# Data package: September field notes (2026-09-12 to 2026-09-21) This package backs the page at `/field-notes-2026-09/`. It is a leak-swept selection of the run records the page cites, with derived summaries where the run produced no public primary record. If a number on the page disagrees with a file in this package, the file is right and the page is wrong. Revised 2026-09-26; the files added or changed in that revision are marked "(2026-09-26 fix)". ## Where this package will look like it argues with itself 1. **Two files give two reading rates for the Qwen3.8-Flash-Next 65,536 preset.** `context-sweep/MEASUREMENTS.tsv` row `Qwen3.8-Flash-Next card-light-64K` prints `247.00`; its own `source` column marks that row `earlier-0912`, a rounded restatement of a measurement taken earlier the same day. The primary record, `flashnext-install/MEASUREMENTS.tsv`, gives 246.9 t/s on a 2,624-token prompt. The page prints 246.9. 2. **`wave2/WAVE2_SUMMARY.json` records a timeout as a recall miss.** Its `minimax-m2.7` entry has the 96,000-token leg as `hits = 0`, "l128 recall 0/3", verdict FAIL. That leg ended at 1,800.4 seconds with no answer; the harness left `error: null` and counted the recall as 0 of 3. The same entry's `l64.cold` (1,662.91) is the time to first content; the whole leg took 1,682.9 seconds, which is the figure the page compares with the 21 September re-run. `wave2/CORRECTION.md` annotates both, with the fields from the per-leg record. 3. **`minimax-m27/relaunch/WARM.console.log` glues two clocks together.** Its line `cold: 85763 tokens in 147.49s (589.5 t/s)` pairs the client wall time of the whole request (147.49 s, including a 16-token answer) with the server's prompt-eval rate, which `WARM.server.log` computes over 145,476.38 ms. 85,763 / 147.49 is about 581.5. The page prints both clocks separately. 4. **`minimax-m27/relaunch/192K.console.log` prints "FIT" after rungs that did not fit.** The `^ FIT at -ub 4096` and `^ FIT at -ub 2048` lines follow a load that failed and a load that aborted with a core dump (the `VOID` lines above them). The `VOID` lines are right; only `-ub 1024` loaded at that window. 5. **One label, two runs.** `one_ub4096_cmoe_ctx131072` is the 3,658-token read in `minimax-m27/relaunch/CONFIRM.console.log` (564.0 t/s) and the 43,909-token read in `DEPTH4096.console.log` and the `PHASE_B.console.log` tables (640.2 t/s). The harness reused the label; the token counts tell the two apart. 6. **The MiniMax M2.7 records of 2026-09-13 and 2026-09-15 predate the launch fix.** Everything under `minimax-m27/MEASUREMENTS.md` and `minimax-m27/budget/` ran before the fix of 2026-09-19 (the 13 September gate at the inherited `-cmoe -ub 128`, per the WHY comment in `minimax-m27/relaunch/m27_sweep.sh`). Their wall times and reading rates belong to that week; `minimax-m27/relaunch/` is the fixed setting. 7. **The MiniMax M3 sweep files carry an impossible decode figure.** The `minimax-m3/A_*.json` results print `"decode_tps": 1000000.0`: the probe divided by a zero-length generation. Decode was not measured in those runs. 8. **The two Ling-3.0-flash reading baselines were measured at different served windows.** `batch-sweep/ling_base.result` (238.2 t/s, default flags) was served at 131,072 on 2026-09-20; `batch-sweep/ling_ub2048_confirm.result` (727.7 t/s, `-b 4096 -ub 2048`) at 262,144 on 2026-09-21. The one same-window pair is the 3,007-token read at 262,144: 201.5 t/s at `-ub 512`, 444.0 at `-ub 2048`. 9. **A crash result labelled with the wrong default.** `batch-sweep/Ling-3.0-flash_ub4096_CRASHPROMPT.result` says "ubatch=skip (llama.cpp default 512)". The run went through the model's start script, whose default at the time was `-ub 4096` (`diff_Ling-3.0-flash_crashfix.txt` shows it changing to 2048 afterwards). The label is the harness's. 10. **The Qwen3.8-Flash-Next batch files are 131,072-window reads.