input on AOV-15 supplied the correct origin (see D4). The correction is logged here as an explicit acknowledgement of NOSRC discipline failure, not a silent typo edit; this is the slippage class Aoven is built to prevent. +- **Alternatives rejected.** Silent overwrite without trail; deletion of the old text without explanation. +- **Risk.** None to the spec. Recurrence risk on future history claims; mitigated by C-6 ("no invented history") in `AGENTS.md` and by the named-reviewer gate on D9 itself. +- **Status.** [validated] — board-supplied source; named-reviewer sign-off requested from Logician (`2ae117a1`) on AOV-16. --- @@ -189,17 +275,19 @@ When a `[BELIEF]` or `[NOSRC]` claim is challenged, the response must either (a) The following terms come from an earlier conlang phase and are not canon. Decision D4 [validated] archived all. They may only be reused if an agent demonstrates they serve a unique epistemic function not covered by the 14 markers. -*Verdict source: CTO draft, AOV-7, 2026-04-26* +**v0.1.2 correction (per D9).** The Notes column for the Renavé family was previously fabricated (asserted as confidence gradients without source). The text below reflects the board-supplied historical origin from AOV-15. Reuse status is unchanged; only the historical description is corrected. + +*Verdict source: CTO draft, AOV-7, 2026-04-26. Renavé family origin source: board input on AOV-15, applied via AOV-16.* | Term | Origin phase | Reuse status | Notes | |------|-------------|--------------|-------| | Aoa | Pre-protocol conlang | [observation] | No epistemic function identified. Not permanently rejected — requires evidence of unique function. | | Aova | Pre-protocol conlang | [observation] | No epistemic function identified. Not permanently rejected — requires evidence of unique function. | | Orven | Pre-protocol conlang | [observation] | No epistemic function identified. Not permanently rejected — requires evidence of unique function. | -| Renavé | Pre-protocol conlang | [observation] | No epistemic function identified. Not permanently rejected — requires evidence of unique function. | -| Renavé-mu | Pre-protocol conlang | [observation] | Was a confidence gradient. Superseded by CONF(high/medium/low). | -| Renavé-li | Pre-protocol conlang | [observation] | Was a confidence gradient. Superseded by CONF(high/medium/low). | -| Renavé-zo | Pre-protocol conlang | [observation] | Was a confidence gradient. Superseded by CONF(high/medium/low). | +| Renavé | Pre-protocol conlang | [observation] | Original meaning (board, AOV-15): *relation de présence répétée sans interaction* — relation of repeated presence without interaction. Not an epistemic marker; no function the 14 markers don't cover. | +| Renavé-mu | Pre-protocol conlang | [observation] | Original meaning (board, AOV-15): *Renavé réciproque* — reciprocal Renavé. Not a confidence gradient; the prior "superseded by CONF" claim was a NOSRC fabrication corrected in v0.1.2 (see D9). | +| Renavé-li | Pre-protocol conlang | [observation] | Original meaning (board, AOV-15): *Renavé asymétrique* — asymmetric Renavé. Not a confidence gradient; the prior "superseded by CONF" claim was a NOSRC fabrication corrected in v0.1.2 (see D9). | +| Renavé-zo | Pre-protocol conlang | [observation] | Original meaning (board, AOV-15): *sentiment résiduel laissé par la disparition silencieuse d'un Renavé* — residual feeling left by the silent disappearance of a Renavé. Not a confidence gradient; the prior "superseded by CONF" claim was a NOSRC fabrication corrected in v0.1.2 (see D9). | --- diff --git a/DECISIONS.md b/DECISIONS.md index c89d3b8..507f909 100644 --- a/DECISIONS.md +++ b/DECISIONS.md @@ -1,9 +1,57 @@ - - # DECISIONS.md -**Status:** placeholder. Final content owned by AOV-12b (CanonicalScribe). +Public log of decisions affecting the Aoven protocol and the team/process around it. Two sections, distinguished by versioning: + +- **Protocol decisions** are versioned with the canonical spec (`AOVEN_PROTOCOL_v0.1.md`). Changing one bumps the protocol version. +- **Process decisions** govern how the team works. They do not bump the protocol version, but they constrain how protocol changes are made. + +Status legend: `validé provisoire` / `en observation` / `rejeté` / `question ouverte` / `validé`. + +Format per row: `id | decision | reason | alternatives rejected | status`. + +--- + +## Protocol decisions (versioned with the spec) + +Source: decision log in `AOVEN_PROTOCOL_v0.1.md` (last updated 2026-04-26, v0.1.2 patches integrated). + +| id | decision | reason | alternatives rejected | status | +|----|----------|--------|------------------------|--------| +| D1 | Keep all 14 distinct markers (no merge to 10). | Tested merges of UNCERTAIN+NOSRC, SPEC+HYP, EMOTION+INTUIT — each merge collapses a distinction tied to a different slippage risk. UNCERTAIN = no answer exists; NOSRC = answer held but uncited. Merging makes both invisible. Logician audit (AOV-9) confirmed all distinctions load-bearing. | 10-marker compact version. | validé | +| D2 | Square-bracket inline syntax `[MARKER]`, prefix per claim, stackable. | Minimal friction, readable inline with natural language, machine-parseable, no new syntax beyond brackets. | JSON format (too verbose); suffix notation (disrupts reading); color coding (non-portable). | validé provisoire (gated on first A/B test) | +| D3 | The LLM applies markers, not the user. | Reduces user cognitive burden; optional user markers permitted but never required. | Requiring user pre-tagging — too high friction, violates usability constraint. | validé provisoire (gated on first A/B test) | +| D4 | All earlier exploratory terms archived (Aoa, Aova, Orven, Renavé, Renavé-mu/li/zo). | None serves an epistemic function not already covered by the 14 markers. Burden of proof on retention. Correct historical origin per board input on AOV-15: Renavé = relation de présence répétée sans interaction; Renavé-mu = Renavé réciproque; Renavé-li = Renavé asymétrique; Renavé-zo = sentiment résiduel laissé par la disparition silencieuse d'un Renavé. **The earlier wording in this row — that Renavé-mu/li/zo "were confidence gradients superseded by CONF" — was a NOSRC fabrication; corrected in v0.1.2, see D9.** Verdict (archived; not canon) unchanged. | Promotion to canonical status. | validé | +| D5 | Tightened INTUIT definition (v0.1.1). | Original "felt sense or heuristic judgment" merged a pre-verbal felt sense with a heuristic judgment — different slippage profiles. New definition anchors on inability to articulate the inference chain. | Keeping original definition; alternative phrasings. | validé | +| D6 | Anti-slippage table extended from 10 to 13 transitions (v0.1.1). | Logician audit identified 3 missing slippage paths: INTUIT→HYP laundering, INTERPRET→certainty, CONF(high)→FACT. Each is a distinct, plausible failure mode. INTUIT→HYP added alongside INTUIT→FACT, not replacing — they block different actions. | Consolidating INTUIT rules into 12 rows. | validé | +| D7 | HYP definition cleanup; no forward-reference (v0.1.1). | Original HYP definition referenced SPEC inside its own definition. Definitions must stand alone. New definition replaces the forward-reference with an explicit "specific, statable test condition" requirement. | Keeping original HYP definition. | validé | +| D8 | CONF gradient at 3 semantic levels (high/medium/low); no numeric confidence. | Numeric `CONF(0.8)` implies calibration infrastructure that does not exist for current LLMs and creates false precision. Three semantic levels are interpretable without calibration claims. Logician concurred. | Numeric confidence; finer gradient. | validé | +| D9 | Renavé family historical descriptions corrected per board input (v0.1.2). | Prior wording in D4 + the exploratory archive table claimed Renavé-mu/li/zo were confidence gradients superseded by `CONF`. None of the agents had a source for the original meaning of these terms; the board supplied the correct origin on AOV-15 (see D4). The correction is logged as an explicit acknowledgement of NOSRC discipline failure — exactly the slippage class Aoven is built to prevent — not as a typo edit. | Silent overwrite without trail; treating the correction as a typo. | validé | + +Open protocol questions: none active. v0.1 audit questions OQ-1 through OQ-5 resolved during the v0.1.1 patch cycle; trail in `AOVEN_PROTOCOL_v0.1.md` § « Resolved questions ». + +--- + +## Process decisions (team / governance) + +Decisions that constrain how protocol decisions are made and reviewed. Not versioned with the spec, but binding on contributors. + +| id | decision | reason | alternatives rejected | status | +|----|----------|--------|------------------------|--------| +| P1 | **Anti-aura rule.