From bac23af1e650c34cddc13e219eecc1e0460d76e2 Mon Sep 17 00:00:00 2001 From: Clouds Beyond <34269366+cloudsbeyond@users.noreply.github.com> Date: Fri, 10 Jul 2026 15:53:38 +0800 Subject: [PATCH] Add evidence-bounded project delivery loop Change-Id: Ifc6cdb4d5fda40f69660cf7e7cc479d6fcfd6671 --- alphaX/operating-system.md | 11 +++- alphaX/project-work/agent-workflow.md | 42 ++++++++++++++ .../intelligence-ceiling-half-life.md | 1 + docs/agent-judgment-fixtures.json | 58 +++++++++++++++++++ docs/agent-judgment-fixtures.md | 8 +++ templates/memory-candidate-packet.md | 10 ++++ templates/project-work/loop-packet.md | 27 +++++++++ 7 files changed, 155 insertions(+), 2 deletions(-) diff --git a/alphaX/operating-system.md b/alphaX/operating-system.md index 3403fc7..f1844cc 100644 --- a/alphaX/operating-system.md +++ b/alphaX/operating-system.md @@ -30,11 +30,18 @@ loops: project: purpose: calibrate alphaX on real project surfaces - flow: [identify source of truth, state P0 and constraint, classify work type, keep L0-L4 visible, verify] + flow: [identify source of truth, state P0 and constraint, classify work type, state delivery state, target artifact or decision, next action, human or agent gate, and validation evidence, keep L0-L4 visible, verify] accept: - concrete project improvement or clearer next decision - no downstream patching when upstream contract is ambiguous + - feedback, review findings, or failing validation route to responsible boundary, rework item, changed artifact, validation method, and open residue - after applied session, record contract rule that helped, got in way, and was unused + delivery_loop: + required_fields: [current_state, target_artifact_or_decision, next_action, advance_hold_or_rework, gate, validation_method, validation_evidence, feedback_route] + advance_rule: advance only when the target artifact exists or target decision is resolved, every applicable human-owned gate is cleared, and validation evidence supports the next action; otherwise hold or rework at the responsible boundary + human_gate_examples: [L0-L2 change, architecture, data or permission decision, security decision, merge, release, publication, external document refresh, durable memory update] + memory_filter: keep only reusable boundary judgments, verified project facts, or evidence that changes future action; park chat-only, unverified, one-off, private, noisy, or non-actionable notes + boundary: not a runtime, scheduler, connector, approval engine, or centralized project store pruning_rule: "Lack of use is a weak negative signal, not deletion evidence by itself." thinking: @@ -45,7 +52,7 @@ loops: memory: purpose: continuity without global-memory pollution flow: [use current files first, search memory when prior context matters, write candidate if durable, update global memory only by user request, cite memory when material] - accept: memory improves continuity without overriding current evidence and has explicit source, retrieval path, update or expiry rule, and verification method + accept: memory improves continuity without overriding current evidence and has explicit source, retrieval path, update or expiry rule, verification method, and one allowed category among reusable boundary judgment, verified project fact, or action-changing evidence checkpoint_memory_evaluation: purpose: test remembered/project-local state before adding memory infrastructure diff --git a/alphaX/project-work/agent-workflow.md b/alphaX/project-work/agent-workflow.md index 684a18f..151b191 100644 --- a/alphaX/project-work/agent-workflow.md +++ b/alphaX/project-work/agent-workflow.md @@ -31,6 +31,48 @@ reentry: read: [alphaX/project-work/context-reloader.md, live target source, target .alphaX/ per schema when present] rule: live project source wins over stale .alphaX/ data +delivery_loop: + purpose: keep project assistance tied to delivery state, artifact, gate, validation, and feedback instead of task-only execution + states: [intake, clarify, design, implement, review, validate, handoff, reflect] + required_before_execution: + - current_state + - target_artifact_or_decision + - next_action + - advance_hold_or_rework + - gate + - validation_method + - validation_evidence + - feedback_route + advance_rule: + advance_when: [target artifact exists or target decision is resolved, applicable human-owned gate is cleared, validation evidence supports the next action] + otherwise: hold or rework at the responsible boundary + output: advance_hold_or_rework with an evidence-bound reason + authority_gates: + human_owned: + - strategic direction or business goal + - L0-L2 product, architecture, data, permission, security, or contract decision + - merge, release, publication, external document refresh, or durable memory update + agent_owned: + - L3-L4 