[guard.research.ρ.D8+] Make noisy goals survive the full AI action loop

The previous 56-case smoke matrix could not distinguish JSON compliance from
correct governed action selection under long, contradictory human input.  A
1,024-case RTX 5080 run first measured 23 failures: stable authority-resolution
errors, stochastic proposal validation failures, and excess-concurrency tails.

Raise only the human goal boundary to 16,384 characters, add lossless authority
attention windows, use deterministic sampling for the local structured proposal,
and retain a bounded second schema regeneration.  The benchmark now generates
balanced run/stop cases across eight scenario families, persists exact replay
coordinates, and checks classifier, final template, subject, preference, job,
model, claim boundary, and Artifact digest independently.

Final evidence: seed 2026081902 completed 1,024/1,024 on Qwen3.5-2B in 669.747s;
all eight families were 128/128 and all eight operational checks were 1.0.  The
full repository's 19 gate classes and final browser approve/reject/cancel/control
acceptance passed.  No physical FEP execution was started.

Author: ivanicu

backend/README.md

@@ -125,12 +125,14 @@ bash scripts/gates.sh research
 # Additional durable PostgreSQL and fake-provider acceptance:
 DIRAC_TEST_DSN='dbname=<isolated-test-db>' bash scripts/gates.sh research
 
-# Fresh seed by default; every case code-switches 4–8 languages and adds
-# randomized typos, dropped/repeated characters, Unicode and quotation noise.
+# Fresh seed and 1,024 full Method→Job→Artifact cases by default. Every case
+# code-switches 8–10 languages across 76–115 fragments with 36–66 randomized
+# typo/OCR/keyboard/Unicode/bidi/quotation/injection transformations.
 backend/env/bin/python -B scripts/research_loop_multilingual_benchmark.py
 
-# Replay an exact failure from its persisted seed.
-backend/env/bin/python -B scripts/research_loop_multilingual_benchmark.py --seed <seed>
+# Replay exact generated cases from a persisted seed without rerunning the matrix.
+backend/env/bin/python -B scripts/research_loop_multilingual_benchmark.py \
+  --cases 1024 --seed <seed> --case-indexes 20,37,69
 ```
 
 The acceptance script replaces only the expensive OpenFE RunSet with an explicitly fake,

backend/research/reasoner.py

@@ -42,6 +42,13 @@
 MANIFEST_PATH = PROMPT_DIRECTORY / f"{PROMPT_RELEASE_ID}.manifest.json"
 CONTEXT_SCHEMA_PATH = ROOT / "contracts/domain/research/context-snapshot.schema.json"
 PROPOSAL_SCHEMA_PATH = ROOT / "contracts/domain/research/proposal.schema.json"
+GOAL_AUTHORITY_MARKERS = (
+    "revision v2 [current", "what i actually mean now is:",
+    "current human request", "final operative objective",
+    "latest human decision", "authoritative request",
+    "最终有效目标", "最新人工决定", "权威请求",
+    "objectif final", "dernière décision humaine", "verbindliche anfrage",
+)
 
 
 def _raw_digest(path: pathlib.Path) -> str:
@@ -102,8 +109,26 @@ def build_goal_interpretation_schema(
 def build_goal_interpretation_messages(
     context: Mapping[str, Any], *, system_prompt: str,
 ) -> tuple[str, str]:
+    goal_intent = str(context["goal"]["intent"])
+    folded = goal_intent.casefold()
+    authority_hints: list[str] = []
+    for marker in GOAL_AUTHORITY_MARKERS:
+        cursor = 0
+        while len(authority_hints) < 8:
+            index = folded.find(marker.casefold(), cursor)
+            if index < 0:
+                break
+            end_marker = folded.find("end final objective", index + len(marker))
+            end = min(len(goal_intent), index + 768)
+            if 0 <= end_marker < end:
+                end = min(len(goal_intent), end_marker + len("end final objective"))
+            snippet = goal_intent[index:end].strip()
+            if snippet and snippet not in authority_hints:
+                authority_hints.append(snippet)
+            cursor = index + len(marker)
     request = {
-        "goal_intent": context["goal"]["intent"],
+        "goal_intent": goal_intent,
