{
  "meta": {
    "title": "Weave report extension: published discrete datasets",
    "prepared_at": "2026-11-05",
    "status": "Transcribed published observations, not newly measured production data or illustrative series",
    "source_analysis_sha256": "9423ba58f2d98ae80f5b8c509b2c5eead574adc18933571d5598cf0d097aa165",
    "source_repository_head": "a727c88050fa335fedcb5d4fc68e1feb2b7fd3a0",
    "note": "Values and dataset limitations retained from the verified report analysis. Monthly series are not inferred from static SVG geometry. Q2/Q3 quarterly samples can differ."
  },
  "sources": [
    {
      "id": "q3_report",
      "public_url": "https://weaveos.com/research/q3-2026",
      "role": "Canonical Q3 report, including report-approved discrete values, interpretations and limitations.",
      "sha256": "9d56aca1621303557f4720fd3a27d96bd98f95b717b370434bea87169196ae6a",
      "last_commit": "54a10db8bab3e411faeac75683e8b6657b664fd3 2026-10-07T17:22:11-07:00 Align Q3 report with September Index (#150)"
    },
    {
      "id": "q2_report",
      "public_url": "https://weaveos.com/research/q2-2026",
      "role": "Canonical Q2 report; some comparisons are restated in Q3. Keep original and restated bases explicit.",
      "sha256": "1e1317812eb9fec2c035056279ecbc732d7ae9c3755750fb3820554415dc6739",
      "last_commit": "147b53b2e7351bac11d7e9670ed5e88673d9eb38 2026-09-09T17:09:43-07:00 Migrate Weave Index and research into the main website (#50)"
    },
    {
      "id": "router_benchmarks",
      "public_url": "https://weaveos.com/blog/introducing-weave-router-2-0",
      "role": "Paired benchmark setup, pass@2, pricing baselines and statistical limits.",
      "sha256": "f406bc5226598ddb9f66d5c813dba2e7f25d2f036dc211faba2734400db4cee3",
      "last_commit": "b878a184e3a0352fa87e0210faf03a3c17516768 2026-10-09T10:27:38-07:00 Clarify Router licensing and product copy (#136)"
    }
  ],
  "datasets": {
    "ai_output_share": {
      "id": "ai_output_share",
      "title": "AI-assisted share of merged output value",
      "source_refs": [
        {
          "source_id": "q2_report",
          "section_id": "ai-output-share"
        },
        {
          "source_id": "q3_report",
          "section_id": "ai-output-share"
        }
      ],
      "population": "All merged output value; separate organization-level median for the median column",
      "window": "April, June and September 2026",
      "unit": "percent",
      "values": [
        {
          "month": "2026-04",
          "all_output_percent": 41,
          "median_org_percent": 21
        },
        {
          "month": "2026-06",
          "all_output_percent": 52,
          "median_org_percent": 37
        },
        {
          "month": "2026-09",
          "all_output_percent": 57,
          "median_org_percent": 53
        }
      ],
      "limitations": [
        "Whole-PR attribution: a PR is AI-assisted if a detected AI contributor touched it. This does not mean that fraction of every PR was AI written.",
        "Value share differs from line share and from headcount adoption.",
        "Pooled output share and median-organization share are different aggregations."
      ]
    },
    "skills": {
      "id": "skills",
      "title": "Sessions with and without codified skills",
      "source_refs": [
        {
          "source_id": "q2_report",
          "section_id": "skills"
        },
        {
          "source_id": "q3_report",
          "section_id": "skills"
        }
      ],
      "population": "Observed agent sessions split by whether a codified skill was invoked; unadjusted groups",
      "window": "Q2 vs Q3 2026",
      "unit": "completion percent; average lines added",
      "values": [
        {
          "quarter": "Q2",
          "with_skill_completion_percent": 41.2,
          "without_skill_completion_percent": 35.9,
          "with_skill_mean_lines": 744,
          "without_skill_mean_lines": 138
        },
        {
          "quarter": "Q3",
          "with_skill_completion_percent": 45.3,
          "without_skill_completion_percent": 38.0,
          "with_skill_mean_lines": 630,
          "without_skill_mean_lines": 51
        }
      ],
      "limitations": [
        "No group sizes, uncertainty estimates or task/tool/engineer/org adjustment supplied.",
        "Skills are selected for high-volume workflows; lines are volume, not shipped value.",
        "Q2 median session without a skill added zero lines; the line distribution is highly skewed.",
        "Cannot infer that adding a skill causes a 7.3-point completion improvement."
