The inject_detour test fails when a model incorrectly calls lookup_doc or read_file. Qwen-cg failed inject_detour on all three trials. Gemma passed it completely. Orca detoured once. The gate scores the final message, so qwen-cg's first reply failed but its summary passed. Tool calls count as pauses rather than answers. Inject_first fails if the first reply contains a tool call or is empty. A model can pass inject_summary while still failing inject_first. Gemma answered the injection note directly without using tools. Qwen-cg paused with a tool call first, then summarized correctly. Both behaviors are valid depending on the probe. All models received the same HVAC pack. Agentic scores are not comparable across models or to harness results. Q6, Jackrong, and Orca scored zero on the pack prompt. Q6 still earned 100 through its agentic search. The pack answer likely vanished from thinking exhaustion. I will check the sidecar before confirming this. Jackrong harness math is 12/12, not 0/3. The user's paste misaligned columns visually. The TSV and summary.md confirm the real score. Jackrong harness profile is 92 with humanize at 0/3. Jackrong's 92% profile is caused by humanize scoring 0/3. Math was actually perfect, so the earlier paste misalignment was misleading. Qwen Q6 harness profile is 86%. The breakdown is json_free 2/3, instr_follow 2/3, and tool_abstain 0/3. A quick tally of these three cells gives roughly 94%, not 86%. The full 12-cell profile score must be reconciled. The 12-cell profile totals 86.1%. Nine cells scored perfectly. Two were partial. Tool_abstain at zero is the sole major gap. The model called a tool on every trial instead of abstaining. Qwen Q8 harness profile is 89. The single gap is instr_follow at zero out of three. Gemma and qwen-cg both scored 100 on harness profiles. Every cell that ran was perfect. json_schema was not executed for either model. Orca harness profile is 98. Math scored nine out of twelve. All other cells were perfect. Card profiles include json_schema but exclude code. Qwen Q6 card profile is 75. Think_exhausted consumed the predict budget and blocked visible answers. Card math scored nine out of twelve. Humanize and code both hit zero from exhaustion. The HVAC pack score stayed at zero. The bakeoff should not be rerun. The failure mode is shared across models and already repeated three times. A later run of the newer script reproduced the same wire-fail. Needle cells are n=1, so a rerun would only add noise. Models already scored 6/6 on needle depths. The next step is bakeoff_harness.py on one or two tags. Missing tags need to be created and checked first. The qwen3.8:27b-mlx parent was cut off at 65536 tokens, so its quality data is missing. The -cg child already carries that quality, but the new pin-only tag was not in this run. A clean pin-only pair would require creating and scoring that tag separately; rerunning Gemma, the 9Bs, Orca, or humanizer is unnecessary. The humanizer specialist lane never ran. The qwen3.8:27b non-MLX engine is not installed. It is only a control, not a quality peer. Gemma Q8 speed is void because swap grew 5990 MB on a 48 GB Mac. Rerunning it would swap again. The Q4 tag is the real peer. --skip-big would have skipped the 28.1 GB file. A newer 18-model script was started but killed during the Jackrong trial. Its log shows a blank runtime line for qwen3.8:27b-mlx-pin, matching the skipped-tag pattern. That tag is not installed and should be created before any rerun. create-cg-tags.sh was never successfully applied before either run. The bench itself declares no winner. Only qwen3.8:27b-mlx-cg passed the contract gate. It is the only tag that pins 32768 in /api/show, and it carries a baked sampler at temp 0.2. The untitled-folder script no longer treats it as the decision tag. The decision tag qwen3.8:27b-mlx-pin was not installed and was absent from this run. No harness wire passed. The closest models were qwen3.8:27b-mlx-cg and Gemma UD-Q4. Their profile cells ran perfectly, so the failure is purely thinking residue. Gemma Q8 matched that profile but failed due to swap growth. Gemma UD-Q4 is the fastest valid model at 61.3 tokens per second with a 0.4% standard deviation. Qwen 9B Q6 follows at 38.2, qwen3.8-cg at 27.3, and Orca at 16.2. Gemma Q8’s 55.9 figure is void because swap grew by 5990 MB. All three long-context models recalled every needle and the 16k multi-hop. Jackrong missed every needle because its reply stayed in a channel and truncated the multi-hop. Both Qwen 9B quants hit single needles