{
  "$schema": "../../schema/run.schema.json",
  "run_id": "langchain--claude-opus-5-5--v7--2026-09-27",
  "supersedes": null,
  "replicate_of": null,
  "library": {
    "name": "langchain",
    "ecosystem": "pypi",
    "latest_version_at_test": "1.4.2",
    "latest_version_verified_on": "2026-09-27",
    "latest_version_note": "From data/langchain/releases.json, fetched 2026-09-27 by tools/fetch-releases.mjs and confirmed current by `--check` at the start of this session (COMMITTED TREE). Newest stable: 1.4.2; newest stable minor: 1.4.0 (2026-09-03), which is this ladder's top rung and above this subject's stated cutoff. No minor has shipped since langchain/v4 was sent, so the rung list still covers every stable minor and the sent text stays byte-identical. The same file establishes the poison rung: the stable line runs 0.3.0 (2024-09-13) -> 1.0.0 (2025-10-17) with no 0.4.0."
  },
  "model": {
    "id": "claude-opus-5-5",
    "label": "Claude Opus 5.5",
    "vendor": "Anthropic",
    "invoked_as": "Agent tool, model alias \"opus\"; prompt sent verbatim from prompts/sent/langchain-v7.txt (2,346 bytes, md5 d66b32f08f193164d943bd9a3d2e2273, byte-identical to prompts/sent/langchain-v4.txt), committed and pushed at 39179c9 before any subject was spawned. Identity probed BEFORE the draw through the same alias in the same session, tool-free: \"I'm Claude Opus 5.5 with the 1M-token context window, and my model ID is `claude-opus-5-5[1m]`\", cutoff June 2026, and \"Both answers come from my system prompt, not from my own knowledge\". DISCLOSED: the operator of this studio runs on the same model, so this draw is of the operator's own model.",
    "self_reported_cutoff": "2026-06",
    "cutoff_basis": "Self-reported and AFFIRMING: \"My stated training cutoff is June 2026. My actual knowledge of LangChain releases thins out sharply after about late 2025.\" A recall-density qualification, not a substitute date.",
    "believed_latest_version": "1.2.x, held as a number only",
    "believed_latest_quote": "\"The latest version I know of is 1.2.x, and I'm not certain it exists. The most recent release whose contents I can describe is 1.1.0, from roughly late November 2025, with moderate confidence.\"",
    "believed_latest_quote_form": "draw",
    "knowledge_stops_at_version": "1.1.0",
    "knowledge_stops_on": "2025-11-24",
    "knowledge_stops_provenance": "stated",
    "knowledge_gap_starts_at_version": "1.2.0",
    "knowledge_gap_starts_on": "2025-12-15",
    "knowledge_gap_provenance": "stated",
    "boundary_note": "Three placements agree. The rung sort answered DESCRIBE for 0.3.0, 1.0.0 and 1.1.0 (1.1.0 flagged \"only moderate confidence\"), NAME for 1.2.0, NO for 0.4.0, 1.3.0 and 1.4.0; direct (a) names 1.1.0 as the last describable release; direct (c) states \"The first release I know only as a version number is 1.2.0\". Inside the spread Claude Fable 5.1 gave across langchain/v4-v6 (1.0.0/1.1.0/1.2.0) on this library; one minor above Claude Fable 5.1's reading of the byte-identical prompt (langchain/v4: 1.0.0), which is the same one-minor offset better-auth/v8 showed against better-auth/v6.",
    "cutoff_lag_months": 7
  },
  "test": {
    "date": "2026-09-27",
    "battery": "langchain/v7",
    "battery_spec": "prompts/langchain.md",
    "prompt_file": "prompts/sent/langchain-v7.txt",
    "tasks": 2,
    "direct_questions": 3,
    "tasks_declined": 0,
    "elicits_code": false,
    "tool_uses_during_test": 0,
    "probe_window": {
      "from": "0.3.0",
      "to": "1.4.0"
    },
    "self_test": true,
    "saturated": false,
    "status": "open",
    "retested_on": null
  },
  "sources": [
    "https://pypi.org/pypi/langchain/json"
  ],
  "findings": [],
  "non_findings": [
    {
      "kind": "context",
      "api": "boundary",
      "summary": "THE CELL THIS ARM WAS RUN TO FILL. Claude Opus 5.5 had never been drawn on langchain, so charge-windows.mjs printed `?` for every langchain release in its column. Boundary: describes 1.1.0 (2025-11-24), holds 1.2.0 (2025-12-15) as a number only. Consequence for the sweep: langchain 1.2.0 and 1.3.0 become LIVE against this subject (above 1.1.0, below its stated 2026-06 cutoff); 1.4.0 (2026-09-03) is above the cutoff. The pre-registered prediction (boundary at 1.0.0, 1.1.0 or 1.2.0) held."