** `batch-sweep/flashnext_base.result` (203.2 t/s) and `flashnext_ub2048.result` (511.7 t/s, `-b 4096 -ub 2048`) were read at 131,072 on 48,069-token prompts. At the 262,144 window the model serves, the only reads after the change are 45.4 t/s on a 3,041-token prompt (2026-09-20) and 112.3 t/s on a 2,999-token replay (2026-09-21); those records ship with the batch-size study, not here. The page prints neither 203.2 nor 511.7 and points to that study. 11. **The rope-trap logs show the cap, not the allocation, and one probe says TIMEOUT.** `flashnext-context/e_512k_norope.log` carries the lines that cap the slot to 262,144; the "still allocates the big window" half is in the card readings (`e_512k_norope.vram` 23,816 MiB, `b_512k_unlock.vram` 23,814 at a genuine 524,288, `a_256k_base.vram` 15,394 at 262,144). `a_256k_base.vram` records `result=TIMEOUT` at 600 s; `a_256k_base.log` (2026-09-26 fix) shows that server loaded and listened, so its card reading is of a loaded 262,144 server. 12. **`256k-sweep/CONTEXT_256K_MEASUREMENTS.tsv` has rows the page does not cover.** It now ships in its final eleven-row form. The page prints six rows (`gemma26_262144`, `glm47flash_202752`, `flashnext_262144`, `ling_262144`, `inkling_262144`, `qwen397_262144`); the other five belong to models this roundup has no entry for. 13. **`context-sweep/MEASUREMENTS.tsv` has rows for models covered by other pages** (Qwen3.8-27B, GLM-4.7-Full, Qwen3-235B), including one REFUSED rung. The file is the sweep's record of record and ships whole. 14. **The budget probe's "first answer token" column is an estimate, not a timing.** `minimax-m27/budget/budget_probe_log.txt` prints it with a `~`; the probe takes the wall time and subtracts the answer's own decode time, with the answer's token count taken as its character count divided by four. ## File map and provenance ### `context-sweep/`: the 2026-09-12 context sweep - `MEASUREMENTS.tsv`: the sweep's record of record. Its header block is the method statement (card and MiB rule, speaking as the best of two warm ~200-token prose replies, the ~21,000-token ledger, the 97 GB download caveat). Redaction: the `body` column held internal shorthand keys, renamed to public model names, and one internal gate-check nickname was genericised. - `logs/`: the per-rung server logs and the per-probe response bodies for the four models this page covers (GLM-4.7-Flash, Gemma-4-26B, Gemma-4-31B at two rungs, Qwen3.8-Flash-Next). `*.probeA*.json` are the warm speaking generations, `*.probeB.json` the long-prompt read with the planted code, `*.warm.json` the warm-up request. The `*.req` request bodies (about 86 KB of synthetic ledger text each) are omitted. ### `flashnext-install/` - `MEASUREMENTS.tsv`: the Qwen3.8-Flash-Next install-day measurements, 2026-09-12, including the 65,536-token preset figures the page quotes. Its header carries the flags, the machine line and the upstream llama.cpp commit. ### `256k-sweep/`: the 2026-09-15 long-window sweep - `CONTEXT_256K_MEASUREMENTS.tsv` (2026-09-26 fix): the record of record, now the final eleven-row version. The five rows shipped before are unchanged; the six rows added (`qwen122_262144`, `inkling_262144`, `dsv4split_262144`, `qwen397_262144`, `dsv4q8_262144`, `dsv4split_131072_120k`) were each checked, field by field, against their server log's `[result]`, `[short rep]` and `[deep]` lines before shipping. Its header block is the method statement. - `gguf_context_header_reader.py`: the standalone, dependency-free GGUF header reader behind every arithmetic figure on the page. It reads the header block and never touches tensor data. - `header-reads/`: that reader's output, one file per