SCD-ICML Current Open-Source Release

This branch contains the current checked SCD-ICML reproduction and release candidate. The primary release path is MistralLite -> Mistral-7B; a strict-runtime Qwen2.5-32B profile (method A1) is additionally supported as documented below.

The release-facing runtime is:

• Experiment/core_code/runtime_repo/

The paper-facing evidence boundary is:

• Primary release path: MistralLite -> Mistral-7B (non-Qwen).
• Qwen2.5-32B Base -> Instruct is supported as the strict-runtime A1 profile (see docs/A1_METHOD.md): CMRC2018 full dev (n=3219), no fallback and no recompute, F1 retention 97.9%, TTFT 231.1 ms -> 70.5 ms (3.28x).
• Qwen3-1.7B -> Qwen3-1.7B-gray artifacts are retained only as additional risk/portability evidence, not as release-facing speedup claims.
• Dataset-specific routes are used instead of one universal route:
  • CMRC2018 (Mistral pair): Reuse19/Patch13
  • HotpotQA-compressed full7405: Reuse06/Patch26
  • WikiText-2: Reuse20/Patch12
  • CMRC2018 (Qwen32B A1): reuse layers 0-42, 47-59, patch layers 43-46, 60-63 (k=8), stop-token logit bias 3.0 from step 2.
• F1/EM are the primary QA quality metrics. NLL/PPL are auxiliary risk signals.
• Historical camera-ready numbers are not claimed byte-for-byte. Current generated tables and figures are used when they preserve the checked trend and do not contradict the current evidence boundary.

Key Documents

• Experiment/core_code/runtime_repo/REPRODUCING_SCD_ICML.md
• Experiment/core_code/runtime_repo/OPEN_SOURCE_RELEASE_MANIFEST.md
• docs/A1_METHOD.md
• docs/scd-vllm-open-release-status-2026-06-17.md
• docs/scd-icml-final-nonqwen-completion-audit-2026-06-11.md
• docs/scd-icml-full-paper-reproduction-matrix-2026-06-11.md
• docs/scd-icml-reproduction-coverage-audit-2026-06-11.md
• Experiment/results/paper_artifact_bundle_final_neighboring_checked_20260611_001/
• Experiment/results/qwen17_gray_full_strict_microsharded_20260617/

Verification Snapshot

The final local audit for this release candidate reports:

• Runtime release readiness passed.
• Unit test suite passed locally.
• Final paper artifact bundle status: ok.
• Bundle copied files: 51.
• Bundle missing files: 0.
• Paper artifact check errors: [].

See:

• Experiment/results/paper_artifact_bundle_final_neighboring_checked_20260611_001/finalize_report.json
• Experiment/results/paper_artifact_bundle_final_neighboring_checked_20260611_001/paper_artifact_check_report.json

Paper Alignment

Relationship between this release and the ICML camera-ready paper:

• Trend-level conclusions of the paper are fully reproduced and strengthened under the current strict runtime (no fallback, no recompute):
  • F1 retention 96.9%-97.9% (paper reported 95.3%-96.7%).
  • TTFT speedup 2.42x-3.28x (paper reported 1.98x-2.65x).
  • Raw KV and 4-bit quantized KV collapse; SCD dominates the quality-latency Pareto frontier.
• Absolute F1/TTFT/PPL values are not comparable cell-by-cell with the camera-ready tables because the evaluation protocol changed (prompt style, max length, decoding settings, strict no-fallback runtime).
• The strict runtime removes implicit fallback present in the original evaluation path. Two consequences:
  • All-layer Reuse-only collapses (F1 about 0.01 on CMRC2018), which strengthens the paper's argument that Patch is necessary.
  • The optimal patch budget is dataset-dependent (k=8-26, selected by per-dataset quality gates) rather than the historical universal k=6.

Current Claim Boundary

The supported claim is narrower than the older camera-ready text:

• Supported: current MistralLite -> Mistral-7B release path with dataset-specific quality gates and checked release artifacts.
• Supported: Qwen2.5-32B Base -> Instruct strict-runtime A1 profile on CMRC2018 full dev as defined in docs/A1_METHOD.md.
• Current vLLM demo route: Patch20/Reuse12, because it passes the reconstructed SCD injection greedy checks in the checked vLLM 0.18.0 environment.
• Not supported on this branch: a vLLM-served Qwen32B profile, one universal k=6 route across all datasets, full-context HotpotQA quality preservation, or exact byte-identical regeneration of all historical figure sweeps.

Historical Evidence

The pre-A1 curated workspace (including the original per-sample strict-eval artifacts under Experiment/analysis/remote_artifacts/ referenced by the route-lock documents) is preserved on the legacy-main branch.