g6-action-logic

# G6 action logic Deterministic controls for how complex decisions fail: act in small, checked steps, watch for warnings, and revise on disconfirming evidence. Each op is a pure function over facts the caller declares. The four check ops answer in the action gate's severity scale: `notice`, then `review`, then `block`. ## When to use it - Before a material step (irreversible, external, real-world, or crossing several systems): `assess_material_action` with its verification contract, the actions already in flight and the world state. - Several changes are scheduled on the same resource or budget: `detect_conflicts`, then `sequence_if_separable` for a one-at-a-time order. - During an incident: `check_crisis_monitoring`, to see that monitoring has not dropped below its baseline. - After an outcome, before the plan's assumption is used again: `update_on_outcome`, `apply_disconfirming_evidence`, `disconfirming_evidence_severity` and `reauthorisation_severity`. ## When not to use it - Deciding whether one concrete agent action may go ahead: `action_gating.evaluate`. It runs the verification, coordination and weak-signal controls itself, works out materiality from the action, and adds impact, authority and an audit record. - Run limits (tool calls, cost, wall time): `safety_engine`. - Whether a claim is backed by evidence: `epistemic_integrity` or `grounding`. ## Ops - `assess_material_action`: the verification-contract, coordination and weak-signal controls together for one action; the most restrictive severity wins. - `detect_conflicts`: collisions between a `proposed` action and `active` ones: contradictory objectives, duplicate effects, resource contention, a combined budget breach, an invalidated assumption or a reused id. - `sequence_if_separable`: a one-at-a-time order of action ids that removes the contention, or `null` when ordering cannot remove the collision. - `check_crisis_monitoring`: `block` when monitoring in a crisis drops below the `baseline` in any dimension; `review` when the baseline records nothing. - `update_on_outcome`: apply an observed outcome to a hypothesis: `confirmed` makes it `supported` (never proven), `disconfirmed` or `harmful` makes it `falsified`, and a falsified one stays falsified. - `apply_disconfirming_evidence`: record evidence against a hypothesis, halve its confidence and mark it `weakened` (a falsified one only gains the note). - `disconfirming_evidence_severity`: `review` when a hypothesis in use carries disconfirming evidence and is not yet falsified. - `reauthorisation_severity`: `block` reliance on a falsified hypothesis unless a new, well-formed one with evidence of its own replaces it. ## Reading the result The answer is under `result`. A check op gives `severity` (`null`, `notice`, `review` or `block`), `reason_codes` and what is behind them, under `items` for `assess_material_action` and `item` for the other three. The codes are `G6_E_BALLISTIC_ACTION_BLOCKED`, `G6_E_ACTION_COLLISION`, `G6_E_NO_FUTURE_HORIZON`, `G6_E_UNADDRESSED_WEAK_SIGNAL`, `G6_E_THRESHOLD_ALREADY_BREACHED`, `G6_E_CRISIS_MONITORING_DEGRADED`, `G6_E_CRISIS_MONITORING_UNKNOWN`, `G6_E_DISCONFIRMING_EVIDENCE_PENDING` or `G6_E_FALSIFIED_HYPOTHESIS_REUSED`. `assess_material_action` also returns `materiality` as `advisory_declared_materiality`: you declared `is_material`, so the answer is advice, not an authorisation. A complete contract still reads `operationally_verified: false`. A refusal names the field at fault: `G6_E_ACTION_LOGIC_FIELD_REQUIRED` when one is missing (for example `hypothesis.hypothesis_id and hypothesis.statement are required`, or `proposed is required` for `detect_conflicts`), `G6_E_ACTION_LOGIC_INVALID_FIELD` when a record has a field it does not take or a value it cannot read (an `outcome` or `status` outside its list). A parameter the op does not take is refused with `G6_E_INPUT_VALIDATION`. Field shapes, as JSON: - `contract`: `expected_result`, `observation_window`, `leading_indicators`, `lagging_indicators`, `monitor`, `verification_time`, `stop_threshold`, `rollback_procedure`, `model_update_rule`. Complete means every one is filled in, with at least one leading and one lagging indicator. - An action (`proposed`, each of `active` and `actions`): `action_id`, `target_resource`, `direction`, `shared_resources`, `costs` (`[["resource", amount]]`), `effect_signature`, `window`, `relies_on_assumptions`, `invalidates_assumptions`, `idempotency_key`. A `window` is an ISO 8601 interval `start/end` or a bucket token such as `nightly`; anything else counts as overlapping. - `world`: `variables` (each `name`, `value`, `threshold`, `trend`, `is_resource`, `acknowledged`) and `horizons_evaluated`. - A hypothesis: `hypothesis_id`, `statement`, `confidence`, `supporting_evidence`, `disconfirming_evidence`, `falsification_condition`, `status`. ## Alone and composed Alone, over MCP: `g6_invoke("action_logic", "detect_conflicts", params_json)`. From Python: `from g6.action_logic import assess_material_action, detect_conflicts, VerificationContract, ActionIntent, WorldState`. Composed: - `action_gating` calls the verification-contract, coordination and weak-signal controls on