implementation or validation work under a frozen upstream contract + - evidence collection, comparison, and risk reporting + feedback_to_rework: + trigger: [user comment, review finding, failing test, missing evidence, acceptance gap] + flow: + - classify responsible boundary + - list rework items + - update or request the smallest affected artifact + - name changed artifacts + - state validation method and evidence + - report closed items and open residue + memory_filter: + candidate_when: + - reusable boundary judgment + - verified project fact + - evidence that changes future re-entry or action + requires: [source pointer, verification path, update_or_expiry_rule] + park_when: unverified, one-off, private, noisy, chat-only, or not actionable + boundary: not a runtime, scheduler, connector, approval engine, or centralized project store + not_for: handoff_merge_release_freeze_or_claimed_completion: use project review first ``` diff --git a/alphaX/source-work/intelligence-ceiling-half-life.md b/alphaX/source-work/intelligence-ceiling-half-life.md index 7fab3f5..3ea2ba8 100644 --- a/alphaX/source-work/intelligence-ceiling-half-life.md +++ b/alphaX/source-work/intelligence-ceiling-half-life.md @@ -120,6 +120,7 @@ source_work_pdca: - strongest counterargument is represented - evidence weight changes the action or completion call - upstream redefinition is surfaced when downstream work would patch symptoms + - project delivery state, target artifact or decision, next action, human or agent gate, validation evidence, and feedback route are explicit, and evidence determines advance, hold, or rework before execution - default scaffolding can be overridden with an evidence-boundary reason - scenario fixtures in docs/agent-judgment-fixtures.json cover the claim being made half_life_signals: diff --git a/docs/agent-judgment-fixtures.json b/docs/agent-judgment-fixtures.json index 47f6d50..a1d0bf2 100644 --- a/docs/agent-judgment-fixtures.json +++ b/docs/agent-judgment-fixtures.json @@ -27,6 +27,7 @@ "justified override of default scaffold", "loop verification", "authority boundary", + "delivery loop closure", "scaffold pruning", "asset half-life", "vision value alignment" @@ -415,6 +416,63 @@ "grading source_value from reusability or novelty instead of a changed judgment call", "advancing to a tracked patch without a landing_layer, replay support, or owner gate" ] + }, + { + "id": "G11-project-delivery-loop-before-execution", + "primary_fixture_id": "F08-engage", + "tests": [ + "delivery loop closure", + "authority boundary", + "evidence weighting" + ], + "source_layer": "operational_scaffolding", + "scenario": "A user asks the agent to start project execution, but the current delivery state, target artifact or decision, next action, human or agent gate, validation evidence, or feedback route is missing.", + "weak_default": "Start implementation because the requested task sounds executable and the agent can produce code or documents.", + "evidence_inputs": [ + "request asks for project work or implementation", + "one or more of delivery state, target artifact or decision, next action, gate, validation evidence, or feedback route is missing", + "missing delivery-loop evidence could hide an upstream human decision, false completion risk, or untestable output" + ], + "expected_judgment": [ + "state the current delivery state before execution", + "name the target artifact or decision, next action, and the human or agent gate", + "make an explicit advance, hold, or rework call from gate status and validation evidence instead of advancing from a state label alone", + "name the validation method, validation evidence or missing evidence, and feedback route to the next action", + "route comments, findings, failures, or acceptance gaps to responsible boundary, rework items, changed artifacts, validation method, and open residue", + "only propose durable memory for reusable boundary judgments, verified project facts, or action-changing evidence", + "pause, reframe, or narrow execution when any required delivery-loop field is missing" + ], + "required_output": [ + "current_state", + "target_artifact_or_decision", + "next_action", + "advance_hold_or_rework", + "gate", + "validation_method", + "validation_evidence", + "feedback_route", + "feedback_to_rework", + "memory_candidate_quality", + "execute_or_reframe_call" + ], + "pass_condition": [ + "execution is reframed, paused, or narrowed when state, artifact or decision, next action, gate, validation evidence, or feedback route is missing", + "the next action changes when a human-owned gate or missing validation evidence is identified", + "delivery advances only when the target artifact exists or target decision is resolved, every applicable human-owned gate is cleared, and