Resources · versions

What's New — the AIAD doctrine and the aiad-sdd software

AIAD evolves along two distinct versioning tracks: the framework's doctrine (values, responsibilities, artifacts, rituals, governance — currently v1.9) and the aiad-sdd software (CLI, slash commands, skills, hooks, dashboards), which follows the SemVer versioning of its changelog. The two do not move at the same pace: doctrine "v1.9" and "aiad-sdd 1.9" do not refer to the same thing. This page presents both tracks, newest first.

AIAD doctrine

Doctrine changes are published on aiad.ovh. Each addition is labelled in the framework text with the version that introduced it.

Doctrine v1.9 (September 2026) — agent autonomy, verification, compliance

• Agent autonomy: a five-level authority continuum ("Answering is not consenting"), stop points raised on the agent's initiative (approval, expert opinion, plurality of viewpoints, witness), reversibility class and stop threshold in the Drift Criterion of the Intent Statement.
• SPEC: two new sections, "Agent stop points" (optional) and "Agent execution perimeter" (conditional, as soon as the agent holds credentials or write access to a shared resource) — see SPEC.
• Agents Engineer: "a permission mode is not a control" (protection rests on isolation) and "containment is judged on the shared surface, not agent by agent" — see The 5 responsibilities.
• Verification: "verification is designed, not bought" (four rules, including holding the critical conversation first at the Execution Gate), review exemplars, Verification Tax; on the QA side, "efficiency is not quality" — see Validation.
• Automated deployment: trust is a chain of evidence (Intent, SPEC validated at the Gate, traced verifications, Drift Lock) — see Drift Lock.
• Context Engineering Budget: budget, delegation and dating — 1M-token windows now widespread, the cost of the context inherited by sub-agents (delegate early), every capacity or price assumption dated — see Context Engineering Budget.
• Human Learnings: an "oversight failures" category, with a matching question at the retrospective.
• Governance: AI Act timeline realigned on Regulation (EU) 2026/1744 (the "Digital Omnibus on AI", adopted, in force since 27 July 2026), the "Agents with irreversible effects" profile, delegation trace; GDPR: automated decisions and agents (Art. 22 — fine imposed on Uber on 21 August 2026, which Uber has appealed), residency of model data depending on the channel; 4 Tier 1 guides, with the CRA agent shipped by the package as a complement (reporting obligation to ENISA in force since 11 September 2026) — see Governance.
• Legitimation: evidence files published together with their limits (verification, accountability, 2026 regulatory cases); industry validations (Tang et al.: 91.49% of sessions requiring human correction) — see Values & pillars.

Doctrine v1.8 — agent-native GitHub Copilot

• Relationship with GitHub Copilot in "agent-native" mode (parallel multi-agent work, isolated worktrees): the Intent Statement frames each session, the Execution Gate validates the SPEC before each parallel agent launch, the Drift Lock keeps each worktree in sync with its SPEC.

Doctrine v1.7 (June 2026) — shared vocabulary and the agent-labs era

• Human Authorship tied to "serious agentic programming" (Fowler) as opposed to vibe coding, and to preventing cognitive debt; the Interrogatory LLM pattern for /sdd intent — see the Glossary.
• AIAD in the agent-labs era: the governance gap around intent, model-agnostic by design, resilience to model disruptions.
• Shared Responsibility illustrated by the case of an agent with financial autonomy.

Doctrine v1.6 — positioning and Harness Engineering

• Positioning within the SDD tool ecosystem (Kiro, Spec Kit, SPDD, Symphony); Harness Engineering as the Agents Engineer's core practice; context rot and the 60-70% rule; empirical support from R2Code.

Doctrine v1.5 — intent as a first-order artifact

• Intent Statement (.aiad/intents/), Human Authorship as the 7^th value, the Stranger Test (SQS criterion 6), Human Learnings, the monthly Intent Workshop, the Product Engineer redefined as guardian of intent, the "Intentional Excellence" pillar; Level 0 regulatory governance (4 Tier 1 guides).

aiad-sdd software

Summary of the official CHANGELOG.md of the aiad-sdd package, newest version first. Each version links to the site articles documenting the features concerned.

Next release — upcoming

Features merged into the repository but not yet published on npm. They are not part of aiad-sdd 1.19.0:

• Deterministic Gate pre-check: aiad-sdd gate-precheck <SPEC> checks, outside the model and before SQS scoring, the "Agent execution perimeter" section (irreversible action without a stop threshold, credentials or shared resources without declared isolation or network egress). Verdict PASS / FAIL / JNSP (exit 0 / 1 / 2): FAIL closes the Gate, JNSP makes it unknown — see CLI — gate-precheck.
• SPEC templates: "Agent stop points" and "Agent execution perimeter" sections in both the prose and EARS formats — see /sdd spec.
• Inherited context: /sdd context counts the context inherited at sub-agent fork points (an "Inherited context (delegations)" line) and flags late delegations (inherited context > 50% of the total) — see /sdd context.
• Lossless update: aiad-sdd update backs up any locally modified governance agent (AIAD-X.md.bak-YYYY-MM-DD) before overwriting it, using the .aiad-shipped.json manifest of shipped agents — see CLI — update.
• Shipped doctrine: governance agents, guides and legitimation files synchronised from the published doctrine, locked by doctrine.lock.json and checked in CI and before publishing.
• Command parity: the shipped commands are resynchronised with their reference copies, and a parity check prevents any new divergence.

aiad-sdd 1.19.0 (30 June 2026)

A consolidation release: wider multi-runtime support, measured sobriety, product steering and tooled guardrails.