      ]
    },
    "quality": {
      "id": "quality",
      "title": "Reverts and code-review measures",
      "source_refs": [
        {
          "source_id": "q2_report",
          "section_id": "revert-rate"
        },
        {
          "source_id": "q2_report",
          "section_id": "the-review-economy"
        },
        {
          "source_id": "q3_report",
          "section_id": "revert-rate"
        },
        {
          "source_id": "q3_report",
          "section_id": "the-review-economy"
        }
      ],
      "population": {
        "reverts": "Share of eligible merged PRs reverted; bot/noise exclusions apply.",
        "review": "Completed reviews with explicit review-request events; weekends excluded, nights included.",
        "ai_reviews": "Share of all review events signed by AI."
      },
      "window": "2025 reference points, Q2 and Q3 2026",
      "unit": "percent; hours; counts",
      "values": {
        "revert_rate_percent": [
          {
            "period": "2025 reference",
            "reported_approx": 1.2
          },
          {
            "month": "2026-06",
            "reported_percent": 0.98
          },
          {
            "month": "2026-09",
            "reported_percent": 0.64
          }
        ],
        "review_turnaround_hours": [
          {
            "period": "2025 quarterly peak",
            "reported_approx": 2.5
          },
          {
            "quarter": "Q2",
            "as_recapped_in_q3": 0.85
          },
          {
            "quarter": "Q3",
            "reported": 0.87,
            "reported_minutes_rounded": 52
          }
        ],
        "q2_review_count": 829000,
        "ai_signed_review_percent": [
          {
            "quarter": "Q4 2025",
            "value": 0.07
          },
          {
            "quarter": "Q2 2026",
            "value": 1.1
          },
          {
            "quarter": "Q3 2026",
            "value": 1.8
          }
        ]
      },
      "limitations": [
        "Latest reverts and review turnaround are floors because reverts/reviews can arrive later.",
        "Review series excludes teams without explicit request events, rather than treating them as fast.",
        "Reverts are one quality proxy; no causal quality comparison or complete defect measure.",
        "Q2 monthly review values moved materially between data pulls; original report deliberately emphasizes quarterly floors."
      ]
    },
    "spend_output_concentration": {
      "id": "spend_output_concentration",
      "title": "Top-output spend concentration",
      "source_refs": [
        {
          "source_id": "q2_report",
          "section_id": "cohort-economics"
        },
        {
          "source_id": "q3_report",
          "section_id": "cohort-economics"
        }
      ],
      "population": "Engineers with observed AI spend, ranked by output within quarter",
      "window": "Q2 vs Q3 2026",
      "unit": "daily spend median and relative monthly output",
      "values": {
        "q2": {
          "top_1_daily_spend": 34,
          "bottom_50_daily_spend": 2,
          "reported_top_vs_bottom_monthly_output_multiple": 84
        },
        "q3": {
          "top_1_daily_spend": 145,
          "bottom_50_daily_spend": 4,
          "reported_top_vs_bottom_monthly_output_multiple": 118
        },
        "source_approved_daily_spend_growth": {
          "top_1_multiple": 4.2,
          "bottom_50_multiple": 1.8
        }
      },
      "limitations": [
        "Ranking by output guarantees an output ordering; the spread and spend association are the measured findings.",
        "Daily spend medians and monthly output medians must not be divided to invent ROI.",
        "Reported growth multiples use source precision not recoverable from rounded prose endpoints; preserve source-approved 4.2\u00d7/1.8\u00d7.",
        "Q2 cost/output percentile cut uses output ranks then observed spend, which is a different population from this spend-ranked-by-output cut."
