This is the week’s build log in one piece: the two results in full, the rest in a paragraph each, every number linked to the artefact that produced it. The daily version is the build log in the evidence repository.


The runtime phoned home every four hours. The switch that stops it works, measured.

A colleague told us a local Qwen on his machine seemed to open connections while he used it: endpoints in China, then a host in France, “it just searches for data in a search engine.” That is the one failure the local-first thesis cannot survive. If inference is local, nothing leaves. We say so on this site and in Aprimerose’s material, and in exp_054 we had measured it for our own code. We had never watched a third-party model’s runtime while it ran. So we did, in exp_057, registered in the morning and run before lunch.

Model weights are inert. A .gguf file is tensors; it cannot open a socket. If a connection appears during “local inference”, it comes from the runtime that loads the weights (here Ollama’s daemon and the runner it spawns per model), or from a harness around it that reads the model’s output and executes a search request it finds there. The experiment was built to tell those apart: lsof every 0.2 s over the daemon and its runners, nettop every second over every process on the machine, and for each of ten subjects three watched windows: the load, 30 s idle, and ten frozen prompts, five of them written to tempt a search. Eight Qwen builds, gemma4:26b as the in-house control, and the embedding model. Three controls had to pass or the run was void; all three did.

Thirty windows, 29 minutes, no runner opened anything. The colleague’s report did not reproduce here. No model’s answer contained a tool call or a search request; they said they could not browse, or guessed and said so.

One connection did appear, and the daemon’s own log settles it to the millisecond. During one 35B model’s window, the Ollama daemon itself opened a TLS connection to the address ollama.com resolves to: 2,001 bytes out, 7,639 in. Fifty-eight milliseconds earlier the log had written that its “model recommendations” refresh ran and succeeded, and ~/.ollama/cache/model-recommendations.json was rewritten in the same second. The log on this machine holds 372 of those refreshes since 5 March, one every 3.5 to 4.6 hours and one at every daemon start: about six contacts with ollama.com a day, generating or not. It fell into that model’s window by arithmetic.

We had registered two claims: that plain inference opens no endpoint absent at idle, and that no external connection first appears inside a prompt’s window. Both fail. They were written to detect an appearance, not a cause, and they detected one. A 30 s idle baseline cannot see a four-hour timer. The explanation sits beside the fails in the record, not in place of them.

The switch, measured

Ollama documents an off switch for its cloud features: OLLAMA_NO_CLOUD=1, or "disable_ollama_cloud": true in ~/.ollama/server.json, then a restart, after which the log reads Ollama cloud disabled: true. The documentation does not say whether that covers the recommendations refresh. So we registered a second run before making it: set the flag, restart, watch the daemon’s tree and the whole machine with nothing generating for longer than the longest interval we had seen, and count any refresh line in the log. On 1 November the production daemon was restarted with the flag at 07:25 and watched until 12:25. One refresh came due in that window, at 11:20.

a call to ollama.com, and the file it rewrotea scheduler line in the daemon logwatched window
Watched: 07:25:51 to 12:25:57, the daemon tree every 0.2 s, every process every 1 swatched for five hours, nothing generating06:0007:0008:0009:0010:0011:0012:00daemon lognetworkcache file06:11:43 — refresh ran and succeeded (cloud on), next in ~4 h06:11 — TLS connection to the address ollama.com resolves to06:11:43 — model-recommendations.json rewritten, 2,041 bytes06:11 last refresh, cloud on07:25:51 — restart; log: Ollama cloud disabled: true; 'cache sleep scheduled', 1 ms after listening07:25 restart, cloud off11:20:48 — 'cache sleep scheduled' again: the timer ran out; no fetch line11:20 refresh dueno connection from the daemon tree; no process reached ollama.comCache file kept its 06:11:43 write through both scheduler linesunchanged since 06:11
1 November, CET. Before the flag, a refresh means three things at once: a log line, a connection and a rewritten cache file. After it, the log line at 11:20 came alone. Hover a mark for the record behind it.
Result
The log beside the new startOllama cloud disabled: true
Connections from the daemon’s tree in five hours, sampled at 0.2 sits local listener, nothing else, including at 11:20
Any process on the machine reaching the address ollama.com resolves tonone
The recommendations cache file, rewritten by every past refresh in the same secondkept its 06:11 write, the last refresh before the flag
Refresh lines in the logtwo “cache sleep scheduled” lines: one at the start, one when that wait ran out at 11:20