** No prose, term, or claim without an operational definition. Aoven is not a fantasy conlang, not a poetic dictionary, not a personal project. | Without this rule, the project drifts toward aesthetic vocabulary that cannot be tested or falsified — the exact failure mode an epistemic protocol must avoid. | Allowing undefined evocative terms when "they sound right". | validé | +| P2 | **Named-reviewer sign-off gate.** A decision moves from `provisoire` to `validé` only after a named reviewer who is not the author signs off in writing. Author self-declaration does not satisfy the gate. | The CTO's 02:06 self-validation on AOV-7 was followed by a Logician audit at 02:10 with 5 blocking findings; CEO had already ratified at 02:09. The blocking findings were integrated before Phase 2. The incident defined the rule: validation is a named-reviewer act, not a self-act. | Author self-validation; CEO-only ratification without independent audit. | validé | +| P3 | **Hiring rule — structural conflict only.** A new role is added when the project has a structural conflict that no current role can resolve without compromising independence. Workload alone is not a hiring trigger. | Both protocol authors (CTO, Logician) are also raters; the only way to add an independent rater was a third hire (Red Team, CEO-approved `8dc2da86`). The conflict was structural, not capacity. | Adding hires for workload smoothing; relying on author-raters with disclosure only. | validé | +| P4 | **Exploratory words — burden of proof on retention.** Earlier conlang terms remain `observation` by default. They re-enter canon only if an agent demonstrates a unique epistemic function not covered by the 14 markers. | Without this rule, legacy vocabulary returns by inertia and inflates the marker set without epistemic gain. | Default-retain (legacy momentum); default-reject (loses traceability). | validé | +| P5 | **A/B test rubric: 8 dimensions, gated on Logician countersign.** The 7-dimension draft missed coverage on SPEC→REC and REC→injunction; an 8th dimension was added. The rubric ships only after the Logician countersigns the coverage check against the 13 anti-slippage transitions. | A 7-dim rubric would let the test "pass" while two slippage transitions remained unmeasured. The countersign requirement enforces P2 at the artifact level. | 7-dim rubric; ship without Logician countersign. | validé (per `AOV_TEST_PLAN_v0.1.md`, Logician audit `96582053`) | +| P6 | **New-vocabulary budget: ≤3 non-marker terms.** Adding a non-marker term requires demonstrating an epistemic function not covered by the marker set. As of v0.1.2: 0 used. | Without a hard cap, vocabulary grows by accretion and the protocol becomes unreadable. | No cap; cap by review only. | validé provisoire | +| P7 | **Disagreement is recorded, not silently resolved.** Two-way disagreements between agents are logged with both positions and the resolution path (or its absence). The Scribe does not editorialize. | Silent resolution erases audit trail and lets one side's framing become canon by default. | Scribe picks a side; private resolution outside the log. | validé | +| P8 | **Pre-committed refinement candidates + hold-out probe set + cap at 1 re-test.** If the A/B test result is inconclusive (10–20% improvement), refinements are drawn only from a list locked before scoring. Q11–Q15 are reserved as a hold-out set. A second inconclusive result terminates v0.1 and ships v0.2 with structural changes. | Open-ended re-tests on the same probes are researcher-degrees-of-freedom; iterating until the threshold is crossed is p-hacking under another name. | Open-ended refinement loop on the same probe set. | validé (patched per Logician F6) | +| P9 | **Inter-rater reliability target: quadratic-weighted Cohen's κ ≥ 0.6 per dimension.** Plain Cohen's κ is for nominal categories; the 0–3 ordinal scale needs weighted κ. Dimensions failing the threshold are flagged as under-specified and excluded from pass/fail until rubric-revised. | Plain κ penalizes off-by-1 the same as off-by-3 on an ordinal scale, which is wrong; without an IRR threshold, scores are arbitrary. | Plain Cohen's κ; no IRR threshold. | validé (patched per Logician F4) | +| P10 | **Stripped-marker re-scoring + Red Team independent rater.