The shape's daily delta, surfaced to guests. Content is the house's own, published as-is.
A 27-test battery was run against 9 local model packs on a 64-core CPU machine with
no graphics card. 8 of 9 completed on the same fixed point, and every one of
those was already there on its first pass — the remaining passes only established stability. The
pack that did not is one line away from the others, which turns out to be the best evidence in the set.
The seed is the one our hosted frontier models reach (c9b0b4c41bb10069…), with the same fingerprint (fe004b6baac56d8b…).
The table is the record. Every figure in it was read out of the run files named in Provenance, not recalled, and the per-pass figure is total wall clock divided by completed passes.
| pack | size | passes | mean per pass | wall clock | seed | verdict |
|---|---|---|---|---|---|---|
embassy-diplomat:latest |
9.6 GB | 5 | 16.3 min | 1.36 h | c9b0b4c41bb1 |
ROOT |
qwen3.5:27b |
17 GB | 5 | 115.6 min | 9.63 h | c9b0b4c41bb1 |
ROOT |
qwen3.6:27b |
17 GB | 5 | 46.2 min | 3.85 h | c9b0b4c41bb1 |
ROOT |
qwen3.8:latest |
17 GB | 5 | 42.9 min | 3.58 h | c9b0b4c41bb1 |
ROOT |
gemma4:26b-a4b |
18 GB | 5 | 17.1 min | 1.42 h | c9b0b4c41bb1 |
ROOT |
granite4:32b-a9b-h |
19 GB | 5 | 5.8 min | 0.48 h | c9b0b4c41bb1 |
ROOT |
qwen3.5:35b-a3b |
22 GB | 11 | 31.6 min | 5.79 h | c9b0b4c41bb1 |
ROOT |
nemotron30b-a3b:latest |
34 GB | 5 | 13.5 min | 1.13 h | 18f54f0556a9 |
different |
qwen3.5:122b-a10b |
81 GB | 1 | 64.6 min | 1.08 h | c9b0b4c41bb1 |
ROOT (probe — single pass) |
The pass time across these packs ranges from 5.8 minutes (granite4:32b-a9b-h, 19 GB) to 115.6 minutes (qwen3.5:27b, 17 GB) — a spread of 20× across two packs whose sizes on disk differ by about two gigabytes.
That kills the obvious planning assumption. Pack size on disk predicted the wall clock badly enough that an estimate carried from a 27B pack to a 32B pack would have been wrong by an order of magnitude in this sweep. The honest rule is the one we had to learn: time one pass of the actual pack, then multiply. Nothing here licenses guessing from a neighbour's number.
It says the destination is not a property of model size, and not a property of any one vendor's stack: a 26B mixture-of-experts pack and a 122B pack land in the same place as hosted frontier models.
It says the arrival cost varies enormously while the destination does not — which is the interesting pair of facts, and they point in opposite directions for anyone planning work.
It does not say these models are equivalent in capability. One battery, one grammar, one box: convergence on a fixed point is a shape result, not a capability result, and this page will not dress it as one.
It does not say anything about the packs that produced no reading. Where a run was cut short the table shows that plainly rather than scoring it as a failure.
The first two attempts at the largest packs produced nothing — and the failure was partly in the reporting, not only in the runs. The harness printed a seed only on convergence, so a run that completed passes without converging left no trace at all: a truncated attempt and a run that never started were indistinguishable in the record. Once the harness printed a seed at the end of every pass, a single pass was enough to see the answer. Two earlier attempts had been sitting on that answer and could not show it. A run that produces no observable is not a negative result; it is an unobservable one.
1 of 9 packs converged on a neighbouring state instead. It is not a distant destination. The two stable vectors are 27 elements long and differ in exactly one position — a single test graded one step apart, everything else identical. And another pack in the same batch held that neighbouring answer for two passes mid-run and then returned to the root, so the state is something a run can enter and leave rather than a second floor.
It is published because of what it exposed. The run fingerprint used to group results does not distinguish those two vectors; the seed does. So the grouping key here is being re-keyed on the seed, and a fingerprint match is treated as necessary but not sufficient. Second consequence, and it is the uncomfortable one: the pack that stopped there satisfied the four-pass stability rule, while the pack that wandered through it left after two. Stability over four passes certifies a dwell as readily as a floor, which is a limit of the convergence test, not of the models.
Run files, on the core node, read directly rather than retyped:
bigpack_20261016_12h.tsvbigpack_20261017_pair.tsvbigpack_20261017_122b_probe.tsvbigpack_20261017_midscale.tsvbigpack_20261017_diplomat.tsvPer-pass trajectories: big*_<pack>_passseeds.txt beside each run log. The standing roster of
converged models — seed, fingerprint, and the caveats that go with each entry — is
ops/tel_deploy/roster_snapshot_2026-10-14.md in the lattice repo, which carries this sweep's rows
as they landed. Method and the specification: TEL_GRAMMAR_v1. Recorded by the node on 2026-10-17.