model, for every file whose cache cost the page computes: Inkling-Small, Qwen3.8-Flash-Next, Ling-3.0-flash, Qwen3.5-397B, Qwen3.8-27B, GLM-4.7-Flash and Qwen3-32B. Read 2026-09-16. - `logs256/`: the full server log and the short, warm and deep response bodies for Gemma-4-26B, GLM-4.7-Flash, Qwen3.8-Flash-Next and Ling-3.0-flash; and (2026-09-26 fix) the full server log and the deep response body for Inkling-Small and Qwen3.5-397B, whose measured cells the page now prints. The Qwen3.5-397B log records its speculative decoding (`--spec-type draft-mtp`, draft acceptance lines); its deep body carries `draft_n` 29 and `draft_n_accepted` 29. The deep `*.req` prompt bodies (800 KB to 1 MB of synthetic text each) are omitted. ### `wave2/`: the 2026-09-13 wave-two check - `WAVE2_SUMMARY.json`: one entry per model per arm: verdict, card use, speaking, the cold and warm times to first content, the reading rate and the code hits for the 48,000 and 96,000-token legs, and the tool-call result. Redaction: the `key` field held internal shorthand and was renamed to public model names. This is a derived summary; the per-model records behind it do not ship, because they carry the private runtime's tool and agent names throughout. - `CORRECTION.md` (revised 2026-09-26 fix): the annotation of the `minimax-m2.7` entry, with the per-leg fields the page uses (first content 1,662.91 s, whole leg 1,682.9 s, 53,581 prompt tokens, 160 completion tokens; the 96,000-token leg's 1,800.4 s with no answer events), the corrected measurements, the harness fix and the two other instances of the defect. ### `minimax-m27/`: the MiniMax M2.7 audition, 2026-09-13, and the budget test, 2026-09-15 - `MEASUREMENTS.md`: the measurement tables of the audition, extracted from the session write-up: file and shard verification, header read, chat template read, the two served rungs, the coherence probe and its empty answer (with three sentences of the 4,096-token answer), the four thinking-switch variants, the tool call, the budget-test rows, and the Ling, Qwen3-32B and Qwen3.6-27B tables behind three other entries. It also carries the machine line (CPU and RAM). No measured number was changed in the extraction. - `budget/budget_probe_log.txt` and `budget/probe_budget.py`: the probe's own log of the `--reasoning-budget` test of 2026-09-15, and the probe that wrote it. - `gguf_header.py`, `gguf_header.txt`: the standalone header reader used during the audition and the raw header dump for the MiniMax M2.7 file, read before any load. - `ling_gguf_header.txt`, `ling_chat_template.txt`: the Ling-3.0-flash header dump and chat template behind the architecture and thinking-switch claims. - `qwen36_upstream_header.txt`: the header dump of the upstream Qwen3.6-27B file, including `qwen35.rope.dimension_sections = [11, 11, 10, 0]`. - `cpu_load_test.log`: the processor-only load proof for that file (`model loaded` at 6.39 s, graphics initialisation failed by design). - `hf_config.json`, `hf_tree.json`, `hf_tree_ling.json`, `hf_tree_Qwen3.6-27B-GGUF.json`: public repository metadata used for the byte-for-byte shard verifications (`hf_tree.json` sums to 108,413,781,312 bytes over the four UD-IQ4_XS shards). - `probe.py`, `recall_probe.py`: the speed and recall probes. ### `minimax-m27/relaunch/`: the 2026-09-19 re-launch measurements - `m27_sweep.sh`, `run_warm.sh`: the harness scripts, paths redacted. The sweep script's WHY comment records that the 13 September gate ran with `-cmoe -ub 128` copied from MiniMax M3's start script with no recorded rationale (2026-09-26 fix: the redaction of that one path now keeps its public model directory name, `MiniMax-M3`, because the page's claim rests on it). - `A_ub128_cmoe.json`: the control at the inherited `-ub 128` (every expert in system memory, 