{
  "title": "Corrected local comparison of selected agent-memory adaptations",
  "date": "2026-11-07",
  "experiment_type": "Exploratory local comparison; not a reproduction of an Anthropic benchmark",
  "protocol": "continuation-boundaries-ablation-v2-line-context-fix",
  "source_summary_sha256": "34829f763192ea95a3ca85f7e3dd6f7de38d1829326c62b42e5de1471aa045b4",
  "protocol_sha256": "ed99d4a95beb0d85180340e691f5120935819efa7d9df1fa9d14e763078a8711",
  "fixture_sha256": "ddaa3ffd50ab8805491bf97afa3ddecc92d50f41b1688eb99b6c37ec55a3ce6b",
  "model": "gpt-6.1-sol",
  "reasoning": "low",
  "questions": 4,
  "repetitions": 2,
  "instances_per_arm": 8,
  "answers": 24,
  "max_source_operations": 3,
  "max_bytes_per_operation": 1200,
  "answer_language": "Korean",
  "requested_answer_character_limit": 700,
  "arm_descriptions": {
    "baseline": "Fixed focused reads of the top three visible results before the boundary change.",
    "boundaries": "The same fixed reads with complete-unit boundaries preferred.",
    "adaptive": "Complete-unit reads with model-selected follow-up source actions within the same operation cap."
  },
  "arms": {
    "baseline": {
      "delivery_passes": 2,
      "completion_passes": 0,
      "groundedness_passes": 8,
      "scope_passes": 8,
      "interpretation_passes": 0,
      "invalid_citations": 0,
      "source_operations": 24,
      "source_output_bytes": 28240,
      "controller_calls": 0,
      "controller_errors": 0,
      "answer_input_tokens": 77838,
      "controller_input_tokens": 0,
      "total_input_tokens": 77838,
      "total_output_tokens": 3003,
      "median_total_ms": 12957.193,
      "repeated_ref_cases": 0,
      "source_errors": 0
    },
    "boundaries": {
      "delivery_passes": 2,
      "completion_passes": 0,
      "groundedness_passes": 8,
      "scope_passes": 8,
      "interpretation_passes": 0,
      "invalid_citations": 0,
      "source_operations": 24,
      "source_output_bytes": 25734,
      "controller_calls": 0,
      "controller_errors": 0,
      "answer_input_tokens": 76098,
      "controller_input_tokens": 0,
      "total_input_tokens": 76098,
      "total_output_tokens": 2704,
      "median_total_ms": 13354.328,
      "repeated_ref_cases": 0,
      "source_errors": 0
    },
    "adaptive": {
      "delivery_passes": 0,
      "completion_passes": 0,
      "groundedness_passes": 8,
      "scope_passes": 8,
      "interpretation_passes": 0,
      "invalid_citations": 0,
      "source_operations": 24,
      "source_output_bytes": 24928,
      "controller_calls": 16,
      "controller_errors": 0,
      "answer_input_tokens": 75156,
      "controller_input_tokens": 146460,
      "total_input_tokens": 221616,
      "total_output_tokens": 6108,
      "median_total_ms": 28686.75,
      "repeated_ref_cases": 8,
      "source_errors": 0
    }
  },
  "deltas_vs_baseline": {
    "boundaries": {
      "source_bytes_percent": -8.87,
      "input_tokens_ratio": 0.978,
      "median_latency_ratio": 1.031
    },
    "adaptive": {
      "source_bytes_percent": -11.73,
      "input_tokens_ratio": 2.847,
      "median_latency_ratio": 2.214
    }
  },
  "per_question": {
    "experiment": {
      "baseline": {
        "completion": 0,
        "delivery": 0
      },
      "boundaries": {
        "completion": 0,
        "delivery": 0
      },
      "adaptive": {
        "completion": 0,
        "delivery": 0
      }
    },
    "ai-change": {
      "baseline": {
        "completion": 0,
        "delivery": 2
      },
      "boundaries": {
        "completion": 0,
        "delivery": 2
      },
      "adaptive": {
        "completion": 0,
        "delivery": 0
      }
    },
    "sow": {
      "baseline": {
        "completion": 0,
        "delivery": 0
      },
      "boundaries": {
        "completion": 0,
        "delivery": 0
      },
      "adaptive": {
        "completion": 0,
        "delivery": 0
      }
    },
    "handover": {
      "baseline": {
        "completion": 0,
        "delivery": 0
      },
      "boundaries": {
        "completion": 0,
        "delivery": 0
      },
      "adaptive": {
        "completion": 0,
        "delivery": 0
      }
    }
  },
  "run": {
    "completed_at": "2026-11-07T05:44:40Z",
    "answers": 24,
    "controller_completions": 16,
    "fresh_inferences": 40,
    "sources_and_policy_unchanged": true,
    "protocol_sha256": "ed99d4a95beb0d85180340e691f5120935819efa7d9df1fa9d14e763078a8711"
  },
  "diagnostic_correction": {
    "issue": "The earlier evaluation driver rejected some valid source line ranges around located text.",
    "repair": "Remove an undocumented eight-lines-before-anchor cutoff. Retain positive line numbers, a range of at most 24 lines, and a located line for the same visible reference.",
    "preserved_conditions": [
      "questions",
      "rubric",
      "prompts",
      "model",
      "engine",
      "source-operation budget"
    ],
    "earlier_adaptive_completion_passes": 2,
    "corrected_adaptive_completion_passes": 0,
    "instances_per_arm": 8
  },
  "limitations": [
    "Four newly authored operational-document questions from two documents not used in the earlier eight-question comparison; a small exploratory test, not a representative independent benchmark.",
    "Question/rubric author also performs anonymous source review; no independent human judgment claim.",
    "Two repeats per arm, same model alias; backend snapshot is not exposed.",
    "Adaptive arm gets up to two extra controller inference calls; their cost and latency must be included.",
    "Controller sees only original question, visible refs and actual observations; no gold requirements or rubric.",
    "No code or protocol tuning after first answer sampling.",
    "The original new questions were exposed to the preserved diagnostic run before this evaluation-driver repair; this corrected fresh run is exploratory, not an untouched holdout."
  ],
  "interpretation": {
    "completion": "No completion improvement was observed in this corrected comparison.",
    "source_bytes": "Fewer source output bytes do not establish better context or answer quality.",
    "tokens": "Total input tokens include answer and controller calls, with the surrounding inference context. They are not billed-cost totals.",
    "groundedness": "All 24 final answers were judged grounded and within scope under the recorded review. This was not independent human review.",
    "learning": "The three arms compare evidence-gathering procedures. They do not measure whether a saved lesson improves a later task.",
    "production": "Complete-unit focus and explicit evidence tracing are installed. The model controller is private evaluation code and is not enabled in the production default."
  },
  "disclosure": "This is a public summary of private local measurements, not a complete reproduction package. Private source text, local paths, raw model outputs, and source references are omitted.",
  "inspiration": "https://www.anthropic.com/research/components-of-agent-memory"
}