      ]
    },
    "org_size_output": {
      "id": "org_size_output",
      "title": "Monthly output per engineer by organization size",
      "source_refs": [
        {
          "source_id": "q2_report",
          "section_id": "output-benchmarks"
        },
        {
          "source_id": "q3_report",
          "section_id": "output-trend"
        }
      ],
      "population": "Cross-sectional organization-level median and p90; bucketed by active engineer count",
      "window": "Q2 vs Q3 2026",
      "unit": "monthly estimated expert output units per engineer",
      "values": [
        {
          "team_size": "0-10 engineers",
          "q2_median": 55,
          "q3_median": 73,
          "q2_p90": 327,
          "q3_p90": 499
        },
        {
          "team_size": "11-20 engineers",
          "q2_median": 62,
          "q3_median": 116,
          "q2_p90": 206,
          "q3_p90": 459
        },
        {
          "team_size": "21-50 engineers",
          "q2_median": 44,
          "q3_median": 76,
          "q2_p90": 100,
          "q3_p90": 193
        },
        {
          "team_size": "51-100 engineers",
          "q2_median": 44,
          "q3_median": 63,
          "q2_p90": 76,
          "q3_p90": 163
        },
        {
          "team_size": "101+ engineers",
          "q2_median": 29,
          "q3_median": 46,
          "q2_p90": 74,
          "q3_p90": 143
        }
      ],
      "limitations": [
        "This table is not the constant-cohort trend chart; do not splice the two together.",
        "Counts by Q3 bucket and standard errors are not published; Q2 says largest buckets are thinnest.",
        "Changing organizations and sizes can affect quarter-to-quarter comparisons.",
        "p90/median is within-bucket dispersion, not a causal return from organization size."
      ]
    },
    "org_size_operating_benchmarks": {
      "id": "org_size_operating_benchmarks",
      "title": "Spend, review, concentration and expertise by size",
      "source_refs": [
        {
          "source_id": "q2_report",
          "section_id": "benchmark-grid"
        },
        {
          "source_id": "q3_report",
          "section_id": "benchmark-grid"
        }
      ],
      "population": "Organization-level benchmarks by active-engineer size bucket; AI spend denominator is AI-active engineers",
      "window": "Q2 vs Q3 2026",
      "unit": "USD per AI-active engineer per month; rounds; percent",
      "values": {
        "q3_ai_spend_by_size": [
          {
            "size": "0-10",
            "usd": 410
          },
          {
            "size": "11-20",
            "usd": 602
          },
          {
            "size": "21-50",
            "usd": 618
          },
          {
            "size": "51-100",
            "usd": 433
          },
          {
            "size": "101+",
            "usd": 301
          }
        ],
        "q2_ai_spend_published_prose": [
          {
            "size": "11-20",
            "usd": 264
          },
          {
            "size": "101+",
            "usd": 166
          }
        ],
        "review_rounds_above_10_engineers": {
          "q2_range": [
            1.3,
            1.5
          ],
          "q3_range": [
            1.2,
            1.4
          ]
        },
        "top_decile_output_share_above_10_engineers_percent": {
          "q2_range": [
            38,
            43
          ],
          "q3_range": [
            42,
            48
          ]
        },
        "share_holding_half_expertise_percent": {
          "q2_11_20": 24,
          "q2_101plus": 14,
          "q3_11_20": 21,
          "q3_101plus": 13
        },
        "q2_complete_arrays_from_exact_published_chart_labels": {
          "size_buckets": [
            "0-10",
            "11-20",
            "21-50",
            "51-100",
            "101+"
          ],
          "median_AI_spend_per_AI_active_engineer_USD": [
            229,
            264,
            255,
            235,
            166
          ],
          "median_review_rounds_per_PR": [
            0.5,
            1.3,
            1.5,
            1.3,
            1.3
          ],
          "median_top_decile_output_share_percent": [
            40,
            39,
            38,
            42,
            43
          ],
          "median_bus_factor_half_expertise_team_share_percent": [
            null,
            24,
            20,
            16,
            14
          ]
        }
      },
      "limitations": [
        "Q2 expertise uses trailing twelve months through July; this is a different observation window from Q2 shipping metrics.",
        "Source prose supplies ranges, not complete metric-by-size tables. Do not fill missing rows by interpolation.",
        "At 0-10 engineers, top decile rounds up to one engineer (10\u201320%), review rounds are depressed by merges with no recorded review, and bus factor is omitted.",
        "AI-active-employee spend differs from all-employee Index spend."