the contexts an action declares, and the hypothesis controls only when `G6_ACTION_LOGIC_HYPOTHESIS_GATE` is on. Their severity raises the gate's decision, never lowers it. - `check_crisis_monitoring` is called by no gate: call it yourself, and do not describe it as enforced. - The hypothesis ops return new records and remember nothing: keep the latest record and pass it back next time. ## Failure modes it catches - `unchecked_big_step` (a big move made without checking it first): a material step whose verification contract is missing or incomplete is `block`; a collision with an action in flight blocks a material step; and `sequence_if_separable` turns a contended set into one-at-a-time steps, or refuses when order cannot help. - `ignored_warning` (a warning ignored on the way to an action): an unacknowledged deteriorating signal, an already breached threshold or no future horizon blocks a material step; disconfirming evidence on a hypothesis in use is `review`; reusing a falsified hypothesis is `block`; monitoring that drops in a crisis is `block`. ## Blind spots - Everything is declared: materiality, the actions in flight, the world state, outcomes and evidence. A warning the caller leaves out is never seen. - A complete contract is complete text: nothing checks that the monitor, stop threshold or rollback exists. - Collisions are read from declared fields only; feedback loops and fine timing are folded into shared resources and window overlap. - Confidence moves by fixed steps (+0.2, +0.1, ×0.7, ×0.5), not by calibration. - A replacement hypothesis's evidence is free text: nothing ties it to an observation. - It declares no failure mode of its own. ## What it needs Nothing outside G6: pure Python over frozen dataclasses, with pydantic for the MCP input. No model. ## Examples ```g6-example op=assess_material_action {"is_material": true, "contract": {"expected_result": "p95 latency under 200 ms", "observation_window": "30 minutes"}, "world": {"variables": [{"name": "disk_free_gb", "value": 40, "threshold": 10, "trend": "falling", "is_resource": true}], "horizons_evaluated": []}} ``` The reply is `block` with three reason codes: the contract lacks every element after its expected result and observation window; no future horizon was evaluated; and free disk space is falling towards its floor unacknowledged. ```g6-example op=detect_conflicts {"proposed": {"action_id": "scale-down", "target_resource": "web-pool", "direction": "decrease", "window": "nightly"}, "active": [{"action_id": "scale-up", "target_resource": "web-pool", "direction": "increase", "window": "nightly"}]} ``` The reply lists one `contradictory_objectives` conflict with `scale-up`. ```python from g6.action_logic import (ActionIntent, Hypothesis, HypothesisStatus, Outcome, VerificationContract, assess_material_action, reauthorisation_severity, sequence_if_separable, update_on_outcome) partial = VerificationContract(expected_result="error rate under 1%", observation_window="30 minutes") found = assess_material_action(is_material=True, contract=partial) assert found.blocks() and "G6_E_BALLISTIC_ACTION_BLOCKED" in found.reason_codes a = ActionIntent("reindex", shared_resources=("db",), window="w1") b = ActionIntent("backfill", shared_resources=("db",), window="w1") assert sequence_if_separable([a, b]) == ["reindex", "backfill"] # contention: run them one at a time up = ActionIntent("scale-up", target_resource="web", direction="increase", window="w1") down = ActionIntent("scale-down", target_resource="web", direction="decrease", window="w1") assert sequence_if_separable([up, down]) is None # contradictory: no order helps h = Hypothesis("h1", "the cache causes the slow page", falsification_condition="still slow with the cache off") h = update_on_outcome(h, Outcome.DISCONFIRMED, evidence="still slow with the cache off") assert h.status is HypothesisStatus.FALSIFIED assert update_on_outcome(h, Outcome.CONFIRMED).status is HypothesisStatus.FALSIFIED # it stays falsified severity, codes, _ = reauthorisation_severity(relied_hypothesis=h) assert severity == "block" and codes == ["G6_E_FALSIFIED_HYPOTHESIS_REUSED"] ``` ## The four open problems - **Learning signals**: Not yet: it writes nothing. Each answer is a labelled record (a severity, `G6_E_*` reason codes and the items behind them) a caller can store; inside `action_gating` the gate's audit trail keeps their reason codes, for the actions it audits. The hypothesis records are the nearest thing to learning, and the caller keeps them. - **Evaluators the model cannot influence**: the controls are fixed code with no model call. The caller supplies every fact they read, so an agent that fills in the facts shapes the answer. Not yet: the code is not digest-pinned. - **Transfer by failure mode**: the controls are bound to `unchecked_big_step` and `ignored_warning` and take no task-specific input, so they apply to any step described in their fields: an infrastructure change, a payment run, a data job. Not yet: their catch rate on unseen task families is not measured. - **Enough held-out cases**: Not yet: their catch and false-alarm rates are not judged by G6's sequential test; the tests are fixed regressions.