Those two lines make one of our registered claims fail: we had written “no refresh line after the restart”, because in every earlier log the line followed a successful fetch. With cloud disabled, the scheduler keeps its timer and skips the request. The start line was written one millisecond after the daemon began listening, too soon for a TLS exchange, and nothing on the network or on disk matched either line. The verdict stands as registered, with the cause next to it. For anyone checking a machine by its logs: a log line is not a call. The cache file’s write time and the network are the evidence.

What this does not show. Sampling at 0.2 s and 1 s can miss a shorter connection; the packet-capture arm did not run. The request’s content is unseen: two kilobytes out is a TLS handshake plus one HTTP request, and whether it carries anything about this machine beyond a timestamp, the bytes do not say. One runtime, Ollama 0.40.2; LM Studio, a bare llama.cpp server, MLX and desktop chat apps are untested. The harness layer, an app executing a model’s search request, is the layer most likely behind the report we started from, and it is the next experiment. Nine quiet models are nine quiet models, not “these models are safe”.

So what. “Nothing leaves the machine during local inference” is false for Ollama 0.40.2 as installed. The sentence our pages can carry is the measured one: the model sends nothing, and the runtime sends nothing once its cloud features are off. Every place we wrote the short version gets the long one. It is a documented feature with a documented switch; what was undocumented is the cadence, and that the switch covers the refresh. A baseline must outlast the longest timer. And read the daemon’s log before guessing at an address: it named the moment, the purpose and the URL.


The reader’s 22 empty answers were loops

In exp_055 we built 220 messages against a fictional person’s public footprint, 110 lures and 110 genuine, and asked gemma4:26b, alone on the Mac mini, to grade each one. It beat our engineered check (balanced accuracy 0.882 against 0.814). It also returned 22 empty answers. Every one of them took about 150 seconds, which is 8,192 tokens at the mini’s speed: the model thought until the context was full, and no answer token was ever written. Under the rule we had registered, an empty answer counts as “not flagged”, so the model’s catch rate was understated and its false-alarm rate flattered.

Our October post on a model that thinks met the same wall on a different task and drew the obvious lesson: thinking needs room. Four of thirty calls stopped before answering at 8K; at 16K they finished. Before building the next thing on this reader, we tested that lesson properly, in exp_060, registered before the kit was written.

Three fixes, because the runtime offers three

Ollama 0.40.2 has no thinking budget for this model. ollama show lists the thinking levels as false and true; a request for "low" is accepted and ignored. So a budget has to be built around the model, and the mechanisms the runtime does offer are these:

ArmThe one changeThe question
Tthinking offdoes the reader need to think at all?
Fthink as before; if the answer comes back empty, one retry of the same prompt with thinking offcan the loss be closed without touching any answer that worked?
Cthe same call with a 32,768-token contextdoes the thinking finish when given room, and what does it then say?

Everything else was exp_055’s: the same 220 messages, the same prompt, temperature 0, seed 55, one call per message. Every call recorded what exp_055 had not: the token counts, why generation stopped, and the thinking text.

model step, thinking onmodel step, thinking offcode step

T — thinking off

modelgemma4:26b, thinking off, 8Kthe same prompt; answers in about 3 s
→ a verdict on every message

F — think, then retry without

1 · modelgemma4:26b, thinking on, 8Kexp_055's call, unchanged
2 · codeDid an answer come out?a parsable verdict ends here; an empty one goes on
3 · modelthe same prompt, thinking offone retry, only for the empties (21 of 220)
→ a verdict on every message; the 199 that answered first are untouched

C — more room

modelgemma4:26b, thinking on, 32Knum_ctx 32,768; the daemon reloads the model for it
→ 198 verdicts, 22 empty
Every arm sends exp_055's prompt with the footprint, the registry facts and the message, at temperature 0 with seed 55. Each changes one thing. Arm F's retry is arm T's call, made only when the first pass returns nothing.

