# Data package: the roster, as of 26 September 2026 This package holds the redacted primary files behind the page "The roster, as of 26 September 2026": the runs of 12 to 15 September that the first list (15 September) was built from, and the runs of 16 to 26 September that this snapshot adds. It also holds a few clearly identified extracts, three tables rebuilt from the page and the files, a listing of the model files taken on 26 September, and the vendor documents the page cites at pinned revisions. The machine is one RTX 5090 desktop with 188 GiB of RAM; one row also uses a laptop joined to it by a direct Thunderbolt cable. **If a number on the page disagrees with a file in this package, the file is right and the page is wrong.** Write to hello@graphometer.ai and the correction goes on the page with its date. `NUMBERS.md` maps every figure on the page to its file and field. `ROSTER_23.csv` is the page's table, machine readable, with the files behind each row. ## Where this package will look like it argues with itself Read this section first. Every one of these is real, and none of them is a mistake in the files. **From the runs of 16 to 26 September** 1. **A speaking figure in the sweep files is not the one on the page.** The batch-size sweep times a short answer: Qwen3-235B's 8.18 in `batch-sweep-2026-09-21/Qwen3-235B-A22B-Instruct-2507_2048.result` is an 11-token code. The page prints its letters of about 350 words from `letters-2026-09-26/` instead, and says so in section 05. 2. **MiniMax M2.7 has two speaking figures for the same request on 13 September.** In `minimax-m2.7/2026-09-13_ctx65536_f16.server.log`, 900 tokens at 9.80 were all hidden reasoning, cut off at the length limit with no answer; in `2026-09-13_ctx131072_q8_0.server.log` the same request, allowed more tokens, finished a prose reply in 1,522 tokens at 9.73. The page prints the second. `2026-09-13_extract.txt` holds the request and both outcomes. 3. **Two reading figures were measured at a smaller window than the model serves.** Inkling-Small's 337.9 and Qwen3.8-Flash-Next's 511.7 (`batch-sweep-2026-09-20/*_2048.result`) were read at 131,072; both models serve 262,144. At 262,144 with the same setting only reads of about 3,000 tokens exist (`*_2048_ctx262144_3k.result` on 20 September and `batch-sweep-2026-09-21/*_ctx262144_3k.result` on 21 September), each the first request after the server started; the long reads at 262,144 in this package are from before the setting changed. The page prints each figure with its window. (26 September update: for Qwen3.8-Flash-Next, reads of 3,019, 48,075 and 229,981 tokens at 262,144 with that setting now exist, in `qwen3.8-flash-next-2026-09-26/`; see point 23.) 4. **MiniMax M3's "194 at depth" is in no file as a rate.** It is 58,307 tokens divided by the whole request's 300.2 seconds (`minimax-m3/2026-09-20_q2kl_needle_58k.json`); the server's own timing for the same read is 208.11 (`2026-09-20_q2kl_ctx131072_ub2048.server.log`). 5. **The Qwen3.8-Flash-Next records contain a failed question.** Each `qwen3.8-flash-next-2026-09-20/*.result` ends with a third question that needs facts combined from across the ledger, answered wrongly at every length, including the 5,986-token control. The page prints only the reading and speaking rates and the three-code retrieval, which passed. 6. **Laguna S 2.1's baseline says one setting in its header and ran another.** `batch-sweep-2026-09-21/ Laguna-S-2.1_baseline_ub128.result` begins `ubatch=skip (llama.cpp default 512)`: that is what the harness printed for any run without an override. Laguna's own start script at the time set `--batch-size 512 --ubatch-size 128` itself (`file-listing-2026-09-26.txt`, section C, the copy saved before the change), so the 72.2 is a `-ub 128` figure, as the file name says. 7. **"skip" in a file name means "no override".