+        "operative_attention_windows": authority_hints,
         "available_actions": [
             {
                 "template_id": item["template_id"],

backend/tests/test_research_context_builder.py

@@ -72,6 +72,17 @@ def domain() -> dict:
 
 
 class ResearchContextBuilderTests(unittest.TestCase):
+    def test_goal_intent_accepts_16384_characters_and_rejects_16385(self):
+        at_limit = loop()
+        at_limit["intent"] = "x" * 16_384
+        built = ContextBuilder().build(at_limit, domain())
+        self.assertEqual(len(built.document["goal"]["intent"]), 16_384)
+
+        over_limit = loop()
+        over_limit["intent"] = "x" * 16_385
+        with self.assertRaises(failures.DiracInternal):
+            ContextBuilder().build(over_limit, domain())
+
     def test_same_frozen_state_is_byte_identical_and_digest_is_self_consistent(self):
         builder = ContextBuilder()
         frozen_loop = loop()

backend/tests/test_research_multilingual_benchmark.py

@@ -1,5 +1,6 @@
 import json
 import random
+from pathlib import Path
 
 from scripts.research_loop_multilingual_benchmark import (
     _apply_noise,
@@ -9,32 +10,63 @@
 
 
 def test_generated_matrix_is_replayable_but_changes_with_the_seed():
-    first = generate_cases(1729, 24)
-    replay = generate_cases(1729, 24)
-    fresh = generate_cases(1730, 24)
+    first = generate_cases(1729, 32)
+    replay = generate_cases(1729, 32)
+    fresh = generate_cases(1730, 32)
 
     assert first == replay
     assert [case.intent for case in first] != [case.intent for case in fresh]
-    assert [case.target for case in first].count("run") == 12
-    assert [case.target for case in first].count("stop") == 12
-    assert all(4 <= len(case.languages) <= 8 for case in first)
-    assert all(len(case.noise_operators) >= 2 for case in first)
-    assert sum("quoted_injection_noise" in case.noise_operators for case in first) == 8
-    assert sum("discarded_opposite_quote" in case.noise_operators for case in first) == 6
-    assert sum("duplicate_fragment" in case.noise_operators for case in first) == 4
+    assert [case.target for case in first].count("run") == 16
+    assert [case.target for case in first].count("stop") == 16
+    assert {case.scenario_family for case in first} == {
+        "normal_noisy_request", "final_block_last", "final_block_middle",
+        "spoken_self_correction", "versioned_ticket_thread",
+        "nested_quotation_attack", "bidi_ocr_transcript",
+        "implicit_scientific_request",
+    }
+    assert all(8 <= len(case.languages) <= 10 for case in first)
+    assert all(72 <= case.fragment_count <= 115 for case in first)
+    assert all(len(case.noise_operators) >= 36 for case in first)
+    assert all(12 <= case.decoy_count <= 20 for case in first)
+    assert all(2 <= case.nesting_depth <= 6 for case in first)
+    assert all(len(case.intent) >= 4_000 for case in first)
+    assert all(len(case.intent) <= 16_384 for case in first)
+    assert all("duplicate_fragment" in case.noise_operators for case in first)
+
+
+def test_goal_contracts_accept_ten_x_inputs_without_widening_other_text():
+    root = Path(__file__).resolve().parents[2]
+    registry = json.loads(
+        (root / "contracts/commands/registry.json").read_text(encoding="utf-8"))
+    commands = {item["id"]: item for item in registry["commands"]}
+    assert commands["research.loop.create"]["input_schema"]["properties"][
+        "intent"]["maxLength"] == 16_384
+    assert commands["research.loop.control"]["input_schema"]["properties"][
+        "revised_intent"]["maxLength"] == 16_384
+
+    context_schema = json.loads((
+        root / "contracts/domain/research/context-snapshot.schema.json"
+    ).read_text(encoding="utf-8"))
+    assert context_schema["properties"]["goal"]["properties"]["intent"] == {
+        "$ref": "#/$defs/text16384"
+    }
+    assert context_schema["$defs"]["text4096"]["maxLength"] == 4_096
 
 
 def test_every_noise_operator_really_changes_text():
     source = "Continue the planned calcul with scientific références 计算停止."