131,072 window, 8-bit cache, `-b 4096`): 32.278 t/s reading a 3,658-token prompt, 9.92 t/s decode on a 32-token answer, 22,530 MiB on the card. - `PHASE_B.console.log`: the `-ub` ladder on the 3,658-token prompt (128 to 2,048) and the placement ladder at `-ub 4096` on 43,909 tokens (every expert in RAM, then 61, 60, 59 and 58 of 62 layers' experts in RAM), with card use and 32-token decode rates. - `CONFIRM.console.log` (2026-09-26 fix): the `-ub 4096` read of the same 3,658-token prompt, every expert in RAM: 564.0 t/s, 24,359 MiB, 9.3 t/s decode on 32 tokens. - `DEPTH4096.console.log` (2026-09-26 fix): the 43,909-token read at `-ub 4096` with every expert in RAM: 640.2 t/s in 72.1 s, 24,296 MiB. - `one_ub4096_59_ctx131072.json`, `one_ub4096_58_ctx131072.json`: the 43,909-token reads at the installed placement (59 of 62: 656.5 t/s, 70.13 s, 29,094 MiB) and the fastest placement tried (58 of 62: 666.1 t/s, 69.1 s, 30,695 MiB). - `WARM.console.log`, `WARM.server.log`: at `--n-cpu-moe 59`, the 85,763-token cold read (server prompt eval 145,476.38 ms at 589.53 t/s; the whole request 147.49 s of client wall time with a 16-token answer), then the save-and-restore across a real server restart: 85,778 tokens saved, 85,778 restored, a 2.19 s restore call, and the identical request after it in 2.17 s with 1 token counted as new work. - `192K.console.log`, `one_ub1024_cmoe_ctx196608.json`: the load ladder at the native 196,608-token window with every expert in RAM: fails at `-ub 4096`, aborts at `-ub 2048`, loads at `-ub 1024` (31,560 MiB, 188.8 t/s on the 3,658-token prompt, 10.3 t/s decode on 32 tokens). See item 4 above for the "FIT" labels. ### `minimax-m27/gate0921/`: the 2026-09-21 gate re-run and the installed script's tests - `M27_GATE_SUMMARY.json`: a derived summary of the re-run through the same agent framework as wave two: PASS, 3 of 3 codes at 53,584 and 103,931 prompt tokens in 122.1 and 251.2 seconds (whole-leg times), decode 8.6 and 7.2 on 180- and 182-token completions, the tool-call legs, the warm legs. The per-leg records do not ship (they carry the private runtime's tool and agent names throughout). - `server.log`: the installed launch script's banner (`ctx 131072 · experts in RAM 59 of 62 · -b 4096 -ub 4096 · q8_0 KV`) and its restart test: after the restart the server evaluated 1 prompt token (136.79 ms) and 16 answer tokens, total 1,969.28 ms, stop count `n_tokens = 42898`. - `needle48k.json`: the acceptance test's recall probe through the installed script, 43,679 tokens in 108.9 seconds, 3 of 3. `second.log`: the second-instance refusal from the same test. ### `minimax-m3/`: the MiniMax M3 measurements, 2026-09-19 to 20 - `MEASUREMENTS.md` (2026-09-26 fix to sections 4 and 5): the extracted measurement note: the zero-indexer-tensors inventory, the micro-batch sweep on the old file, the correct-file figures, the 262,144 load ladder and the two-machine attempt, with the decode-artifact caveat. Two sentences were corrected against the logs and say so in place: the 262,144 failure is a compute buffer (the log does not attribute it to the indexer), and the two-machine run was stopped, not a server crash. - `A_ub128_cmoe.json`, `A_ub512_cmoe.json`, `A_ub1024_cmoe.json`, `A_ub2048_cmoe.json`, `M3_128K_ub4096.prefill.json`, `M3_128K_ub4096.decode.json`: the sweep on the old dense-fallback file, 2026-09-19 (22.5615 to 393.7783 t/s on 3,658-token prompts) and its decode. - `DOWNLOAD_Q2_STATUS.txt`: the byte-for-byte verification of the correct file, bartowski's Q2_K_L (4 shards, 153,086,988,768 bytes, 2026-09-20 02:23). Every correct-file measurement below is later than this verification. - `msa_q2kl_128k_ub2048.prefill.json`, `msa_q2kl_128k_ub2048.decode.json`, `msa_192k_ub512_4k.json`: the correct file's prompt