** In `batch-sweep-2026-09-21/`, `_skip` is a run through the model's start script without a batch override, so the script's own setting applied, which for these models was llama.cpp's default `-b 2048 -ub 512` on 21 September (the saved copies of their scripts set no batch flag; `file-listing-2026-09-26.txt`, section C). The numbered files (`_1024`, `_2048`, `_4096`, `_8192`) are the micro-batch the run exported. Files named `_series` read several prompt lengths in one load. `Ling-3.0-flash_skip`, added in the fix pass for the window its server reported, is the exception: no saved copy of Ling-3.0-flash's script is in section C, so this package does not show the batch setting that run used, and the page quotes no speed from it. 8. **Some runs went through our own model manager, and say so in their names.** `Laguna-S-2.1_ctx262144_model- manager.result` and `MiniMax-M3_model-manager_3k.result` were started by the service that launches models on this machine, as a user would start them, rather than by the sweep. Their figures are close to, not identical with, the sweep's own (Laguna at 24,013 tokens: 898.8 through the manager, 901.9 in the series run). 9. **Qwen3.6-27B's speaking figures fall with depth, and its row has no 60,000-token letter.** Its letters were written at 20,067, 99,934 and 229,564 tokens; the Qwen3-235B letters at 20,063, 60,180 and 99,864. Different depths, on purpose: Qwen3.6-27B serves 262,144. **From the runs of 12 to 15 September** (unchanged from the first list; these files are as they were packaged) 10. **The same model has two different speeds, and both are correct.** Gemma-4-26B-A4B reads at 10,593 tokens a second in `context-sweep-2026-09-12.tsv` and at 6,727 in `tool-recall-wave1-2026-09-12.tsv`. The first is the server answering a prompt directly; the second is the same server with an agent framework holding the conversation. The page keeps these in separate columns and never averages them. 11. **The summary files and the per-model records disagree about prompt sizes, on purpose.** The ledgers carry the seeded token counts, about 48,000 and 96,000; the records under `gate-records/` carry what the server read, 52,820 to 54,138 and 101,684 to 115,508, because the framework adds its own prompt. The page prints the second. 12. **Round one of the tool and recall check disagrees with round two, and the page uses round two.** `tool-recall-wave1-round1-2026-09-12.tsv` is kept because shipping only the round that agrees with the page would be the kind of selection this site exists to avoid. 13. **One number in round one is impossible and we left it in**: Gemma-4-26B-A4B's 987.9 tokens a second of generation at the 96,000-token leg. Round two records 97.3. 14. **Two framework records carry a FAIL verdict for runs the page quotes.** `gate-records/deepseek-v4-flash-0731-two-machines.json` and `gate-records/inkling-small.json` fail on a separate leg of our own runtime that tests neither speed nor recall nor the tool call; their recall and tool results at both depths passed. `gate-records/glm-4.7-full-358b.json` fails on the tool call at depth, and the page prints it as a failed check. 15. **Graphics memory readings for the same model differ slightly between files.** Raw readings at different moments of different runs, never divided by 1000, never rounded. 16. **One file changed its name on 26 September and nothing else.** `install-runs-qwen3.8-flash-next-2026-09-12.tsv` was packaged on 15 September under a name that carried our internal key for the model; it is renamed, its contents untouched, and the one reference to it in `ROSTER_22.csv` changed to match. **Added in the fix pass of 26 September** 17. **The pair's own chain script calls 262,144 its served window; the page does not.** `batch-sweep-2026-09-21/DeepSeek-V4-Flash-pair_chain.sh` says "its served 262144 window" in its opening comment, and passes 262144 to the start script itself (`SCRIPT_ARGS`). Every run in this package that loaded 262,144 on the pair passed it, and the split script's fallback is 131,072, so the page gives the fallback, says the installed setting may raise it, and gives the two dates 262,144 was loaded (`installed-windows-2026-09-21.txt`, section C). 18. **Inkling-Small served two windows a week apart.