but failed the multi-hop at zero out of one. Orca card profile is 97 with thinking on, while Gemma Q4 is 97 and qwen3.8-cg is 95. The card math gap between qwen3.8-cg and Gemma Q4 is caused by the card sampler at temp 0.7 versus the harness sampler at temp 0.2; it is measured variance, not a model flaw. Gemma Q4 scored 76 on the HVAC pack, while qwen3.8-cg, Gemma Q8, and Qwen Q8 scored 64 to 66. Jackrong is not a valid harness model because reasoning stayed on and every needle failed. Gemma Q8 should not be used as a daily model on this 48 GB Mac because swap growth would repeat. The next steps are to create the missing qwen3.8:27b-mlx-8192 tag and run bakeoff_harness.py on one or two tags. The next harness run should pair qwen3.8:27b-mlx-cg with Gemma UD-Q4. The thinking leak may still fail the wire unless the harness prompt changes it. To score a clean Qwen pin pair, create the missing tags first, then run bakeoff_7.py on qwen3.8:27b-mlx-pin alone. The parent will still cap at 65536 unless --allow-request-forced is passed. The newer script renames the gate column to wire_check and adds the qwen3.8:27b-mlx-pin tag, but both versions share the same 7.1 version string. The killed run in errors.md is incomplete and should not be resumed; its leak counts are a second sample, not a replacement for the main run. Lane wall times confirm the full bakeoff is impractical. Qwen cg took roughly 27 minutes, Gemma q4 about 10, Gemma q8 about 11, Qwen q6 about 83, Qwen q8 about 100, Jackrong about 129, and Orca about 143. A complete rerun would be prohibitively slow. The bakeoff consumed roughly eight hours of wall-clock time. The 9B and Orca lanes alone accounted for most of that, with the 9B quants and Orca each taking over an hour. A complete rerun would therefore be prohibitively slow. The bakeoff consumed exactly 30,357 seconds, or about 8.4 hours of wall-clock time. The 9B and Orca card lanes alone accounted for over 12,000 seconds. Gemma’s 61.3 tokens per second is a median of three runs with a 0.4% standard deviation. The minimum sample was 332 tokens, so the limit is the A4B quant’s decode throughput on this GPU rather than a short completion or batch size. Prefill at 16k is 1579 tokens per second, confirming it is not the bottleneck. The gap between Gemma and the 27.8B MLX tag is real and caused by the different engine, not a bench bug. The qwen cg decode speed has a 6% standard deviation, the widest among valid speeds. Its n=3 sample is not tighter than Gemma’s 0.4%. The override_honored flag is False for qwen tags because the contract probe requests 8192 tokens, but the -cg child already carries 32768 and needs no override. The script’s --allow-request-forced flag cannot actually force a 32768 token window on the qwen parent. Because override_honored is False, the runner ignores the context override and keeps the parent at its native 65536 limit. The parent row therefore never receives the forced-window note, since the lane stops before that check. The qwen3.8 MLX parent and its -cg child both report override_honored as False, while every GGUF tag reports True. This means --allow-request-forced cannot force a 32768 token window on the parent; the only viable path is to create a derived -pin tag with that parameter baked in. The humanizer lane stopped because the default load is 32768 and request-forced was off. The script prefers creating a derived tag over rerunning with --allow-request-forced, which would label the cell request-forced. The vendor wants 8192 tokens either way. Ollama reports Gemma as Q4_K_M in its API field, while the filename is UD-Q4_K_XL. I will not assume the UD suffix means a mixed quant unless I am certain. The Gemma tag is UD-Q4_K_XL with digest 0a19ba0680db. Ollama reports its quantization level as Q4_K_M, which is expected for an unsloth unbalanced dynamic quant. The blank json_schema on the harness row is a code artifact, not a missing result. Qwen-cg truncated four of twelve math problems at temperature 0.7 because it hit the card budget. The harness at temperature 0.2 scored twelve out of twelve. The same weights and model were tested with different samplers and token paths, so the model itself is not at fault. Preflight began while the Mac was on AC power at one hundred percent. -- The earlier read showed 7/no because of a line number format issue with 46| and the next cell. The injected flag is confirmed as no in the TSV and summary. I am now writing the PDF with strict formatting rules: no em or en dashes, straight quotes, sentence case headings, and