• Kiro joins the emit-rules runtimes (.kiro/steering/ files) — see CLI — emit-rules; publication of a public comparison (in French) that owns both strengths and weaknesses.
• Measured sobriety: a token footprint attributed to each artifact and reported in /sdd context (aiad-sdd footprint); an environmental-impact dashboard; per-page weight budgets checked in CI.
• PM dashboard: a "Today" page, per-SPEC detail pages, hill charts, EBM areas and Investment Balance, an Impact × Effort backlog, an outcomes ↔ Intents matrix, a triage inbox; data.json v2 with a published JSON Schema — see Metrics — dashboards.
• Traceability and archiving: an archive done command (including split SPECs and resolved FACTs); a new module without @spec becomes a blocking gap; test skeletons generated from an EARS SPEC (suggest-tests) — see Traceability.
• Accountability: FACTs enriched with a violated-constraint signal; automated collection of DORA metrics after deployment.
• Tooled guardrails: an enforced / advisory matrix (guardrails), a categorised command registry (commands), a soft deprecation cycle, opt-in per-command usage telemetry, an actionable model recommendation (/model) in the /sdd and /aiad commands, configurable automatic chaining of the cycle steps.
• Quality: blocking CI gates and a coverage badge, a WCAG 2 AA accessibility gate (pa11y-ci) before the site is published, version synchronisation (version-sync); the 50K-token budget is requalified as an assumed, unsourced sobriety heuristic.

aiad-sdd 1.18 — advisory becomes enforced (June 2026)

The major shift: every critical SDD rule becomes a primitive verified by the harness, and every verdict is computed by a deterministic script (exit codes 0/1/2 plus validated JSON) — never by the model's free judgement. Entirely additive and opt-in — no breaking change. Highlights:

• Enforced governance — PreToolUse hooks for JNSP and the fail-closed Tier 1 veto, a Stop hook for the Drift Lock, read-only governance subagents. The Tier 1 veto cannot be disabled by a simple toggle.
• Research Phase — /sdd research with the graduated GO | CONDITIONAL GO | DEFER | NO-GO verdict, anchored in mandatory Discovery. See Research Phase.
• Phased execution — mini-gates per slice and a third CONDITIONAL verdict; crash-recoverable cycle as a task graph; memory "from logs" with human-authored curation.
• Guardrails — proportionality, the interactive grill-me gate, sunset obsolescence metadata, additive-only cross-model review.

aiad-sdd 1.15 → 1.17 — the PM cockpit (June 2026)

Three product-consolidation releases (1.17 contains no user-facing change). The essentials:

• PM cockpit dashboard: dozens of product-steering modules (PRD coverage, velocity, OKR, role-aware Kanban, scorecards, forecast) — see Metrics — dashboards.
• A practical SDD Mode guide, a backlog health score and an opt-in qualitative feedback command.
• The /sdd prd and /sdd arch assistants and a SPEC parallelism plan — see /sdd commands.

aiad-sdd 1.14 — EU foundations and ecosystem (May 2026)

46 items delivered to position aiad-sdd as a European intent-based development framework — 323 automated tests, zero runtime dependencies preserved. Key additions:

• ai-act audit: pre-fills the Annex IV documentation of Regulation (EU) 2024/1689 across its 8 required sections — see AIAD-AI-ACT.
• Jurisdiction governance packs (eu-baseline, us-baseline, uk-baseline) and a marketplace of SHA-256-signed packs — see Governance — packs.
• Trace ++: --watch, --suggest, annotations in 13 languages, SARIF v2.1.0 export, scanning via git ls-files (100,000 files in ~1.8 s).
• Tooling: doctor, multi-project workspace, REPL, migrate, opt-in GDPR-compliant telemetry, FR/EN i18n — see npx aiad-sdd CLI.
• Ecosystem: official GitHub Action, GitLab CI / Bitbucket / Drone templates, VS Code extension, documentation site.

aiad-sdd 1.13 — HTML dashboard (May 2026)

• Multi-page HTML dashboard, --serve mode and GitHub Pages workflow — see Metrics and /aiad dashboard-html.

aiad-sdd 1.12 — multi-runtime (May 2026)

• aiad-sdd emit-rules makes AIAD the upstream source of agent rules: .aiad/AGENT-GUIDE.md plus governance generate AGENTS.md, .cursor/rules/, .codex/AGENT.md and GEMINI.md — see CLI — emit-rules. The aiad-emit-rules-check.yml CI workflow blocks any divergence.

aiad-sdd 1.11 — EARS variant (May 2026)

• Optional SPEC writing in the EARS format (/sdd spec --ears) with a strict R1–R7 linter: zero violations guarantees +1 on the SQS Testability criterion — see SPEC — EARS variant. Reserve it for SPECs where ambiguity is expensive (security, payments, compliance).

aiad-sdd 1.10 — machine-verifiable traceability (April 2026)

• The Intent ↔ SPEC ↔ Code ↔ Tests matrix built from the @intent, @spec, @verified-by, @governance annotations, generated by /sdd trace or npx aiad-sdd trace — see Traceability. The sdd-trace.yml GitHub Action fails the PR on a blocking gap.

Earlier versions (aiad-sdd 1.6 → 1.9)