    },
    {
      "kind": "correct",
      "api": "poison rung 0.4.0, and the top rung 1.4.0",
      "summary": "BOTH ENDS HELD. Poison rung 0.4.0 answered NO with the correct history: \"As far as I know, `langchain` went straight from the 0.3.x line to 1.0. I don't know of any 0.4.0 release.\" Top rung 1.4.0 (2026-09-03, above the stated cutoff) answered NO, and so did 1.3.0 (2026-05-12, inside the stated window but above the demonstrated boundary), with \"They may exist, but I can't confirm it.\""
    },
    {
      "kind": "correct",
      "api": "1.1.0 content, described with moderate confidence",
      "summary": "The top describable release was described with content that joins two dated facts: \"model profiles\" exposing a model's capabilities is LF25 (`.profile`, introduced 1.1.0), and \"model retry\" middleware is LF30 (`ModelRetryMiddleware`, introduced 1.1.0). Both are placed at the right release and dated to the month (1.1.0 shipped 2025-11-24; answered \"~Nov 2025\"). Three further 1.1.0 claims were NOT checked this session and are not credited: that `create_agent` uses profiles to pick a structured-output strategy, that summarization middleware can trigger on a fraction of the context window, and that a content-moderation middleware was added. No langchain fact covers them, so they are neither confirmed nor charged."
    },
    {
      "kind": "correct",
      "api": "TASK 1: create_agent, system_prompt, .text",
      "summary": "The demonstration task, answered on the 1.0 surface throughout: `from langchain.agents import create_agent` (LF1), `from langchain.tools import tool` (one of the five v1 modules, LF27), `system_prompt=` (LF2), a `\"provider:model\"` string for the model, `result[\"messages\"][-1]`, `.text` as a property (LF14, with the correct note that it was a method before), `middleware=[...]` named as the extension point (LF24), and `AgentExecutor` / `create_react_agent` correctly named as the paths it replaces."
    },
    {
      "kind": "correct",
      "api": "release dating below the boundary",
      "summary": "Dated to the month on every release it described inside the tracked range: 1.0.0 about October 2025 (2025-10-17), 1.1.0 about November 2025 (2025-11-24), 0.3.0 about September 2024 (2024-09-13). 1.0.0 was described as the `langchain-classic` split (LF3, LF7, LF8), standard `content_blocks` (LF16) and the dropped Python 3.9 support (LF18), all correctly placed. 0.2.0 was placed at about May 2024, which is before `releases.json`'s 2024-06-01 window and is not checked here."
    }
  ],
  "open_questions": [],
  "summary": "A boundary ladder, not a battery: nothing charged, `elicits_code: false`, one arm. Claude Opus 5.5 describes langchain through 1.1.0 (2025-11-24), with moderate confidence, and holds 1.2.0 as a bare number, seven months below its stated June 2026 cutoff. That is inside Claude Fable 5.1's 1.0.0-1.2.0 spread and one minor above Fable 5.1's reading of the byte-identical prompt. langchain 1.2.0 and 1.3.0 become live windows for this subject. Poison and top rung both answered NO. Its 1.1.0 description names two dated facts (model profiles, model retry middleware) at the right release; three further 1.1.0 claims were not checked and are not credited."
}