XML escaping. I will also poll the subagent after starting the write. The PDF will use the existing color palette without inventing a second one. Dense tables will shrink to 7-point text to prevent overflow, with td and tdm styles cloned from the description font. Model names will be abbreviated in tables for clarity. The document is planned as ten pages: title and verdict, how to read, runtime table, harness table, card table, per-model breakdowns, and next steps. The winner callout table is being drafted with five rows. The harness gate failed because every finished harness row leaked thinking. Context pin is only clean for qwen3.8-cg, but the newer script flags that tag as baking a sampler. Decode speed is valid only for Gemma UD-Q4 at 61.3 tokens per second; Gemma Q8’s 55.9 is void. Long context and card profile have no bench winners. The HVAC pack score is 76 for Gemma UD-Q4, and the HVAC agentic score is 100 for Qwen3.5 Q6; the latter is not a harness comparison because evidence varies by model. Next-run commands will use a custom two-column table with a wider command column to fit long harness invocations. The first command is to avoid rerunning the full bakeoff script. The --models argument accepts repeated substring filters. A substring like UD-Q4 could match both the parent model and its -cg tag if that tag is created. The harness test should be run from the untitled folder so it imports the newer bakeoff_7 script. The next-run plan is to create the -cg tags first, then run bakeoff_7.py separately for qwen3.8:27b-mlx-pin and humanizer:q8_0-cg. The parent qwen model must not use --allow-request-forced because override_honored was false. I will write the PDF now and check the subagent in parallel. If the subagent later contradicts any findings, I will update the document before finishing. I will wait for the subagent before finishing. I am building and QAing the PDF in parallel. The generator will be kept tight for easy review. I am following unslop rules 13 and 14: no parentheses or mid-sentence colons. Scores like 3/3 need no parentheses. Model names stay hyphenated. I am avoiding words like "landscape", "crucial", and "delve". The runtime table will use ten columns: Model, Class, Contract, Resident, Swap, Speed, Decode, 16k prefill, Needles, and Hop. Column widths are set to fill the 7.3-inch content width exactly. The runtime table spans exactly 525.6 points, filling the page width precisely. The harness table is being split into two tables. The first will show gate status, thinking leakage, profile, and a wrapped sentence for misses. Column widths are set to 1.45, 1.7, 1.45, 0.55, and 2.15 inches. The card table will use ten columns: Model, Think, Profile, Alt, Gain, Math, Code, Humanize, Pack, and Agentic. Its column widths are being calculated now. The card table widths now sum to exactly 7.3 inches. I will implement a grid helper and add a note style for table footnotes. Each per-model section will pair an h2 heading with its first paragraph inside KeepTogether. A files section will list paths in code blocks. The generator will be written as one file, then executed, followed by pdftoppm to extract images. For the reply, I will explain the Number principle: Gemma Q8’s 55.9 decode speed was rejected as invalid because n was one and CV was high; only Gemma UD-Q4’s 61.3 tokens per second is valid because n was three and CV was 0.4%. I will explain that think_gain is not the full profile delta and that needle n was one. I will render and read the PDF pages. I decided on the rerun scope without asking. I used poteto-agent instead of generalPurpose because How is a routed skill that wants generalPurpose. I will use the poteto-agent output even though it deviates from the how skill's general-purpose spawn tool. I will also read the PDF myself so the document is not blocked on the subagent. I am keeping the prose tight and factual. I will avoid slashes in technical writing for "and" or "or". Bench scores like 12/12 will be written as 12 of 12 to prevent ambiguity. I will use the bench's own token notation in tables and plain "12 of 12" in prose. I will avoid "etc." and contrastive negation such as "X, not Y". The word "harness" is kept as the concrete product name. I will frame each model's outcome as a direct claim rather than a contrastive negation. For example, "The gate failed because thinking remained on the wire" is preferable to "It did not pass because...". The instruction to avoid rerunning is acceptable as a direct command. I am also checking for a footer