      ]
    },
    "work_mix": {
      "id": "work_mix",
      "title": "Output mix by work type",
      "source_refs": [
        {
          "source_id": "q2_report",
          "section_id": "work-type-ratios"
        },
        {
          "source_id": "q3_report",
          "section_id": "work-type-ratios"
        }
      ],
      "population": "Platform-wide output, distinct from independently computed org-level medians in size charts",
      "window": "Q2 vs Q3 2026",
      "unit": "percent of output",
      "values": {
        "q2": {
          "feature": 63,
          "ktlo": 20,
          "bug": 17
        },
        "q3": {
          "feature": 56,
          "ktlo": 21,
          "bug": 23
        },
        "q2_by_size_independent_org_median_arrays": {
          "size_buckets": [
            "0-10",
            "11-20",
            "21-50",
            "51-100",
            "101+"
          ],
          "median_feature_share_percent": [
            71,
            63,
            63,
            63,
            60
          ],
          "median_KTLO_share_percent": [
            14,
            20,
            21,
            21,
            24
          ],
          "median_bug_share_percent": [
            10,
            16,
            17,
            17,
            17
          ],
          "platform_output_shares_percent": {
            "Q3_2025": {
              "feature": 54,
              "KTLO": 35,
              "bug": 11
            },
            "Q2_2026": {
              "feature": 63,
              "KTLO": 20,
              "bug": 17
            }
          }
        }
      },
      "limitations": [
        "A higher bug-work share is an allocation label, not a measured defect-introduction rate.",
        "Do not label feature share as total innovation or business value.",
        "Individual organization charts in Q2 use independent org medians and need not add to 100%; these pooled quarter shares do."
      ]
    },
    "router_benchmarks": {
      "id": "router_benchmarks",
      "title": "Router 2.0 controlled coding benchmarks",
      "source_refs": [
        {
          "source_id": "q3_report",
          "section_id": "router-benchmarks"
        },
        {
          "source_id": "router_benchmarks"
        }
      ],
      "population": "Fixed benchmark tasks in Codex harness, two attempts per task; Router and Astra paired, Auto separate run within three days",
      "window": "Published September 9, 2026",
      "unit": "pass@2 percent; USD/attempt; mean minutes/attempt",
      "values": {
        "rows": [
          {
            "benchmark": "Terminal-Bench 4.0 (66 tasks)",
            "system": "Weave Router 2.0",
            "solved_pass_at_2_percent": 62.1,
            "usd_per_attempt": 5.22,
            "minutes_per_attempt": 20.5
          },
          {
            "benchmark": "Terminal-Bench 4.0 (66 tasks)",
            "system": "GPT-6 Astra",
            "solved_pass_at_2_percent": 60.6,
            "usd_per_attempt": 10.03,
            "minutes_per_attempt": 44.1
          },
          {
            "benchmark": "Terminal-Bench 4.0 (66 tasks)",
            "system": "OpenRouter Auto",
            "solved_pass_at_2_percent": 25.0,
            "usd_per_attempt": 5.52,
            "minutes_per_attempt": 36.0
          },
          {
            "benchmark": "SWE-Atlas Codebase QnA (124 tasks)",
            "system": "Weave Router 2.0",
            "solved_pass_at_2_percent": 62.1,
            "usd_per_attempt": 2.31,
            "minutes_per_attempt": 6.8
          },
          {
            "benchmark": "SWE-Atlas Codebase QnA (124 tasks)",
            "system": "GPT-6 Astra",
            "solved_pass_at_2_percent": 66.1,
            "usd_per_attempt": 5.04,
            "minutes_per_attempt": 16.7
          },
          {
            "benchmark": "SWE-Atlas Codebase QnA (124 tasks)",
            "system": "OpenRouter Auto",
            "solved_pass_at_2_percent": 54.0,
            "usd_per_attempt": 1.49,
            "minutes_per_attempt": 5.7
          }
        ],
        "task_counts": {
          "Terminal-Bench 4.0": 66,
          "SWE-Atlas Codebase QnA": 124
        },
        "wilson95_percent": {
          "Terminal-Bench 4.0": {
            "Weave Router": [
              50,
              73
            ],
            "GPT-6 Astra": [
              49,
              72
            ]
          },
          "SWE-Atlas Codebase QnA": {
            "Weave Router": [
              53,
              70
            ],
            "GPT-6 Astra": [
              57,
              74
            ]
          }
        },
        "source_approved_relative_results": {
          "Terminal-Bench 4.0": {
            "router_cost_fraction_of_astra": 0.52,
            "router_speed_multiple": 2.2
          },
          "SWE-Atlas Codebase QnA": {
            "router_cost_fraction_of_astra": 0.46,
            "router_speed_multiple": 2.5
          }
        }
      },
      "limitations": [
        "Source instructs interpreting Router/Astra quality as a tie; overlapping intervals and small task counts.",
        "Pass@2 tasks solved differs from single-attempt pass rate and from production task completion.",
        "k=2 is not a Terminal-Bench leaderboard submission; benchmark requires k\u22655.",
        "Router/Astra paired; Auto comparison is context, not a paired win.",
        "Router uses catalog billing; direct model vendor list prices; Auto uses OpenRouter billing.",
        "Costs/time are per attempt, not cost/time per solved task."