Four controls had to pass. The baseline had to reproduce (F’s first pass gave exp_055’s verdict on 193 of 193 unchanged messages, and all 20 unchanged empties were empty again). The wall had to be the cause (all 21 empty first passes ended at exactly 8,192 tokens). Nothing could leave the machine (every connection went to the local daemon). And arm C had to run at 32K (the daemon reported 32,768 and 23 GB, against 18.7 GB at 8K).

What we found

AnsweredLures caughtGenuine flaggedBalanced accuracyMedian95th percentile
exp_055, thinking on, 8K198 of 22092 of 1108 of 1100.88233.6 s
T, thinking off220101240.8502.8 s3.4 s
F, think, then retry without22010090.91436.9 s162.3 s
C, 32K context1989560.90536.8 s884.8 s
Opus, exp_05522010560.950

By the rule we registered, the arm that answers at least 218 of 220, scores at least 0.882, and stays under 45 s median and 180 s at the 95th percentile with no call error, wins. F is the only arm that passes all of them. It takes the gap to Opus from 0.068 to 0.036.

F, the arm chosenthe other two armsfor reference: exp_055's run of the same call, and Opus
Answered, of 220
higher is better
exp_055, 8KAnswered, of 220: exp_055, 8K 198198T thinking offAnswered, of 220: T thinking off 220220F think, retryAnswered, of 220: F think, retry 220220C 32KAnswered, of 220: C 32K 198198Opus, exp_055Answered, of 220: Opus, exp_055 220220
Lures caught, of 110
higher is better
exp_055, 8KLures caught, of 110: exp_055, 8K 9292T thinking offLures caught, of 110: T thinking off 101101F think, retryLures caught, of 110: F think, retry 100100C 32KLures caught, of 110: C 32K 9595Opus, exp_055Lures caught, of 110: Opus, exp_055 105105
Genuine messages flagged, of 110
lower is better
exp_055, 8KGenuine messages flagged, of 110: exp_055, 8K 88T thinking offGenuine messages flagged, of 110: T thinking off 2424F think, retryGenuine messages flagged, of 110: F think, retry 99C 32KGenuine messages flagged, of 110: C 32K 66Opus, exp_055Genuine messages flagged, of 110: Opus, exp_055 66
Balanced accuracy
higher is better
exp_055, 8KBalanced accuracy: exp_055, 8K 0.8820.882T thinking offBalanced accuracy: T thinking off 0.8500.850F think, retryBalanced accuracy: F think, retry 0.9140.914C 32KBalanced accuracy: C 32K 0.9050.905Opus, exp_055Balanced accuracy: Opus, exp_055 0.9500.950
An empty answer counts as not flagged, as registered, so the two arms with 198 answers carry 22 misses or clears they never made. Balanced accuracy is the mean of the catch rate and one minus the false-alarm rate.

More room moved the empty answers. It did not close them.

At 32K, 22 answers were still empty. Not the same 22.

  • 11 of the original 22 finished, all 11 right. None of them needed the room: they finished in 1,026 to 5,304 generated tokens, inside the 8K wall they had hit the first time.
  • The other 11 thought to the 32K wall: about 31,800 tokens and 885 seconds each.
  • 11 messages that had answered fine at 8K now failed. Ten thought to the 32K wall. The eleventh was cut off by Ollama after 96 seconds with prediction aborted, token repeat limit reached, the runtime’s own loop guard.
answeredemptyempty at 8K, finished at 32Kanswered at 8K, empty at 32K

At 8K (exp_055)