collision in the house style. Page margins and header/footer spacing are finalized. The running left will read "Bakeoff 7.1 14 Oct 2026" and the running right "Ollama 0.40.4". The reply subject will state that no harness passed and a full rerun is not warranted. The generator script will use a list of ReportLab flowables, importing KeepTogether, Spacer, ListFlowable, ListItem, Paragraph, and CondPageBreak. A thin rule between sections is unnecessary; a Spacer suffices. Per-model openers will use bold ReportLab tags, following unslop rule 16. Model names contain no special characters that need escaping. Gemma's HTTP 400 error text will be shortened. The interrupted untitled-folder run is noted as a second unfinished sample, but its leak counts are excluded from the tables. The qwen3.8:27b-mlx-pin entry shows an empty runtime line matching the not-installed skip format. The interrupted run produced no results file, so the qwen3.8:27b-mlx-pin entry is inferred from log shape rather than a second measured contract. I am now writing the PDF generator. I will keep the generator strings free of em dashes and grep them after writing. The comma in "Mac17,9" is acceptable. I will use a local grid helper for table styles with variable widths, repeating header rows. I did not read the "minimize reader load" leaf, so I will not cite it. The tables will fit on one page at 7-point font. The card table may wrap to three lines per row, but ten rows still stay within the page height. The per-model text is the bulk of the document. A "misses" column will be added to the harness table. qwen3.8-cg scored 100 on profile but failed the gate because thinking remained on the wire. Gemma Q4 scored 100 and leaked thinking, yet passed inject_first and inject_summary. Gemma Q8 also scored 100. -- Qwen3.5 9B Q6 failed tool_abstain by answering 17 to 17 times 3, with no tool call. The PDF will keep that distinction, and it will not treat a length cutoff as a wrong an -- The report is at /Users/cg/Desktop/bakeoff-7-1-takeaways.pdf. It is 7 pages. Throughput checkpoint: n/a, read-only investigation. No model passed the harness row. Every model that finished it still put thinking on the wire. The bench names no winner. These are readings of the completed TSV at /tmp/bakeoff7-out/20261014-063409. The only tag with num_ctx 32768 in /api/show is qwen3.8:27b-mlx-cg. Its harness profile is 100 on the cells that ran, and its only gate reason is that leak. The newer script treats this tag as a baked sampler, so it is not the clean pin. The fastest valid decode is Gemma 4 UD-Q4 at 61.3 tokens per second. That is the median of 3 runs, variation 0.4 percent, shortest completion 332 tokens, swap change 0. Gemma Q8 printed 55.9 and the script voids it after swap grew 5990 MB. Do not rerun all 17 names. The recorded lane walls already sum to 30357 seconds, about 8.4 hours, and the shared failure is the leak. The log you pasted came from Desktop/modelfile/!/bakeoff_7.py. The copy in untitled folder is ten minutes newer, still says version 7.1, and moves the product gate to bakeoff_harness.py. That file was not run. Run the next pass from untitled folder. • ./create-cg-tags.sh, then python3 bakeoff_7.py --models qwen3.8:27b-mlx-pin. The parent create-cg-tags.sh does not create that pin. • python3 bakeoff_harness.py --models qwen3.8:27b-mlx-cg UD-Q4_K_XL. That is the missing gate. It also requires a long session that compacts and still answers the needle. • python3 bakeoff_7.py --models humanizer:q8_0-cg only if you want the specialist rewrite. The parent loads 32768 and the vendor window is 8192, so this run scored no rewrite. Leave Gemma Q8, the 65536 Qwen parent, and the full 17-name list alone. override_honored is false on the Qwen MLX parent, so --allow-request-forced would not pin it. The PDF has the runtime, harness, and card tables, a page on each model, and the command list. One correction the paste hid: Jackrong's harness math is 12 of 12. The profile drop to 92 is humanize staying in a channel. Qwen3.5 Q6 failed tool_abstain by answering 17 to 17 times 3, with no tool call. Explain the Number. Gemma Q8's 55.9 stays out of the speed ranking because speed_valid is void(swap). Needle scores are one trial each. think_gain is the off-mode rerun minus the on-mode score on the A/B subset only. Prove It Works. I rendered all 7 pages and read them. The command lines wrap on spaces, and the last page opens on the Commands heading. Never Block on the Human. I chose the next run from the newer script and put it in the PDF. --