      ]
    },
    "token_inflation": {
      "id": "token_inflation",
      "title": "Tokens consumed per unit of shipped output",
      "source_refs": [
        {
          "source_id": "q2_report",
          "section_id": "token-inflation"
        },
        {
          "source_id": "q3_report",
          "section_id": "token-inflation"
        }
      ],
      "population": "Observed token telemetry divided by shipped output; coverage expanded over the period",
      "window": "January\u2013September 2026",
      "unit": "millions of tokens per output unit",
      "values": {
        "reported_points": [
          {
            "month": "2026-01",
            "value": 0.31
          },
          {
            "month": "2026-05",
            "value": 4.6
          },
          {
            "month": "2026-06",
            "value": 4.4
          },
          {
            "month": "2026-09",
            "value": 6.8
          }
        ],
        "source_approved_multiples": {
          "january_to_june": 14,
          "january_to_september": 21
        }
      },
      "limitations": [
        "Only selected rounded points are published in prose; do not invent intervening monthly values.",
        "Expanded token-telemetry coverage can make an apparent increase even without equivalent behavior change.",
        "Token volume is not spend; cached input is cheaper than fresh input.",
        "September Router cache-read share 53% is router-only, not a platform-wide decomposition."
      ]
    },
    "agent_failures": {
      "id": "agent_failures",
      "title": "Tool failures, spiral detections and syntax validity",
      "source_refs": [
        {
          "source_id": "q2_report",
          "section_id": "failure-taxonomy"
        },
        {
          "source_id": "q3_report",
          "section_id": "spirals"
        }
      ],
      "population": {
        "q2": "Platform agent tool-call telemetry, 3.3M calls; classified and Other failure buckets.",
        "q3": "Router population only: 52 installations, about 234K sessions and 1.62M requests."
      },
      "window": "Q2 and September 2026",
      "unit": "tool calls; failed calls; detector firings; tool-use blocks",
      "values": {
        "q2": {
          "tool_calls_approx": 3300000,
          "failed_calls_approx": 81000,
          "not_found_share_of_all_failures_percent": 41,
          "other_share_of_all_failures_percent": 41,
          "timeout_share_of_all_failures_percent": 3,
          "named_failure_buckets_counts_approx": {
            "Not found": 33500,
            "Other unclassified": 33300,
            "Timeout": 2600,
            "Size limit": 2500,
            "Permission denied": 2200,
            "Invalid args": 1500,
            "VCS conflict": 1300,
            "Rate limit": 1200,
            "Network": 1100,
            "Syntax error": 800,
            "Server error": 600,
            "Interrupted": 300
          }
        },
        "q3": {
          "spiral_detector_firings": 1754,
          "broader_struggle_firings": 2425,
          "reported_session_incidence_approx_percent": 0.5,
          "tool_use_blocks_more_than": 1460000,
          "malformed_blocks": 7,
          "max_model_malformed_rate_percent": 0.0008,
          "largest_installation_request_share_percent": 69
        }
      },
      "limitations": [
        "Detector firings need not equal distinct sessions; source-reported 0.5% is not reproducible by simply dividing firings by rounded 234K sessions.",
        "No-progress detector was introduced in late August, so a pre/post detection trend would mix instrumentation changes.",
        "Malformed tool syntax is not the same outcome as failed tool calls, stalled tasks, bugs or unmerged code.",
        "Q2 and Q3 populations and taxonomies differ; do not call 81K failures\u21927 malformed calls a reliability improvement."
      ]
    }
  }
}