evaluation on 3,658 tokens (147.5406 t/s at 131,072 and `-ub 2048`; 72.635 t/s at 196,608 and `-ub 512`) and decode (8.8995 t/s on an 82-token answer, 8.9733 on a forced 128-token generation). - `needle_msa_q8_0.json`, `needle_msa_192k.json`: the 3-of-3 recall runs on 58,307 prompt tokens at both windows: whole requests of 300.2 s with 165 generated tokens and 856.3 s with 207. They carry no prefill field; the page's 194 and 68.1 are prompt tokens divided by those wall times, which include generation. - `msa_ctx262144_cmoe_ub2048.server.log`: the 19,619.15 MiB (20,572,169,216-byte) compute-buffer allocation failure. `MSA_256K.console.log`, `MSA_256K_tiny.console.log`: the 262,144 ladder (fails at 2,048, 1,024, 512 and 256; loads at 128 with 31,525 MiB). `msa_256k_ub128.prefill.json`: that rung's 24.228 t/s on 3,658 tokens. - `split_msa_128k_ub1024.prefill.json`, `split_msa_128k_ub1024.decode.json`: the two-machine attempt: 17.5804 t/s reading 3,658 tokens (208.07 s), and the decode attempt's dropped connection. - `split_ctx131072_vl36_ub1024.server.log` (2026-09-26 fix): that run's server log, naming the file (`MiniMax-M3-IQ3_XXS-00001-of-00005.gguf`), the 17.58 t/s prompt eval, and the interrupt that stopped the run during the speaking probe. - `DOWNLOAD_STATUS.txt` (2026-09-26 fix): the verification of that file, bartowski's IQ3_XXS (5 shards, 180,011,108,960 bytes, 167.6 GiB, 2026-09-20 01:17). ### `batch-sweep/`: the 2026-09-20 to 21 batch-size sweep, this page's models - `ling_base.result`, `ling_ub4096.result`: the Ling-3.0-flash 2026-09-20 reads at a 131,072 window: 238.2 t/s on 47,992 tokens at default flags (7,233 MiB at load) and 1,140.4 at `-ub 4096`. - `ling_ub2048_confirm.result`, `ling_ub2048_confirm.vram`: the shipped setting (`-b 4096 -ub 2048`) at the 262,144 window, 2026-09-21: 727.7 t/s on 47,986 tokens, 694.6 on 149,711, 448.6 on 3,005, 8,824 MiB at peak. - `ling_shipped_262k_A.result`: the withdrawn `-ub 4096` (then the script's default) at the same window: 1,173.9 t/s on 48,084 tokens and 1,067.6 on 150,153, five reads all correct, 10,050 MiB at peak. - `w_ling_256k.log`: the original 2026-09-20 crash (the first request after a load at 262,144). `Ling-3.0-flash_ub4096_CRASHPROMPT.log` and `.result`: the 2026-09-21 reproduction with the exact 3,007-token prompt at 262,144, CUDA illegal memory access, the client side sees the connection drop. - `Ling-3.0-flash_2048.result`, `_1024.result`, `_512.result`, `Ling-3.0-flash_skip.result`: the same prompt passing at the lower micro-batches (444.0, 325.7 and 201.5 t/s at 262,144) and through the shipped script. - `stress_ub4096_s777.SUMMARY.json`, `stress_ub2048_s777.SUMMARY.json`, `ledger_ub4096_s4242.SUMMARY.json`, `ledger_ub2048_s4242.SUMMARY.json`: 60 prompts of real code and prose and 50 ledger prompts, run at each of the two settings, zero crashes. - `diff_Ling-3.0-flash_crashfix.txt`: the change that withdrew `-ub 4096`, with its comment block. That comment also describes a same-day diagnostic (other windows and `-ub 3072`); the diagnostic's own logs ship with the batch-size study, not here, and the page does not state those rungs as measured. - `repro_ling_crash.py`: the standalone reproducer (standard library only, loopback only). The crash and pass outcomes are the logs above, not runs of this script. - `flashnext_base.result`, `flashnext_ub2048.result`, `flashnext_ub8192.result`: Qwen3.8-Flash-Next at a 131,072 window. See item 10 above; the page does not print them. - `GLM-4.7-Flash_*.result`, `Gemma-4-26B-A4B_*.result`, `gemma-4-31b-it-qat_*.result` and the three `*_chain.txt`: the measured-and-not-adopted micro-batch rungs for the three card-resident models, ~48,000-token prompts at their long windows, with the chain files recording where each ladder stopped. GLM-4.7-Flash's 2,048 rung returned an empty visible answer with the code in its reasoning (`needle_in_answer: false`). ### `flashnext-context/`: the 2026-09-20 context study - `reliability.out`: the console summary of the five sittings (the short control, 256K at f16 and at q8_0, 384K, 512K), one pass each, temperature 0. - `hard_f16_256k.