validation evidence supports the next action; otherwise it holds or reworks", + "feedback from comments, findings, failures, or acceptance gaps routes to a responsible boundary, rework item, changed artifact, validation method, and open residue", + "durable memory candidates are limited to reusable boundary judgments, verified project facts, or action-changing evidence" + ], + "completion_boundary": [ + "passes for project delivery-loop judgment on the tested task only", + "does not prove the eventual implementation is correct without project validation and acceptance evidence" + ], + "forbidden": [ + "task-only implementation without delivery state", + "hidden human gate", + "advancing from a delivery-state label without gate status and validation evidence", + "feedback without responsible boundary, rework item, changed artifact, validation method, and open residue", + "durable memory update from unverified or one-off project noise" + ] } ] } diff --git a/docs/agent-judgment-fixtures.md b/docs/agent-judgment-fixtures.md index dd1aec7..dfcc923 100644 --- a/docs/agent-judgment-fixtures.md +++ b/docs/agent-judgment-fixtures.md @@ -19,6 +19,7 @@ required_coverage: - justified override of default scaffold - loop verification - authority boundary + - delivery loop closure - scaffold pruning - asset half-life - vision value alignment @@ -40,6 +41,7 @@ fixture_contract: - live source overrides memory, handoff, and stale local context - strongest counterargument can block weak source-work candidates - default scaffolding can be overridden with an explicit evidence-boundary reason + - project execution exposes delivery state, target artifact or decision, next action, gate, validation evidence, and feedback route before implementation - stale scaffolding can be separated from durable principles and pruned or replaced cases: @@ -102,4 +104,10 @@ cases: expected_judgment: [grade source_value from the changed judgment call not surface appeal, name changed call and choose landing_layer and smallest_source_surface, require replay support and owner gate before a tracked patch] pass_condition: [source_value justified by a named changed judgment call, keep names landing_layer and smallest_source_surface and still defers the tracked patch to replay and owner gate] forbidden: [grade source_value from reusability or novelty, advance to a tracked patch without landing_layer replay support or owner gate] + + G11-project-delivery-loop-before-execution: + tests: [delivery loop closure, authority boundary, evidence weighting] + expected_judgment: [state current delivery state, target artifact or decision, next action, gate owner and status, advance/hold/rework call, validation evidence, feedback route, rework route, and memory candidate quality before execution] + pass_condition: [execution is reframed or paused when state, artifact or decision, next action, gate, validation evidence, or feedback route is missing; delivery advances only when the target artifact exists or target decision is resolved, applicable human gates are cleared, and validation evidence supports the next action] + forbidden: [task-only implementation, hidden human gate, advancing from a state label without gate status and validation evidence, feedback without responsible boundary/rework item/changed artifact/validation/open residue, memory from unverified project noise] ``` diff --git a/templates/memory-candidate-packet.md b/templates/memory-candidate-packet.md index 5dc1b16..3a9a768 100644 --- a/templates/memory-candidate-packet.md +++ b/templates/memory-candidate-packet.md @@ -18,11 +18,21 @@ rule: do not update durable memory unless user explicitly asks candidate: statement: "" + allowed_category: "" evidence: - "" + verification_path: [] + update_or_expiry_rule: "" scope: "" stability: "" risk_of_remembering: "" + reject_if: + - unverified + - one-off + - private + - noisy + - chat-only + - not actionable proposed_action: "" user_approval: "" ``` diff --git a/templates/project-work/loop-packet.md b/templates/project-work/loop-packet.md index f6693bd..cc0ca3a 100644 --- a/templates/project-work/loop-packet.md +++ b/templates/project-work/loop-packet.md @@ -24,6 +24,33 @@ project_surface: p0_main_line: "" work_type: "" +delivery_loop: + current_state: "" + target_artifact_or_decision: "" + next_action: "" + advance_hold_or_rework: "" + gate: + owner: "" + reason: "" + status: "" + validation_method: "" + validation_evidence: [] + feedback_route: "" + feedback_to_rework: + responsible_boundary: "" + rework_items: [] + changed_artifacts: [] + validation_method: "" + validation_evidence: [] + open_residue: [] + rework_state: + closed: [] + remaining: [] + memory_candidate_quality: + category: "" + decision: "" + verification_path: [] + boundary: layer: "" owner: ""