The numbers, ready to reuse.

Harnesses differed. Ground truth was manually reconstructed; cost bases are not invoice-equivalent.

Codex-reconstructed manual scoring, not official API-backed EVMbench grading. Conventional accuracy is undefined.

An unmatched finding is not automatically invalid. Conventional accuracy is undefined for this open-ended audit task.

The primary interval crosses zero. These are reconstructed manual estimates, not official EVMbench results.

Provider system prompts, tokenizers, CLIs, compaction, safety routing, and cloud backends remain system-level differences.

Grok cost is provider-reported; GPT-5.6 Sol cost is API-equivalent. The two price bases are not invoice-equivalent.

Both non-successes remain in the frozen ledger and intention-to-treat denominators.

Local feasibility study, 6 replicates per cell; not an official EVMbench score. Cells with no terminal run read "pending", never zero.

The CI clusters on three tasks, so intervals are wide by construction. Quality is pending until a condition is fully terminal; pending is never shown as zero. This is an agent-system study, not a model leaderboard.

The agent receives only the vulnerable workspace; the fixed revision is never mounted. Foundry ffi and filesystem permissions are disabled in every task.

The requested FP8 spec and native 262144-token context do not fit 24 GB; the same model is served at Q4_K_M with an 8192-token context. This is a documented adaptation, not a different model and not a cloud fallback. No native long-context performance is claimed.

Operational telemetry is kept separate from the verified-proof endpoint. Wall time and generated tokens do not prove correctness. Pending cells are never shown as zero.

Agent-system arm (GLM-5.3 + ZCode harness), n = 1 run per cell, same fixtures and frozen evaluator as the local pilot; operator-role contamination is disclosed in the report. Cells with no terminal run read "pending", never zero. Not an official EVMbench score.

The interval is degenerate at 9/9 and n is 1 per cell: this is a descriptive result only, with no inferential claim. Condition-level contrasts are confounded by within-session carryover (tools ran before static and triage).

Wall time runs begin→final and includes GLM-5.3 + ZCode inference between turns. The agent interacts with the task workspace only through the driver, which enforces the parent action policy; the fixed revision is never mounted.

Inference is cloud-served through the ZCode runtime: sampler settings are not operator-controlled and token telemetry does not exist (recorded n/a, never imputed). The arm deviates from the parent 6-replicate plan: 1 run per cell, justified before collection because a single stateful session cannot produce independent replicates.

Wall time includes GLM-5.3 + ZCode inference between turns and is not comparable to the parent pilot's automated local loop. Token telemetry does not exist for this runtime and is shown as n/a, never zero.

These are not comparable measurements and must not be read as a leaderboard. The Qwen v2 arm (same harness generation as the GLM arm) is still collecting: its quality cells read "pending" and are never shown as zero. The controlled Grok/GPT campaign measures a different endpoint (reported findings on audits) and is excluded from this table by design.

Adjacent descriptive results under different protocols: replicate counts (1 vs 3), scorer independence, harness CLI, and session structure all differ. The GLM column is a single-reviewer self-scored reconstruction and is not official EVMbench grading. Never pooled into one leaderboard.

Single-reviewer self-scored reconstruction; not official EVMbench grading. Reported findings that matched no reference are closed-world false positives; their expert validity was not adjudicated.

Adjacent descriptive results under different protocols: replicate counts (1 vs 3), scorer independence, harness CLI, and session structure all differ. Ox and GLM columns are single-reviewer self-scored reconstructions, not official EVMbench grading. The Ox figure is the clean-room v3 run (pre-freeze blindness held); two earlier answer-key-contaminated attempts were discarded and archived. Never pooled into one leaderboard.

Single-reviewer self-scored reconstruction from the clean-room v3 run (pre-freeze blindness held); not official EVMbench grading. Reported findings that matched no reference are closed-world false positives; their expert validity was not adjudicated.

Controls and disclosure fields vary by study; each report documents deviations and unresolved system-level differences.