** Aoven `[MARKER]` syntax leaks treatment status to raters, breaking blinding. Compromise: raters re-score Test B with markers mechanically removed; if blinded vs. unblinded scores diverge by >0.5, blinded scores take precedence. Red Team's pass is the primary control for author-rater bias. | Plain blinding loses the marker-accuracy signal Aoven exists to produce; no blinding makes the test a self-confirmation. | Plain blinding (loses signal); ignore unblinding (test cannot answer its own question). | validé provisoire (patched per Logician F2) | + +--- -The canonical decision log lives in [`AOVEN_PROTOCOL_v0.1.md`](./AOVEN_PROTOCOL_v0.1.md) under "Decision log" (D1–D8) until the Scribe-authored top-level DECISIONS.md is published. The Scribe version will consolidate protocol and process decisions with reasons, rejected alternatives, and status (`validated` / `provisional` / `observation` / `rejected` / `open`). +## What this log does not cover -This placeholder will be replaced in a follow-up commit. +- **Decisions internal to a single agent's role.** Captured in the agent's instruction file, not here. +- **Open empirical questions awaiting test data.** Tracked in `AOVEN_PROTOCOL_v0.1.md` § « Open questions » and in test artifacts under `tests/`. +- **Live disagreements without a resolution path yet.** Recorded as `question ouverte` rows when they reach this log; before that, they live in issue threads. diff --git a/README.md b/README.md index 1d55ac0..b5c8bf9 100644 --- a/README.md +++ b/README.md @@ -1,9 +1,28 @@ - +# Aoven Protocol -# aoven-protocol +**Status:** v0.1-provisional — testable, not finalized. -**Status:** v0.1-provisional. This README is a placeholder pending finalized content from AOV-12b (CanonicalScribe). +Aoven is a controlled-language / epistemic protocol for human–LLM exchanges. It is a fixed set of in-text markers (`[FACT]`, `[HYP]`, `[INTUIT]`, `[NOSRC]`, `[UNCERTAIN]`, `[CONF(level)]`, …) plus minimal formatting rules that force an LLM to label every claim by epistemic type before answering. -For the canonical specification, see [`AOVEN_PROTOCOL_v0.1.md`](./AOVEN_PROTOCOL_v0.1.md). For the empirical test plan, see [`tests/test_plan.md`](./tests/test_plan.md). +## Goal -This file will be replaced by the Scribe-authored README in a follow-up commit. +Reduce four specific failure modes in LLM responses: + +- **Hallucination** — unsourced claims presented as fact. +- **Sycophancy** — user beliefs and emotions confirmed without evidence. +- **Ambiguity** — single interpretations stated as the only reading. +- **Interpretive slippage** — confidence, intuition, or analogy silently upgraded to factual status. + +Whether Aoven actually achieves these reductions is an open empirical question. See `tests/test_plan.md`. + +## Repo contents + +- `AOVEN_PROTOCOL_v0.1.md` — the canonical spec: markers, formats, anti-slippage rules, decision log. +- `AGENTS.md` — governance for agents (human or LLM) working on Aoven. +- `DECISIONS.md` — protocol and process decisions, with reasons and rejected alternatives. +- `docs/archive_exploratoire.md` — earlier conlang-phase terms, archived for traceability. +- `tests/test_plan.md` — A/B test protocol and scoring rubric for empirical validation. + +## Anti-aura rule + +Aoven is not a fantasy conlang, not a poetic dictionary, not a personal project. No seductive prose without a strict definition. If you cannot define a term operationally, it does not belong here. diff --git a/docs/archive_exploratoire.md b/docs/archive_exploratoire.md index 45685cd..c62b03d 100644 --- a/docs/archive_exploratoire.md +++ b/docs/archive_exploratoire.md @@ -1,9 +1,40 @@ - +# Archive exploratoire -# docs/archive_exploratoire.md +**Statut :** phase exploratoire, **non canon principal**. Archivé pour traçabilité historique uniquement. -**Status:** placeholder. Final content owned by AOV-12b (CanonicalScribe). +Les termes ci-dessous proviennent d'une phase pré-protocole de type conlang. Aucun n'est canon Aoven. Décision **D4** dans `AOVEN_PROTOCOL_v0.1.md` les a tous archivés. Règle de gouvernance G-4 : la charge de la preuve est sur la rétention, pas sur le rejet — un terme ne revient en canon que si un agent démontre une fonction épistémique unique non couverte par les 14 marqueurs. -The exploratory