** Its 14 September probes and framework record were at 131,072 (`inkling-probes-2026-09-14.md`, run 1; `gate-records/inkling-small.json`, `S.n_ctx`); the crash check of 21 September, which passed no window, was served 262,144 (`installed-windows-2026-09-21.txt`, section A). Both are right for their dates, and the page gives both. 19. **The letter runs record a code that never came back, and the reason is the cap.** Each depth in `letters-2026-09-26/*.jsonl` has an `edit_reread` line with `code_in_answer` false. That request sends the same ledger and letter request again under a changed instruction and allows 40 tokens of reply (`max_tokens` 40 in `prose_probe.py`; `predicted_n` 40 in each line), while the code belongs on the letter's final line, so the reply stops at the letter's opening (`answer_tail`). It times the second read. The `read` lines, which ask for the code directly, have it at every depth, and the page says both. 20. **One rung of GLM-4.7-Flash has an empty answer.** `batch-sweep-2026-09-21/GLM-4.7-Flash_2048.result` has `needle_found` true and `needle_in_answer` false, with `answer_len` 0 and 3,189 characters of reasoning: the code is in the reasoning and the answer is empty. The sweep allows 900 tokens of reply (`max_tokens` 900 in `batch_sweep.sh`). The page names this rung as the one read that did not return the code in its answer. 21. **Two DeepSeek files say `ubatch=skip` inside and a number in their name.** `batch-sweep-2026-09-21/DeepSeek-V4-Flash-3bit-desktop_4096_150k.result` and `DeepSeek-V4-Flash-pair_512_150k.result` ran through the start script with no micro-batch override. The name gives the script's own setting at the time: 4,096 on the desktop's fast preset (its log's first line says `preset=fast`; section C of `file-listing-2026-09-26.txt`; the card reading at load, 28,570 MiB, matches the explicit 4,096 run's 28,490, not the 2,048 run's 27,092) and 512 on the pair (the split script's micro-batch fallback, with `batch=2048` passed). 22. **"skip" in the crash check means what it means in the sweep.** `STRESS.txt` reports `"ubatch": "skip"` for every model: no batch override, so each script's own setting at the time applied. Those runs and the sweep's `_skip` runs are also the runs that passed no window, which is why `installed-windows-2026-09-21.txt` rests on them. Laguna S 2.1's crash check ran at 14:25, before its window changed later that day, and was served 32,768. **Added in the update of 26 September (Qwen3.8-Flash-Next at its served window)** 23. **Qwen3.8-Flash-Next's first read on 26 September is its slowest, and the page leads with the second.** `qwen3.8-flash-next-2026-09-26/shipped_b4096_ub2048.jsonl` row 1 read 3,035 tokens at 87.7; row 2 read 3,019 at 517.6. During the first request the file's share in memory grew from 43.54 to 57.66 GB (`resident_gb_before`, `resident_gb_after`; `run_fn.out` shows 0.0 GB before that start): the only start recorded with the file out of memory. The page leads with the later reads, made with 57.7 to 60.1 GB of the 89.99 GB file in memory, and gives the first one in its row note. The run at the defaults came second and began with 60.05 GB in memory, so its first two reads are alike (253.7 and 255.7): the two settings' first reads are not a like-for-like pair. The file never became fully resident (at most 60.06 GB). The 20 September 45.4 was also the first request after a start (`load=65s` in its result), but that file records no memory state, so the page gives it no cause. Every read was a new prompt (`fn_reads.py`), so no read reused another's prompt cache. The server was started through a copy of the start script that differs from the installed one only in the line naming its settings file (and in a comment the installed script gained afterwards); `run_fn.sh` points that line at a settings file which, by the run's own notes, held the installed values without the helper process. We did not open it. **Added in the same update (Qwen3.8-27B at 262,144)** 24. **Qwen3.8-27B's run says `Q38_MTP=1` and the page says the draft head was off.