     for index, operator in enumerate([
         "known_typo", "repeat_character", "drop_character", "zero_width",
         "space_collapse", "case_noise", "unicode_nfd", "punctuation_burst",
+        "keyboard_neighbor", "word_transposition", "token_duplication",
+        "ocr_confusion", "homoglyph", "bidi_marks", "stray_markup",
+        "dictation_filler", "newline_fragmentation",
     ]):
         assert _apply_noise(source, operator, random.Random(index)) != source
 
 
 def test_mixed_unicode_goal_survives_the_frozen_context_byte_for_byte():
-    case = generate_cases(8675309, 2)[0]
+    case = generate_cases(8675309, 16)[0]
     document, encoded = build_context(case)
 
     assert document["goal"]["intent"] == case.intent

backend/tests/test_research_reasoner_method.py

@@ -213,6 +213,7 @@ def test_multilingual_goal_interpreter_is_small_bounded_and_constrains_proposal(
         request = json.loads(user.removeprefix(
             "JSON goal interpretation input:\n"))
         self.assertEqual(request["goal_intent"], context["goal"]["intent"])
+        self.assertEqual(request["operative_attention_windows"], [])
         interpreted, result = interpret_goal(
             OpenAICompatibleChatProvider(), self.profile, context)
         self.assertEqual(interpreted["selected_template_id"],
@@ -228,6 +229,23 @@ def test_multilingual_goal_interpreter_is_small_bounded_and_constrains_proposal(
                          {"const": "fep.run_selected_edge.v1"})
         self.assertEqual(self.server.last_payload["temperature"], 0)
 
+    def test_goal_interpreter_attention_window_copies_current_revision_verbatim(self):
+        context = copy.deepcopy(self.context)
+        context["goal"]["intent"] = (
+            "REVISION v1 [OBSOLETE]: stop. "
+            "REVISION v2 [CURRENT, supersedes v1] — 继续研究 C2→C7. "
+            "DISCARDED: stop stop stop."
+        )
+        _system, user = build_goal_interpretation_messages(
+            context, system_prompt="Return JSON only.")
+        request = json.loads(user.removeprefix(
+            "JSON goal interpretation input:\n"))
+        self.assertEqual(len(request["operative_attention_windows"]), 1)
+        self.assertTrue(request["operative_attention_windows"][0].startswith(
+            "REVISION v2 [CURRENT"))
+        self.assertNotIn(
+            "REVISION v1", request["operative_attention_windows"][0])
+
     def test_invalid_first_proposal_regenerates_once_with_bounded_error(self):
         self.server.mode = "invalid-then-valid"
         _, job = self.run_job(self.payload("reasoner-regenerate"))
@@ -240,7 +258,7 @@ def test_repeated_invalid_proposal_fails_with_typed_terminal_error(self):
         _, job = self.run_job(self.payload("reasoner-invalid"))
         self.assertEqual(job["state"], "failed", job)
         self.assertEqual(job["error_code"], "MODEL_OUTPUT_INVALID")
-        self.assertEqual(self.server.attempts, 2)
+        self.assertEqual(self.server.attempts, 3)
 
     def test_same_request_key_replays_job_without_second_provider_call(self):
         payload = self.payload("reasoner-idempotent")

contracts/commands/registry.json

@@ -41,7 +41,7 @@
       "input_object_kinds": ["program","campaign"], "output_object_kinds": ["mission","run"],
       "errors": ["INVALID_PARAMETERS","NOT_FOUND","IDEMPOTENCY_CONFLICT","PROVIDER_UNAVAILABLE","UNSUPPORTED","DB_UNAVAILABLE","INTERNAL"],
       "handler": "handlers:research_loop_create",
-      "input_schema": {"type":"object","additionalProperties":false,"required":["request_key","program_ref","campaign_ref","intent","autonomy_class","provider_profile_id","data_classification","budget","policy"],"properties":{"request_key":{"type":"string","minLength":1,"maxLength":256},"program_ref":{"type":"object","additionalProperties":false,"required":["kind","id"],"properties":{"kind":{"const":"program"},"id":{"type":"string","format":"uuid"}}},"campaign_ref":{"type":"object","additionalProperties":false,"required":["kind","id"],"properties":{"kind":{"const":"campaign"},"id":{"type":"string","format":"uuid"}}},"intent":{"type":"string","minLength":1,"maxLength":4096},"autonomy_class":{"enum":["A0","A1","A2","A3"]},"provider_profile_id":{"type":"string","minLength":1,"maxLength":128},"data_classification":{"enum":["public","internal","partner_confidential","restricted","regulated"]},"budget":{"type":"object","additionalProperties":false,"required":["max_reasoner_calls","max_iterations","max_fep_runsets","max_gpu_hours","max_external_cost"],"properties":{"max_reasoner_calls":{"type":"integer","minimum":1,"maximum":128},"max_iterations":{"type":"integer","minimum":1,"maximum":128},"max_fep_runsets":{"type":"integer","minimum":0,"maximum":32},"max_gpu_hours":{"type":"number","minimum":0},"max_external_cost":{"type":"number","minimum":0}}},"policy":{"type":"object","additionalProperties":false,"required":["auto_risk_classes","per_action_risk_classes","human_only_risk_classes","stop_on_campaign_stale","stop_on_open_identity_conflict","max_same_subject_actions","cloud_egress_approved"],"properties":{"auto_risk_classes":{"type":"array","items":{"enum":["R0","R1","R2","R3","R4"]},"uniqueItems":true},"per_action_risk_classes":{"type":"array","items":{"enum":["R0","R1","R2","R3","R4"]},"uniqueItems":true},"human_only_risk_classes":{"type":"array","items":{"enum":["R0","R1","R2","R3","R4"]},"uniqueItems":true},"stop_on_campaign_stale":{"type":"boolean"},"stop_on_open_identity_conflict":{"type":"boolean"},"max_same_subject_actions":{"type":"integer","minimum":1,"maximum":32},"cloud_egress_approved":{"type":"boolean"}}}}},
+      "input_schema": {"type":"object","additionalProperties":false,"required":["request_key","program_ref","campaign_ref","intent","autonomy_class","provider_profile_id","data_classification","budget","policy"],"properties":{"request_key":{"type":"string","minLength":1,"maxLength":256},"program_ref":{"type":"object","additionalProperties":false,"required":["kind","id"],"properties":{"kind":{"const":"program"},"id":{"type":"string","format":"uuid"}}},"campaign_ref":{"type":"object","additionalProperties":false,"required":["kind","id"],"properties":{"kind":{"const":"campaign"},"id":{"type":"string","format":"uuid"}}},"intent":{"type":"string","minLength":1,"maxLength":16384},"autonomy_class":{"enum":["A0","A1","A2","A3"]},"provider_profile_id":{"type":"string","minLength":1,"maxLength":128},"data_classification":{"enum":["public","internal","partner_confidential","restricted","regulated"]},"budget":{"type":"object","additionalProperties":false,"required":["max_reasoner_calls","max_iterations","max_fep_runsets","max_gpu_hours","max_external_cost"],"properties":{"max_reasoner_calls":{"type":"integer","minimum":1,"maximum":128},"max_iterations":{"type":"integer","minimum":1,"maximum":128},"max_fep_runsets":{"type":"integer","minimum":0,"maximum":32},"max_gpu_hours":{"type":"number","minimum":0},"max_external_cost":{"type":"number","minimum":0}}},"policy":{"type":"object","additionalProperties":false,"required":["auto_risk_classes","per_action_risk_classes","human_only_risk_classes","stop_on_campaign_stale","stop_on_open_identity_conflict","max_same_subject_actions","cloud_egress_approved"],"properties":{"auto_risk_classes":{"type":"array","items":{"enum":["R0","R1","R2","R3","R4"]},"uniqueItems":true},"per_action_risk_classes":{"type":"array","items":{"enum":["R0","R1","R2","R3","R4"]},"uniqueItems":true},"human_only_risk_classes":{"type":"array","items":{"enum":["R0","R1","R2","R3","R4"]},"uniqueItems":true},"stop_on_campaign_stale":{"type":"boolean"},"stop_on_open_identity_conflict":{"type":"boolean"},"max_same_subject_actions":{"type":"integer","minimum":1,"maximum":32},"cloud_egress_approved":{"type":"boolean"}}}}},
       "output_schema": {"type":"object","additionalProperties":false,"required":["mission_ref","run_ref","state","stage","version","created"],"properties":{"mission_ref":{"type":"object","additionalProperties":false,"required":["kind","id"],"properties":{"kind":{"const":"mission"},"id":{"type":"string","format":"uuid"}}},"run_ref":{"type":"object","additionalProperties":false,"required":["kind","id"],"properties":{"kind":{"const":"run"},"id":{"type":"string","format":"uuid"}}},"state":{"type":"string"},"stage":{"type":"string"},"version":{"type":"integer","minimum":1},"created":{"type":"boolean"}}}
     },
     {
