DeepSeek v4.1 flash 最强逆向工具(无限五代) 求 Star 收藏 ⭐欢迎大家提交项目的改进
12 development and 8 holdout, with 3/2 for each category respectively. registry.mjs validates the input SHA, reference digest, budget, and run records, and does not certify the verdict filled in by the caller.
Build: scripts/evals/build_targets.ps1 -Output , then run node scripts/evals/freeze.mjs <targets.json>. Preserve the actual MSVC/C/PE/MAP digests. verify_targets.mjs uses the real static engine and Unicorn under the original approval to check the return/memory writes of the listed inputs, and saves replay evidence of the single-case investigation recovery. This is not native debugging observation, a whole-domain proof, or autonomous VM recovery.
model_compare.mjs [output] [pilot-evidence.json] executes baseline/brief three times each via the installed DSH SDK. Prepare an independent SDK profile: dsh --profile ig5-eval-v2 --from-default-profile sdk --help. Use IG5_EVAL_PROVIDER/IG5_EVAL_MODEL to specify an existing route; IG5_EVAL_PROVIDER_PATCH can be used to pass in a private overlay for an existing route. Credentials are resolved by the original host, are not written into source code/reports, and no mock models are used.
model_guard.mjs only allows fixed JSON queries, with no paths, commands, or oracle actions; all other tool dispatches are rejected. Actual DSH SDK tests verify rendering, rejection of extra paths, and cross-tool rejection. Runs within the boundary may be marked isolated; runs in an ordinary source workspace still count as exposed. Retaining partitions is not proof of blinding.
Freeze the prompt, dataset, driver/guard summary, budget, model, and provider; save the raw events and every tool request. First run audit_model_compare.mjs to preserve a snapshot of the script under test, then modify the evaluator; existing snapshots support subsequent repeat audits. The audit re-verifies format, return values, position, old evidence markers, current evidence read-back, budget, tokens, and cleanup from the raw assistant/tool events. Null values, missing fields, and extra body text after JSON do not count as success; caller accepted and independent results are counted separately. Format failures are not re-run as replacements; environment/evaluator failures preserve the original batch.
This round completed six real model replays of one development task; the remaining tasks remain unrun. Each fresh session cannot control provider caching, so cacheRead tokens are recorded separately; monetary cost was not calculated, and the success rate of the whole project is not inferred from a single case.
The independent entry point for PR09 is memory_resume.mjs, with three strategies—direct log/LCM/brief+LCM—three times each under the same budget; --models uses the real SDK. It uses synthetic investigations and simulated compression, keeping the statistics separate from the original real static task replays. Run audit_memory_resume.mjs to re-verify conditions, counterexamples, source read-back, and invocation cost from the raw SDK events; environment blocks are retained, and unknown effects are not padded into successes. See ../../docs/LCM_MEMORY.md for details.