** `qwen3.8-27b-2026-09-26/Qwen3.8-27B_ctx262144_draft-off.result` asked for the draft head, and at 262,144 on this file the start script turns it off (`file-listing-2026-09-26.txt` section C, lines 150 to 155): its `cmdline:` line has no draft flags and every JSON line has `draft_n` null. The run went through a copy of the start script, named in its first line, pointed at a settings file named in its `env:` lines, which we did not open; the command line it produced is the one the installed script produces for this file at this window. 25. **One structured reply in that file says `code_in_answer` false and quotes a code.** At 229,152 tokens the ledger holds three codes; the structured request asks for "the sealed reference", and the reply ends with the first of the three, `CEDAR-4418-HARBOR`, while the check looks only for the middle one. The page prints no structured figure; its `read` line at that depth has all three (`codes3_hits` 3). 26. **Qwen3.8-27B's window differs by date.** The 21 September sweep served it 131,072 (`installed-windows-2026-09-21.txt` section B); it was installed at 262,144 on 26 September, and no run in this package started the installed script without a window after that (section C). The page gives the 262,144 in the install-record form and keeps the 131,072 figures, a window at which the default file keeps its draft head. The start script can also serve two smaller files, UD-Q4_K_XL and UD-Q5_K_S, both on disk (`file-listing-2026-09-26.txt` section A), and at 262,144 it keeps the draft head for both (section C, lines 150 to 158; the two `*_draft-on.result` files in `qwen3.8-27b-2026-09-26/`). Our operator records call these configurations presets; the start script does not name them, so the page does not either. **Added on the evening of 26 September (Qwen3.5-122B-A10B, Ornith-1.5-35B-A3B and GLM-5.3-Flash at 262,144; MiniMax M2.7 at 196,608; the starts with no window passed)** 27. **At 229,090 tokens their letters and structured replies say `code_in_answer` false and quote a code.** As in point 25, the ledger at that depth holds three codes, the request asks for "the sealed reference", and the reply names the first or the third while the check looks only for the middle one. The `read` line at that depth has all three (`codes3_hits` 3) in every 262,144 file, and the page prints no structured figure. 28. **Each model has three runs, and the page's lead is only one of them.** The lead is the setting installed at 262,144 (`Qwen3.5-122B-A10B_ctx262144_b2048_ub512`, `Ornith-1.5-35B_ctx262144_b2048_ub2048`). The 122B's `-ub 1024` run reads faster and was not adopted (its card peak left about 0.8 GB, the start script's comment, section C line 135); Ornith's `-ub 512` run is the setting its script used at that window before the change. The 131,072 runs are the same day's comparison, at the other window and batch setting, and the page says so. 29. **Their windows differ by date.** The crash check of 21 September served Qwen3.5-122B-A10B and Ornith-1.5-35B-A3B 131,072 (`installed-windows-2026-09-21.txt` section A); both were installed at 262,144 on 26 September, and a start of each installed script with no window passed was served 262,144 that evening (section E, and `installed-starts-2026-09-26.txt`), as were MiniMax M2.7 (196,608) and GLM-5.3-Flash (262,144). The 122B also loaded 262,144 on 15 September, by passing it (`long-reads-2026-09-15/`). 30. **MiniMax M2.7's 26 September letters say `code_in_answer` false, and two of them hold no letter at all.** It always thinks. At 20,039 and 47,933 tokens the letter request's whole 4,096-token budget went to hidden reasoning (`answer_words` 0, about 14,700 characters of reasoning), so those speaking figures time reasoning only, and the page says so. At 189,491 it wrote a 275-word letter that ends with the first of the three codes, the artifact of point 25. Its `read` lines quote the code at every depth, 3 of 3 at 189,491. The window came from the run's `env:` line through the installed start script, which sets the one shape that loads at 196,608 itself. 31. **GLM-5.3-Flash's installed setting was measured only to about 48,000 tokens.** `glm-5.3-flash-2026-09-26/` holds its `-ub 1024` run at about 20,000 and 48,000 tokens; the deeper reads and letters, and the 3 of 3 at 230,039, are the `-ub 2048` run, which loaded but was not adopted (31,951 MiB at peak). The page labels each figure with its setting. The runs set the reasoning effort to none, and the model still reasoned: several letters stop at their 1,200-token budget with little or no visible text (`predicted_n` 1200, `answer_words` 0 to 119), and `code_in_answer` is false there because the letter never reached its last line. The `-ub 4096` start that did not load has a result and a server log of its own. 32. **Its crash check counts 36 "empty answers" out of 40 and no crash.** `empty` is true when either of a prompt's two replies has no visible text; every such prompt used its whole budget (160 tokens, then 80). The four with text in both replies are prompts 0, 4, 8 and 11; prompt 0 is a 3,014-token ledger whose reply quoted its code. 33. **The four starts of the evening are records of another tool, cut down.** `installed-starts-2026-09-26.txt` keeps the fields the page cites. Their card readings at load (30,848, 31,256, 31,512 and 28,714 MiB) sit close to the measured runs' at the same settings (30,898, 31,270, 31,709 and 28,834), with the settings each script applies by itself at that window. ## What was removed, and why Everything in this package was recorded on a personal machine that also does private work. The redactions remove where and how, never what was measured. **Files added on 26 September** (218 copied records, plus `ROSTER_23.csv`, `summary-table-corrections-2026-09-26.csv`, `file-listing-2026-09-26.txt` and `installed-windows-2026-09-21.txt`, which were written for this revision; the fix pass of 26 September added ten copied records and the last of the written files, and the update of the same day the six files of `qwen3.8-flash-next-2026-09-26/` and the five of `qwen3.8-27b-2026-09-26/`, and that evening the six each of `qwen3.5-122b-2026-09-26/` and `ornith-1.5-35b-2026-09-26/`, three in `minimax-m2.7/` and eleven in `glm-5.3-flash-2026-09-26/`), replacements by kind, as counted by the copying scripts: - 164 absolute or working-tree paths to `` (the file name is kept when it is a model shard, a library or a source file); - 152 bridge and loopback addresses to ``; 4 link addresses to ``; 26 uses of the laptop's short name or the desktop's host name to `` or ``; - 30 port numbers to ``, two of them in the usage lines of `batch_sweep.sh`; - 371 of our internal keys for a model to its public name, in run tags, labels and script text (file names were renamed the same way; 67 of them are the Qwen3.8-27B runs' tags in their four result files and 138 the evening runs' tags in their twelve; in the MiniMax M2.7 run of 26 September the rule also renamed the settings file named in its `env:` line), 22 serving aliases to ``, and 5 build folder names to ``; in `qwen3.8-flash-next-2026-09-26/run_fn.sh` the key rule also matched `grep`'s `-m3` flag, which was put back and is not counted; - 56 request identifiers removed from response JSON; - **lines removed:** 104. Of these, 94 were a helper process's start-up and status lines in the server logs, each kept as its timestamp followed by ``, so the logs still show when each run began; four are an `env:` line in each letters result that named a private folder; six are comment or command lines in the scripts that described a private use (two of them the lines in `stress_model.py` and `glm-5.3-flash-2026-09-26/stress_fw.py` that stopped that helper). The three sweep drivers and the two crash-check scripts also lose a block (24 lines in all) that pointed that helper at an empty folder. `installed-windows-2026-09-21.txt` quotes its lines through the same tool, which replaced 2 port numbers and 1 address in them. `installed-starts-2026-09-26.txt` is written from four start records and keeps only the fields it names in its header: the name of the tool that issued the starts, the service unit names, the tool's own label for each model and its two status lines, the time each server took to become healthy, and the unit state at stop are not reproduced (a label held an em dash). **Words changed, not measurements.