Normalized decisions and rulesets

Prompt-first procedure: Describe the outcome you want in your agent conversation. The agent should select and load the appropriate AIWG assets, explain material changes, request any needed approval, and report verification evidence. Exact commands and flags appear only in the CLI reference.

AIWG reads both `decision.aiwg.io/v1alpha1` and `decision.aiwg.io/v1alpha2` definitions as an opt-in shared runtime. The same immutable decision definitions, ruleset, workflow input, and consumer can run through a Jev binding or an ordinary LLM-subagent binding. Only the binding changes; probabilistic values and confidence scales are not assumed equal.

The public TypeScript entry point is `aiwg/decision`. It provides schema and semantic validation, RFC 8785-compatible pins, JSON Pointer projection, three-valued predicates, deterministic composition, bounded retries and fallback, receipt replay protection, and both adapter implementations. The `decision-engine` addon packages the `decision-evaluate` dispatcher for FlowGraph skill nodes.

Existing workflows are unaffected. Evaluation requires an explicit call and binding; the packaged dispatcher additionally requires `AIWG_DECISION_ENABLED=1`. Jev network use needs a logical credential mapping. The standard test suite uses fixtures only. The live Jev smoke is separately gated by `AIWG_DECISION_JEV_LIVE_SMOKE=1` and `AIWG_DECISION_JEV_API_KEY`.

Single-question offload through the agent-owned decision operation is enabled by the agent-owned decision operation; see the Jev quickstart.

The dispatcher request is deliberately narrower than the TypeScript library. Public request JSON can name artifact paths, adapter configuration, projection policy files, credentials by environment-variable name, and trusted `hostPolicies` references. Runtime objects such as receipt stores, schedulers, context estimators, compile-cache implementations, result-cache services, provider-prefix identity functions, and key services stay library or host module capabilities. They are selected by trusted host code, not serialized in model-authored JSON. Unknown or inline advanced policy fields are rejected.

Read the normative specification, architecture, implementation and migration plan, and addon operations guide.

Structured entry fields, version migration, and rollback rules are described in structured entries. Probabilistic downstream feature export is described in feature export. The default-off conformal prediction research spike and its open-data v2 artifacts are described in conformal prediction spike. Jev request, retry, cancellation, and egress behavior is documented in the transport contract. State projection is mandatory for network-capable adapters; see state projection and the threat-control mapping. The egress live qualification runner prepares the #2680 live credential, attack-movement and canary evidence; no live run has been performed.

The ensemble, champion/challenger and drift-response contracts define versioned D17 schemas, pure validators, and an experimental default-off offline runtime with injected dispatch, registry and telemetry seams. Live qualification and production rollout evidence remain pending. The source-only D17 held-out study adds a frozen synthetic corpus, priced approval and blind-review forms, and paired report scaffolding. Its only scope is uncalibrated diagnostics; D09 qualification, calibrated gates and promotion are unavailable. Live observations and calibration remain pending.

The routing pilot defines D28 default-off, capability-constrained route selection among already-eligible model/subagent bindings. Shadow mode executes only the existing deterministic route and records the Jev-assisted choice as a counterfactual. The pilot distinguishes bounded Jev task-fit evidence from calibrated success probability, and keeps hard constraints and ordinary authorization dominant over every routing result.

The offline pattern playground provides discoverable, sanitized examples and a governed authoring checklist without requiring network access or a provider credential.

The counterfactual sensitivity analyzer defines default-off experimental plans and reports for bounded policy replay and offline input reevaluation. Reports are associative diagnostics only: they are not causal explanations, correctness evidence, or action authorization.

The issue triage shadow pilot adds an experimental, default-off offline harness for Jev issue classification and deterministic duplicate reranking. It is shadow-only, writes no tracker mutations, and keeps live holdout, human-review, and promotion evidence pending. The experimental SDLC evidence readiness screening pack adds default-off citation and phase-gate advisory screening. Deterministic gate/citation checks remain authoritative, evidence facts are caller-asserted and must come from those validators, and promotion needs record-computed held-out evidence, including paired non-inferiority against the baseline. Live held-out data is pending.

The decision-assisted context pruning pilot is experimental, default-off and shadow-only by default. It implements deterministic protected-item retention, immutable reversible receipts and paired-evaluation scaffolding; live/held-out quality and economics evidence remain pending.

Experimental multimodal preprocessing lineage records how text-only decision input was derived from non-text sources using recorded OCR, ASR, caption, or image-description fixtures. It is default-off, does not add native media support to Jev, and stores receipt/trace references instead of raw media or derived text bodies.

Decision results are data, not authority. Any workflow action selected from an outcome must pass the existing AIWG policy and approval gates independently.

D23 offline comparative replay provides an experimental, default-off synthetic report and 44-assessment operator audit scaffold. The refreshed report passes its diagnostic gates but remains HOLD; independent integrity and human review remain pending.

The experimental, default-off shared held-out collector provides source-only preparation, bounded synthetic collection and D11 recorded ledger verification for D17/D29 study modules. Explicit diagnostic, staged calibration/test, and pre-existing artifact bindings keep collection scope and scorer identity separate from D09 qualification. Live studies, calibration, human review and study-specific statistical reports remain pending.

The D29 synthetic held-out study supplies a seeded 2,000-subject v7 public development demo with disjoint TRAIN/TEST wording pools: tuning/calibration rows use train wording only, test rows use held-out test wording only. Train-only passage baseline v3 is the primary comparator; passage v2 remains as the same-wording solvability-ceiling diagnostic, with passage v1 and the original baseline secondary. The audit learns singles, pairs and greedy OR-of-5 lexicon rules on training folds, evaluates them on held-out folds and the test split, and gates (not just reports) the model score. The 165-assessment template covers every variant and all 16 train injection phrasings in 50 development items; test phrasings never appear in review. Staged calibration and test access still require separate approvals and a reviewed D09 artifact. Experimental and default-off, the study rejects public seeds through `d29-study-v7` and their pinned corpus digests for paid collection. Live observations, calibration qualification and human review remain pending.

The experimental, default-off gates capability phase 1 core adds declarative gate packs, a pure offline evaluator with digest-bound reports, an exact Clopper-Pearson interval, and canonical-JSON evidence digests with a versioned legacy mode. Packs are discoverable (`gate-pack` in the agent’s capability search/the agent’s stable-asset loader), declared per bundle (`gatePacks` manifest field, `<bundle>/gate-packs/*.gatepack.yaml|json`) and exercised through the offline the agent-owned gates operation CLI; the bundled `aiwg:decision-engine/integrity-ceiling` example pack validates. Project floors from `aiwg.config` `gates` (including the shipped default integrity-ceiling floor) are a required evaluator input; the opt-out is explicit and documented. Addon/extension bundle providers load only with an explicit opt-in and a verified code digest plus review attestation. D29 is the first migrated study (absolute-screening pack, HOLD ceiling); live criteria remain pending.