*`, `hard_q8_256k.*`, `hard_q8_384k.*`, `hard_q8_512k.*`, `hard_q8_short_control.result`: the per-sitting result files and server logs. The 512K sitting is the 459,911-token, 3-of-3 run at 171.90 t/s (44.6 minutes). The `[q3 INTEGRATE]` rows are the integration question: `N` at every sitting, including the 5,974-token short control, always as a 526-token follow-up with the ledger cached. - `q/q8_512k/meta.json`, `q/q8_512k/q1.out.json`: the answer key and the raw answer body for the deepest run. (2026-09-26 fix: `q1.out.json` still carried its `id` and `system_fingerprint` fields; both were removed.) - `e_512k_norope.log`, `e_512k_norope.vram`, `b_512k_unlock.vram`, `a_256k_base.vram`, `a_256k_base.log` (the log added in the 2026-09-26 fix): the rope-flag trap. See item 11 above. - `run_hard.sh`: the probe driver, paths redacted. ### `two-box/` - `maverick-run12-excerpt.log`: lines 74 to 101 of the two-box campaign log for that night, verbatim apart from the redactions below. The full log, and the header read behind the file's 167.21 GiB, ship with the two-box study's own data package. ## Redactions, stated plainly Applied to every file in this package, with counts over the shipped tree. Records shipped before 2026-09-26 (the context sweep, 256K sweep, install run, audition, budget test, wave-two summary, two-box excerpt): - **Every path on our machines was replaced in full** with ``, 35 occurrences across 15 files, keeping only a model file's own name where a record named one. - Network addresses and host aliases became `` (23 occurrences, 11 files) and ports became `` (28 occurrences, 14 files). Loopback keeps its form, `127.0.0.1:`, 5 occurrences, per this site's precedent. - Process ids became ``, 38 occurrences across 11 files. - Serving aliases inside response bodies became `` and `--alias` launch flags became ``. - One line of a probe script's own documentation named the machine and was genericised (`minimax-m27/gguf_header.py`). Records added on the morning of 2026-09-26 (the re-launch, gate, M3, batch-sweep and context-study files): - **Paths**: 62 full internal paths became ``, 18 more with the model file's own basename kept, and 3 relative internal paths. - **Network**: 6 `:` pairs, 18 bare `` replacements, 28 loopback `127.0.0.1:` (form kept per precedent), 12 other port literals. - **Process ids**: 3 became ``. - **Names**: internal serving aliases and systemd unit names became ``; the agent framework is named nowhere in these files; probe, tool and agent names from the private runtime appear in no shipped file (`M27_GATE_SUMMARY.json` was written without them). - **One parenthetical naming an internal coordination document** was removed from `minimax-m27/relaunch/m27_sweep.sh`'s refusal message. Records added or changed in the 2026-09-26 fix (nine new files, seven changed): - **Paths**: 10 ``, 9 of them keeping the model file's or binary's own name. One earlier redaction was relaxed: `m27_sweep.sh`'s WHY comment now reads `/MiniMax-M3/start_server.sh`, keeping the public model directory name only. - **Network and processes**: 4 ``, 9 `` (2 of them loopback `127.0.0.1:`), 8 ``, 2 ``, 2 ``. - **Identifiers**: `id` and `system_fingerprint` removed from three response bodies (the two new deep bodies and `q1.out.json`). All 44 response bodies in the package now lack both fields; `choices`, content, `usage` and `timings` ship as the server wrote them. These are outputs of