** The docstring of `letters-2026-09-26/prose_probe.py` was rewritten to its technical content, two comment lines about a private use became one, and a dash in a help string became a colon; `minimax-m2.7/m27_probe.py` lost three comment passages of the same kind; one sentence of `qwen3.8-flash-next-2026-09-20/hard_recall_probe.py` and one phrase of `run_hard.sh` were reworded; `batch_sweep.sh` lost a clause of its opening comment; one printed label in each of the two `minimax-m3/*.measure.log` files has a phrase replaced with ``; the two `minimax-m3/*launch-check.log` files lose the line telling the operator how to stop the server, which named a process-id file. The docstring of `qwen3.8-27b-2026-09-26/probe38.py` was rewritten to its technical content in the same way as `prose_probe.py`'s, one comment line about a private use was shortened, and a dash in a help string became a colon; `minimax-m2.7/2026-09-26_probe_fw.py`, the same probe with a `--read-max-tokens` option, had the same changes. **Extracts, each saying so in its first lines.** `minimax-m2.7/2026-09-13_extract.txt` (the launch block of a serving script, two probe lines, the result lines of two runs with their texts omitted, and the layer count from a file header), `first-runs-2026-09-16/first-runs-2026-09-16.tsv` (the header and the two rows for the models on this list; the same table held a third model that is not an endpoint here) and `installed-windows-2026-09-21.txt` (two lines from each of twelve crash-check server logs, which are not otherwise in this package, and two lines from each of seven sweep results, which are). **One record reformatted.** `gate-records/minimax-m2.7-2026-09-21.json` is shipped in the format of the other seventeen records in that folder: the private runtime's own blocks (`P`, `fit_provision`, `H1`, `X`), its `key`, `unit`, `port` and `handle` fields, the alias inside `S`, and two fields in each long leg that describe our runtime rather than the model are dropped; the two tools the model was asked to call appear as `file_write` and `file_read`. Every measurement field is kept. **Files from the 12 to 15 September runs** are as they were packaged on 15 September, with the redactions that packaging described: paths, machine identity and network, ports, service and unit names, our internal keys, one column naming private work (`lane`) and four columns describing our runtime dropped from the two ledgers, the runtime's own blocks dropped from the per-model records, private tool names renamed, the agent framework unnamed, `refusal-guards-2026-09-12.txt` redacted with square-bracket placeholders, two extracts, and `ROSTER_22.csv` and `summary-table-corrections.csv` rebuilt or extracted. **Nothing that carries a measurement is altered.** Every token count, time, rate, byte count, memory reading and verdict is as recorded. Units are the units the record used: MiB stays MiB. The typographic characters the recording programs printed, dashes included, are kept in the copied records and quoted script lines (seven quoted lines in `file-listing-2026-09-26.txt` carry an em dash from a script comment); the text written for this revision carries none. **Verification.** The package brief's search over the whole `data/` folder, for working-tree and home paths, the staging folder name, host and tailnet names, link and bridge addresses, the private system's vocabulary and the framework's, and service names, returns nothing. A wider search for our internal model keys and the machine's service port numbers finds one number, 8129, which is a token count in `gate-records/minimax-m2.7-2026-09-21.json` and not a port. ## The files ### Added for the 26 September snapshot | file or folder | what it is | date of the runs | |---|---|---| | `ROSTER_23.csv` | the page's 23-row table, machine readable, built from the page itself, with the files behind each row | written 2026-09-26 | | `summary-table-corrections-2026-09-26.csv` | section 05's table: what our records said, where that figure