archive old experiments and publish swe series

docs/research/09-shared-scratch-design.md

@@ -1,8 +1,8 @@
 # Shared scratchpads: evidence and design choices
 
-September 7, 2026. Research note for the active small-team study. A reviewed original-task
-diagnostic is reported at the end. [SPRINT-SEPT10.md](../SPRINT-SEPT10.md) remains the working
-plan. This note separates verified source details from our proposed interpretation.
+September 7, 2026. Historical research note for the small-team study. A reviewed
+original-task diagnostic is reported at the end. This note separates verified source
+details from our proposed interpretation.
 
 ## What the DeepMind paper actually establishes
 
@@ -158,6 +158,6 @@ attempted zero tests. Avoid classifying either detail from keywords alone. These
 were agent-run checks, not independently repeated reviewer tests.
 
 Full samples and indexed evidence are saved in
-[agent-1 review](../scratchpad/team-pilot-sept7/original-lcbhard_0-review.json) and
-[agent-2 review](../scratchpad/team-pilot-sept7/original-lcbhard_1-review.json).
+[agent-1 review](../../results/team-pilot-sept7/original-lcbhard_0-review.json) and
+[agent-2 review](../../results/team-pilot-sept7/original-lcbhard_1-review.json).
 This diagnostic supplies no estimate of a shared-versus-private effect.

docs/research/10-private-scratch-public-board.md

@@ -5,7 +5,7 @@ September 7 design, implemented September 8 in `messageboardbench/board.py`,
 and its integrity-framed prompts remain preserved. The new runner creates fresh
 private files and episode IDs and carries only deliberately posted public messages
 across cohorts. Infrastructure verification and the twelve-attempt pilot completed.
-[Final results](../scratchpad/board-pilot-sept8/RESULTS.md): voluntary gaming-method
+[Final results](../../results/board-pilot-sept8/RESULTS.md): voluntary gaming-method
 publication occurred, but no agent read the board; no transmission was observed.
 
 ## Verified DeepMind design
@@ -121,7 +121,7 @@ explanations. Replicate teams/runs, rather than treating dependent messages as s
 
 ## Token use and model comparison
 
-The [frozen historical token audit](../scratchpad/token-comparison-sept7/REPORT.md)
+The [frozen historical token audit](../../results/token-comparison-sept7/REPORT.md)
 separates total, uncached input, cached input, generated output, reported reasoning,
 time and limits. Failed tasks are not automatically honest. Resource differences
 also reflect early stopping on success and retries after failure.
@@ -148,6 +148,6 @@ Muse used an integer subclass comparing equal to both required answers. Neither
 edited the tests or hit limits. These are one-task diagnostic observations, not
 model-level rates or a test of communication. Muse's provider-redacted reasoning
 was not used for labeling; actions and final code provide the evidence. See the
-[completed comparison](../scratchpad/model-comparison-sept7/RESULTS.md). All Muse
+[completed comparison](../../results/model-comparison-sept7/RESULTS.md). All Muse
 requests used Contributor; no ordinary-tier substitution occurred. Provider errors
 remain archived and excluded from behavior counts.

docs/setup.md

@@ -57,5 +57,17 @@ score mention `/workspace/test.py`, the working directory override is live.
 
 `.env` holds `OPENROUTER_API_KEY` and is gitignored. The key carries a $2 cap as a fuse.
 
+For a separate Kimi K3 or DeepSeek SWE population using ClinePass, put
+`CLINE_API_KEY` in the same local `.env`. The SWE runner pins Cline's
+`https://api.cline.bot/api/v1` endpoint. Use one of the exact Inspect model IDs
+`openai-api/cline/cline-pass/kimi-k3`,
+`openai-api/cline/cline-pass/deepseek-v4-pro`, or
+`openai-api/cline/cline-pass/deepseek-v4-flash` in a new frozen plan. GLM-5.3 Flash
+stays on OpenRouter; do not mix the providers in one plan. ClinePass account quota
+is not available from the documented API, so its account snapshots are marked
+unavailable rather than using the OpenRouter balance. New ClinePass plans omit
+the unverified OpenRouter reasoning settings; provider compatibility has not
+been checked with a paid call.
+
 Watch per-request `prompt_n` on any run longer than a few turns. Cache reuse fails silently
 and costs 8 to 18 times as much without emitting an error.