open-weight models served on our own machine. - **Our warm-restore program and a build directory**: the warm-wake study calls the program that saves and restores a server's slot "a small program of our own" and does not name it; this package named it in four places. Two of its log lines inside `batch-sweep/Ling-3.0-flash_ub4096_CRASHPROMPT.log` were replaced by a bracketed note saying a line was removed, and its name became `` in `minimax-m27/gate0921/server.log` (the launch banner) and `minimax-m27/relaunch/WARM.console.log`. One comment in `minimax-m27/relaunch/m27_sweep.sh` named an internal build directory; it now reads ``. Punctuation normalisation, disclosed because it is a change: em dashes in copied records were replaced with hyphens (32 occurrences in the morning's files, 4 in the fix's new files, and 1 written as a JSON unicode escape inside `Ling-3.0-flash_ub4096_CRASHPROMPT.result`). No number, token count or verdict was touched. ## What does not ship, and why - **Our session write-ups and working notes**: the MiniMax M2.7 audition write-up, the budget test's verdict paragraph, the MiniMax M3 session write-up, the two overnight notes behind the 700-to-900 correction, and the internal audit and plan documents behind the harness-defect correction. They carry internal names and coordination detail. The measurement tables were extracted into `minimax-m27/MEASUREMENTS.md`, `minimax-m27/budget/budget_probe_log.txt` and `minimax-m3/MEASUREMENTS.md`; the 700-to-900 correction is quoted on the page. - **The MiniMax M3 start script** that M2.7's `-cmoe -ub 128` was copied from: operator inventory. The fact the page uses is in `m27_sweep.sh`'s WHY comment. - **The per-leg records of both agent-framework checks** (wave two, the 2026-09-21 re-run) and of the 2026-09-12 gate round behind the two other defect instances: they carry the private runtime's tool and agent names throughout. Derived summaries ship instead, and `wave2/CORRECTION.md` lists the per-leg fields the page uses. - **The deep-prompt request bodies** named above, the context study's synthetic ledger prompt bodies, and any request capture of an agent-framework prompt. - **The batch sweep's own README and queue files**, and its diagnostic and served-window records for models or rungs this page does not print: they ship with the batch-size study. - **The full two-box campaign logs and header reads**, which ship with the two-box study. ## Verification The packaging rule's leak sweep (one case-insensitive grep over the page and this whole tree for vault and home paths, staging folder names, the hostname and tailnet name, link and bridge addresses, the laptop's names, service and unit names, operator keys, the private runtime's vocabulary and tool names, and the words that name anything of the household) was run after the last edit of the 2026-09-26 revision: **no matches** (exit status 1), with a positive control on a file that does contain those strings. The command, verbatim, is recorded with the page's self-check; it is not reprinted here because the pattern itself would match. A second sweep for response-body identifiers (the `id` and fingerprint fields, and the request-id prefix) over every `.json` here: **no matches**. The only IPv4 address anywhere in the tree is loopback, `127.0.0.1`. The em dash sweep (the character, the HTML entity and its JSON unicode escape) over the tree: **no matches**. Every file named in `NUMBERS.md` exists in this package, and every response body parses as JSON. Independent measurements. Not affiliated with, endorsed by, or connected to MiniMax, Z.ai, Google, inclusionAI, Alibaba Group, DeepSeek, Thinking Machines Lab, Meta, Mistral AI, Moonshot AI, Ollama, Unsloth, bartowski, Resemble AI, NVIDIA, ASUS, Intel, or the llama.cpp project.