stood, what the files show, and the files | written 2026-09-26 | | `file-listing-2026-09-26.txt` | every file each row serves, with byte counts from the file system; the files no longer present; and the start-script lines behind the row notes (windows, batch sizes, cache, file choice), paths and ports removed | read 2026-09-26 | | `batch-sweep-2026-09-21/` | the batch-size sweep through each model's own start script: results and server logs at the defaults and at the setting each model now runs, the other rungs of the models kept at the defaults (results only), the series runs (several lengths per load) for GLM-5.3-Flash, Ling-3.0-flash, Laguna S 2.1 and both DeepSeek services (each DeepSeek service also at `-ub 2048`), the 3,000-token reads at 262,144 for Inkling-Small and Qwen3.8-Flash-Next, two runs through the model manager, the three drivers and the pair's chain script; and the crash check of the same day (its chain `stress_chain.sh`, its per-model script `stress_model.py` and its summary `STRESS.txt`), whose runs passed no window | 2026-09-21 | | `glm-5.3-flash-2026-09-26/` | GLM-5.3-Flash at 262,144, installed as its default that evening at `-b 4096 -ub 1024` (reads and letters at about 20,000 and 48,000 tokens), at `-ub 2048` (measured to about 230,000, not adopted), the `-ub 4096` start that did not load (result and server log), and at 131,072 with `-b/-ub 4096` the same day; the 40-prompt crash check at the installed setting (summary and one line per prompt) and its script | 2026-09-26 | | `installed-starts-2026-09-26.txt` | four starts of installed start scripts with no window passed (Qwen3.5-122B-A10B, Ornith-1.5-35B-A3B, MiniMax M2.7, GLM-5.3-Flash), 21:06 to 21:13, redacted to date and time, the served window from the server and its log, the card before and at load, and the stop time | written 2026-09-26 from records of that evening | | `qwen3.5-122b-2026-09-26/` | Qwen3.5-122B-A10B at 262,144, installed as its default that evening, at the setting installed there (`-b 2048 -ub 512`) and at `-ub 1024` (measured, not adopted), and at 131,072 with `-b/-ub 4096` the same day: each run's result file (window, command line, card memory) and one JSON line per measurement, reads and letters at about 20,000, 48,000, 100,000 and 230,000 tokens (100,000 at 131,072). The probe is `qwen3.8-27b-2026-09-26/probe38.py`, the same file under another name | 2026-09-26 | | `ornith-1.5-35b-2026-09-26/` | Ornith-1.5-35B-A3B at 262,144, installed as its default that evening, at the setting installed there (`-b 2048 -ub 2048`) and at the old one (`-b 2048 -ub 512`), and at 131,072 with `-b/-ub 4096` the same day: result files and JSON lines as above | 2026-09-26 | | `qwen3.8-27b-2026-09-26/` | Qwen3.8-27B at 262,144 with the draft head off, the configuration installed as its default that afternoon: the result file (window, command line, card memory) and one JSON line per measurement, reads and letters at about 20,000, 48,000, 100,000 and 230,000 tokens, with the probe; and the result files of the two smaller files at 262,144 with the draft head on, cited for their command lines | 2026-09-26 | | `qwen3.8-flash-next-2026-09-26/` | Qwen3.8-Flash-Next at its served 262,144 window, the server started once at `-b 4096 -ub 2048` (from a cold start) and once at llama.cpp's defaults: one JSON line per read for each setting (`shipped_b4096_ub2048.jsonl`, `default_b2048_ub512.jsonl`), the run's own log of each load (`run_fn.out`: window, card memory, how much of the model was in memory), the driver, the read probe and the memory-residency script | 2026-09-26 | | `installed-windows-2026-09-21.txt` | for each endpoint that has one, a run on 21 September that passed no window, and the window its server reported; notes for the endpoints these records cannot settle | written 2026-09-26 from runs of 2026-09-21 | | `batch-sweep-2026-09-20/` | the runs started by hand: Inkling-Small and Qwen3.8-Flash-Next at the defaults and at `-b 4096 -ub 2048` (131,072), the same setting at 262,144 on 3,000 tokens, DeepSeek's 8-bit file at the defaults and at 8192 (48,073 tokens at 