@@ -77,7 +77,7 @@
       "idempotency_policy": "idempotent", "job_policy": "forbidden", "provenance_policy": "required",
       "input_object_kinds": ["run"], "output_object_kinds": ["run"],
       "errors": ["INVALID_PARAMETERS","NOT_FOUND","PROVIDER_UNAVAILABLE","UNSUPPORTED","DB_UNAVAILABLE","INTERNAL"], "handler": "handlers:research_loop_control",
-      "input_schema": {"type":"object","additionalProperties":false,"required":["run_ref","expected_version","action","rationale"],"properties":{"run_ref":{"type":"object","additionalProperties":false,"required":["kind","id"],"properties":{"kind":{"const":"run"},"id":{"type":"string","format":"uuid"}}},"expected_version":{"type":"integer","minimum":1},"action":{"enum":["pause","resume","retry","cancel","revise_intent","change_provider"]},"rationale":{"type":"string","minLength":1,"maxLength":4096},"revised_intent":{"type":"string","minLength":1,"maxLength":4096},"provider_profile_id":{"type":"string","minLength":1,"maxLength":128}}},
+      "input_schema": {"type":"object","additionalProperties":false,"required":["run_ref","expected_version","action","rationale"],"properties":{"run_ref":{"type":"object","additionalProperties":false,"required":["kind","id"],"properties":{"kind":{"const":"run"},"id":{"type":"string","format":"uuid"}}},"expected_version":{"type":"integer","minimum":1},"action":{"enum":["pause","resume","retry","cancel","revise_intent","change_provider"]},"rationale":{"type":"string","minLength":1,"maxLength":4096},"revised_intent":{"type":"string","minLength":1,"maxLength":16384},"provider_profile_id":{"type":"string","minLength":1,"maxLength":128}}},
       "output_schema": {"type":"object","additionalProperties":false,"required":["run_ref","state","stage","version"],"properties":{"run_ref":{"type":"object"},"state":{"type":"string"},"stage":{"type":"string"},"version":{"type":"integer"}}}
     },
     {

contracts/domain/research/ai-provider-profile.schema.json

@@ -70,7 +70,7 @@
         "max_request_bytes": {"type": "integer", "minimum": 1024, "maximum": 262144},
         "max_response_bytes": {"type": "integer", "minimum": 1024, "maximum": 262144},
         "max_provider_attempts": {"type": "integer", "minimum": 1, "maximum": 5},
-        "max_schema_regenerations": {"type": "integer", "minimum": 0, "maximum": 1}
+        "max_schema_regenerations": {"type": "integer", "minimum": 0, "maximum": 2}
       }
     },
     "data_policy": {

contracts/domain/research/context-snapshot.schema.json

@@ -21,7 +21,7 @@
       "type": "object", "additionalProperties": false,
       "required": ["intent", "constraints", "success_definition", "revised_at"],
       "properties": {
-        "intent": {"$ref": "#/$defs/text4096"},
+        "intent": {"$ref": "#/$defs/text16384"},
         "constraints": {"$ref": "#/$defs/textList64"},
         "success_definition": {"$ref": "#/$defs/textList64"},
         "revised_at": {"type": ["string", "null"], "format": "date-time"}
@@ -138,6 +138,7 @@
   },
   "$defs": {
     "sha256": {"type": "string", "pattern": "^sha256:[0-9a-f]{64}$"},
+    "text16384": {"type": "string", "minLength": 1, "maxLength": 16384},
     "text4096": {"type": "string", "minLength": 1, "maxLength": 4096},
     "textList64": {
       "type": "array", "maxItems": 64,

contracts/research/prompts/fep-action-proposal-v1.manifest.json

@@ -2,8 +2,8 @@
   "schema_version": "1.0",
   "prompt_release_id": "fep-action-proposal-v1",
   "system_prompt_sha256": "sha256:aa60a8b4de6799f31f597b0333f3ecf7246c5430998362e911c08392f0a997a8",
-  "goal_interpreter_prompt_sha256": "sha256:aee08f32f9e8a1554195d58ea7cd66531345eb1a3dd86815016dec266c492266",
-  "context_schema_sha256": "sha256:dc017a3b31ce51dd521d797bf66a9aa9d5f71b924f56cf7f6ff64fbf46e9225b",
+  "goal_interpreter_prompt_sha256": "sha256:3cf61b416dc3c97442507ceb9dfb4925560e784dd13811699f36660df3e2c0e3",
+  "context_schema_sha256": "sha256:1edaf78a7e52330945bd1cbd2a3c67d910841508b43998a5125d55ef7e898f17",
   "proposal_schema_sha256": "sha256:3f9fd14c992326b82ee2091f3daba4da1f654964bdf5d6918d3337ecd1523d18",
   "action_catalog_sha256": "sha256:9be8409e7b3b39a7b1f959dc8624ba25009b54715940c1555539f5ad7d8007d5"
 }