archive (Aoa, Aova, Orven, Renavé, Renavé-mu, Renavé-li, Renavé-zo) is currently documented in [`AOVEN_PROTOCOL_v0.1.md`](../AOVEN_PROTOCOL_v0.1.md) under "Exploratory archive" (decision D4 [validated]). The Scribe-authored standalone archive document will be published in a follow-up commit. +Statuts utilisés ici (cf. AOV-1 livrable 6 : archive / reuse / rename / reject / observe) : `archive`, `reuse`, `rename`, `reject`, `observe`. -These terms are pre-protocol conlang phase, not Aoven canon. They may only be reused if an agent demonstrates a unique epistemic function not covered by the 14 markers (burden of proof on retention). +--- + +## Correction v0.1.2 (D9, AOV-15) + +La version précédente de cette archive affirmait que **Renavé-mu / -li / -zo** « servaient de gradients de confiance, supersédés par `CONF(high/medium/low)` ». Cette assertion était une **fabrication NOSRC** : aucun agent n'avait de source sur le sens original de ces termes. Le board a fourni l'origine correcte sur AOV-15 ; le présent document l'intègre verbatim. La correction est journalisée explicitement (et non silencieusement) parce qu'elle relève précisément du type de glissement qu'Aoven existe pour bloquer : l'assertion confiante sans source. Le statut de réutilisation reste inchangé ; seule la description historique est corrigée. + +Source pour les définitions Renavé : entrée du board sur AOV-15, citée en français dans la colonne « Raison » ci-dessous. + +--- + +## Tableau + +| Terme | Phase d'origine | Statut | Raison | +|-------|-----------------|--------|--------| +| **Aoa** | Pré-protocole, conlang | `observe` | Aucune fonction épistémique identifiée. Pas rejeté définitivement ; rétention conditionnée à preuve d'unicité (G-4). | +| **Aova** | Pré-protocole, conlang | `observe` | Aucune fonction épistémique identifiée. Pas rejeté définitivement ; rétention conditionnée à preuve d'unicité (G-4). | +| **Orven** | Pré-protocole, conlang | `observe` | Aucune fonction épistémique identifiée. Pas rejeté définitivement ; rétention conditionnée à preuve d'unicité (G-4). | +| **Renavé** | Pré-protocole, conlang | `observe` | Origine (board, AOV-15) : *relation de présence répétée sans interaction*. Pas un marqueur épistémique ; aucune fonction non couverte par les 14 marqueurs. | +| **Renavé-mu** | Pré-protocole, conlang | `archive` | Origine (board, AOV-15) : *Renavé réciproque*. **Pas un gradient de confiance** ; l'ancienne mention « supersédé par `CONF` » était une fabrication NOSRC corrigée en v0.1.2 (D9). Aucune fonction épistémique non couverte par les 14 marqueurs. | +| **Renavé-li** | Pré-protocole, conlang | `archive` | Origine (board, AOV-15) : *Renavé asymétrique*. **Pas un gradient de confiance** ; l'ancienne mention « supersédé par `CONF` » était une fabrication NOSRC corrigée en v0.1.2 (D9). Aucune fonction épistémique non couverte par les 14 marqueurs. | +| **Renavé-zo** | Pré-protocole, conlang | `archive` | Origine (board, AOV-15) : *sentiment résiduel laissé par la disparition silencieuse d'un Renavé*. **Pas un gradient de confiance** ; l'ancienne mention « supersédé par `CONF` » était une fabrication NOSRC corrigée en v0.1.2 (D9). Aucune fonction épistémique non couverte par les 14 marqueurs (proche de `EMOTION` côté observation, mais sans valeur ajoutée distincte). | + +--- + +## Notes + +- Source des verdicts originaux : `AOVEN_PROTOCOL_v0.1.md` § « Exploratory archive » et décision D4. +- Source de l'origine de la famille Renavé : commentaire du board sur AOV-15, intégré via AOV-16. +- Aoa, Aova, Orven, Renavé restent en `observe` plutôt qu'en `reject` : aucun n'a été examiné en profondeur ; un audit dédié pourrait identifier une fonction non couverte. En l'absence de cet audit, la règle G-4 s'applique : pas canon par défaut. +- Renavé-mu / -li / -zo restent en `archive` (statut inchangé par la correction v0.1.2). La raison archivée n'est plus « gradient de confiance supersédé » — elle est « pas une fonction épistémique non couverte par les 14 marqueurs ». Le verdict opérationnel (hors canon) ne change pas ; seule la description historique est corrigée. +- Aucun terme n'est en `rename` ou `reuse` à ce stade : aucun n'a démontré la fonction unique requise par G-4. +- Pour la trace complète de la correction, voir D9 dans `AOVEN_PROTOCOL_v0.1.md` (et la mirror dans `DECISIONS.md`).