131,072; 150,324 at 262,144) with the two result summaries its verify runs printed, and the drivers | 2026-09-20 | | `qwen3.8-flash-next-2026-09-20/` | Qwen3.8-Flash-Next at 262,144 with the 8-bit cache: a 229,982-token read and a 5,986-token control, and the two larger windows (results only), with the driver and the probe | 2026-09-20 | | `minimax-m2.7/` | the 13 September desktop runs (two server logs and an extract), the 19 September micro-batch results with one server log and the probe; and the 26 September run at 196,608, installed as its default that evening (result file with window, command line and card memory, one JSON line per measurement, and its probe, `2026-09-26_probe_fw.py`) | 2026-09-13, 19 and 26 | | `minimax-m3/` | the Q2_K_L file alone at 131,072 (launch check, server log, measurement log, prefill and decode JSON, the 58,307-token recall read), the attempts at 262,144, and the 196,608 rung | 2026-09-20 | | `letters-2026-09-26/` | Qwen3-235B at its served setting and at `-ub 4096`, and Qwen3.6-27B at 262,144 with the draft head off, plus the attempt with it on: results, the probe's JSON lines, and the probe | 2026-09-26 | | `first-runs-2026-09-16/` | Ornith-1.5-35B-A3B and GLM-5.3-Flash on their first day: the table rows, the paragraph and 120,000-token read JSON, the server logs, the driver and the long-prompt probe | 2026-09-16 | | `long-reads-2026-09-15/` | the long-read table at a full window, and the paragraph and deep-read JSON for the eight rows the page draws on | 2026-09-15 to 16 | | `gate-records/minimax-m2.7-2026-09-21.json` | MiniMax M2.7's framework check, in the format of the September records | 2026-09-21 | | `NUMBERS.md`, this file | rewritten for the snapshot | 2026-09-26 | ### From the first list, 12 to 15 September (kept as packaged) | file | what it is | date of the run | |---|---|---| | `context-sweep-2026-09-12.tsv` | the direct sweep at each model's own flags, varying only the context size; its header states the method | 2026-09-12 | | `install-runs-qwen3.8-flash-next-2026-09-12.tsv` | the install runs of Qwen3.8-Flash-Next (renamed on 26 September, see point 16) | 2026-09-12 | | `tool-recall-wave1-2026-09-12.tsv`, `tool-recall-wave1-round1-2026-09-12.tsv`, `tool-recall-wave2-2026-09-13.json` | the framework check: rounds two and one, and the second wave | 2026-09-12 and 13 | | `gate-records/` (seventeen files dated 12 to 14 September) | the per-model framework records behind those summaries | 2026-09-12 to 14 | | `refusal-guards-2026-09-12.txt` | seven recorded refusals to launch while another model was up | 2026-09-12 | | `two-box-deepseek-probes-2026-09-13-to-15.txt` | probe lines from the two-machine runs, extracted | 2026-09-13 and 15 | | `inkling-probes-2026-09-14.md` | Inkling-Small's first-night probe tables and header facts | 2026-09-14 | | `model-file-headers.md` | what the model files themselves say, for the first list's rows | 2026-09-12 to 14 | | `ROSTER_22.csv` | the first list, 15 September, as it was | assembled 2026-09-15 | | `summary-table-corrections.csv` | the first list's four corrections of our summary table | our record as it stood 2026-09-15 | | `vendor/` | the makers' model cards for DeepSeek V4 Flash 0731 (revision 7872f01b, with its licence) and Inkling-Small (revision 8cc5877b), verbatim | vendor documents | ## What is not here - **Model output, beyond what a measurement needs.** What is here: the paragraph replies about a water pump (the 15 and 16 September JSON), the first or last 60 characters of some answers and letters, the code words a model had to quote back, a few short wrong answers, short tool returns and character counts. Nothing any model wrote in real use. - **No prompts from real work.** Every prompt is synthetic: a generated ledger with code words planted in it, a paragraph request, a letter request against the ledger. - **No quality measurement of any kind**, because none was made. - **No hosted-service output.** Every model here ran on this desktop. - **Nothing from the private side of this machine**, and nothing that identifies the machine or its network.