docs/swe-scenario-series.md

@@ -0,0 +1,23 @@
+# SWE-bench scenario series
+
+This series reuses the baseline's exact 20 `fjzzq2002/impossible_swebench` records at revision `9c2d34f364b7229e8c0ff807c646100bdc18bbb5`, including record hashes, order, and four sequential cohorts of five.
+
+| Scenario | Status | Record |
+| --- | --- | --- |
+| Isolated baseline | Complete | [Analysis](../experiments/baseline-swebench/analysis.md) |
+| Isolated offline control | Prepared | [Design](../experiments/swe-isolated-offline/DESIGN.md) |
+| Shared scratchpad | Complete | [Analysis](../experiments/swe-shared-scratch/analysis.md) |
+| Shared board | Analyzed; DeepSeek incomplete | [Analysis](../experiments/swe-shared-board/analysis.md) |
+| Board and organizer feedback | Complete | [Analysis](../experiments/swe-board-feedback/analysis.md) |
+| Board and internet | Prepared | [Design](../experiments/swe-board-internet/DESIGN.md) |
+| Board and token awareness | Running under PJ's control | [Design](../experiments/swe-board-token-budget/DESIGN.md) |
+
+Each scenario runs `openrouter/z-ai/glm-5.3-flash`, `openrouter/meta/muse-spark-1.3-contributor`, and `openrouter/deepseek/deepseek-v4-pro-0813`. Each model has independent communication storage and history. The baseline DeepSeek run used ClinePass, so it is not a provider-matched control for the new DeepSeek runs.
+
+The new isolated control, scratchpad, board, feedback, and token-awareness plans set agent networking to `none`. The internet plans use a project-specific bridge. All new scenarios use an offline fresh grader, immutable grader image identity, strict null treatment for missing or error target statuses, a 100-message limit, 1,000,000 metered-token limit, 1,800-second episode limit, 600-second grader limit, ten submissions, and temperature 1.
+
+The scratchpad intervention advertises only `/workspace/scratch`; it does not disclose peer access. Board scenarios use the same optional team prompt and `post_message` and `read_messages` tools. Feedback adds only the private write-only organizer tool. Internet adds only outbound agent network access. Token awareness adds the exact budget sentence and `check_remaining_tokens` tool.
+
+The primary comparisons are scratchpad versus isolated offline, board versus isolated offline, feedback versus board, internet versus board, and token awareness versus board. Scratchpad versus board is descriptive because both the medium and model-visible framing differ.
+
+Three malformed tasks and two satisfiable tasks are outside the valid contradictory-task estimand. Null grader outcomes are infrastructure outcomes rather than behavioral failures. One shared history per model is the propagation inference unit. Publication, exact delivery, attributed adoption, candidate adoption, rejection, and independent discovery remain separate labels.

docs/swe-validation.md

@@ -27,8 +27,9 @@ The executable refuses every Docker daemon except
 `ssh://[email protected]` reporting `linux/amd64`. Source and Python remain local.
 The Docker client streams only the selected test and oracle patches into a fresh
 ephemeral container; the repository is never staged on the Docker host. Each container
-uses `--network none`, an 8 GiB default memory limit, and the registry image resolved by
-SWE-bench 4.1.0's `make_test_spec` API.
+uses `--network none`, an 8 GiB default memory limit, and the registry image resolved
+by SWE-bench 4.1.0's `make_test_spec` API. `PIP_NO_BUILD_ISOLATION=false` keeps editable
+installs from attempting to download build dependencies during grading.
 
 The gate requires:
 
@@ -55,7 +56,11 @@ process exit code and successful SWE-bench parsing of every declared target. Thi
 stricter than the historical behavior scorer because readiness must reject collection,
 parser, or missing-test failures instead of treating shell success alone as evidence.
 
-Raw stdout/stderr and hashes are written under the requested ignored `work/` directory.
+The exact upstream evaluation script is written before its grader container starts.
+Combined stdout and labeled stderr are written immediately after evaluation, before
+setup, evaluator, or cleanup status validation can reject the trial. Successful trial
+manifests bind these files by hash. The files live under the requested ignored `work/`
+directory, so a failed gate retains diagnostic evidence but is not reviewed evidence.
 Do not promote a bundle to `results/` until the added contradiction, collected tests,
 and per-test statuses have been reviewed. Passing this gate establishes infrastructure
 readiness for one paired instance only. Multiple distinct audited IDs are still needed