X001 (lure): answered at 8KX002 (benign): answered at 8KX003 (lure): empty at 8KX004 (lure): answered at 8KX005 (benign): empty at 8KX006 (benign): answered at 8KX007 (benign): answered at 8KX008 (benign): answered at 8KX009 (lure): answered at 8KX010 (benign): answered at 8KX011 (benign): empty at 8KX012 (lure): answered at 8KX013 (lure): answered at 8KX014 (benign): answered at 8KX015 (lure): answered at 8KX016 (lure): answered at 8KX017 (benign): answered at 8KX018 (benign): answered at 8KX019 (lure): answered at 8KX020 (lure): empty at 8KX021 (benign): answered at 8KX022 (lure): answered at 8KX023 (lure): answered at 8KX024 (lure): answered at 8KX025 (lure): answered at 8KX026 (lure): answered at 8KX027 (lure): answered at 8KX028 (lure): answered at 8KX029 (benign): answered at 8KX030 (lure): empty at 8KX031 (lure): answered at 8KX032 (lure): answered at 8KX033 (benign): answered at 8KX034 (benign): answered at 8KX035 (lure): answered at 8KX036 (lure): answered at 8KX037 (lure): answered at 8KX038 (lure): answered at 8KX039 (benign): answered at 8KX040 (benign): answered at 8KX041 (benign): empty at 8KX042 (lure): answered at 8KX043 (benign): answered at 8KX044 (lure): answered at 8KX045 (benign): answered at 8KX046 (lure): answered at 8KX047 (benign): answered at 8KX048 (lure): answered at 8KX049 (benign): answered at 8KX050 (benign): answered at 8KX051 (lure): answered at 8KX052 (benign): empty at 8KX053 (benign): answered at 8KX054 (lure): answered at 8KX055 (lure): answered at 8KX056 (lure): answered at 8KX057 (lure): answered at 8KX058 (lure): answered at 8KX059 (benign): answered at 8KX060 (lure): answered at 8KX061 (lure): answered at 8KX062 (lure): empty at 8KX063 (lure): empty at 8KX064 (benign): answered at 8KX065 (lure): answered at 8KX066 (benign): answered at 8KX067 (benign): answered at 8KX068 (benign): answered at 8KX069 (lure): answered at 8KX070 (benign): answered at 8KX071 (lure): empty at 8KX072 (lure): answered at 8KX073 (lure): answered at 8KX074 (benign): answered at 8KX075 (lure): answered at 8KX076 (benign): answered at 8KX077 (benign): answered at 8KX078 (benign): answered at 8KX079 (benign): answered at 8KX080 (benign): answered at 8KX081 (lure): answered at 8KX082 (benign): answered at 8KX083 (lure): answered at 8KX084 (lure): answered at 8KX085 (benign): answered at 8KX086 (lure): answered at 8KX087 (lure): answered at 8KX088 (lure): empty at 8KX089 (lure): answered at 8KX090 (benign): empty at 8KX091 (lure): answered at 8KX092 (benign): answered at 8KX093 (benign): answered at 8KX094 (lure): answered at 8KX095 (lure): answered at 8KX096 (benign): answered at 8KX097 (lure): answered at 8KX098 (lure): answered at 8KX099 (lure): answered at 8KX100 (benign): empty at 8KX101 (benign): answered at 8KX102 (benign): answered at 8KX103 (benign): empty at 8KX104 (benign): answered at 8KX105 (benign): answered at 8KX106 (benign): answered at 8KX107 (lure): answered at 8KX108 (lure): answered at 8KX109 (lure): empty at 8KX110 (lure): answered at 8KX111 (lure): answered at 8KX112 (lure): answered at 8KX113 (lure): answered at 8KX114 (benign): answered at 8KX115 (benign): answered at 8KX116 (lure): empty at 8KX117 (lure): answered at 8KX118 (benign): answered at 8KX119 (benign): answered at 8KX120 (benign): answered at 8KX121 (lure): answered at 8KX122 (lure): answered at 8KX123 (benign): answered at 8KX124 (benign): answered at 8KX125 (benign): answered at 8KX126 (lure): answered at 8KX127 (lure): answered at 8KX128 (lure): answered at 8KX129 (benign): answered at 8KX130 (benign): answered at 8KX131 (lure): answered at 8KX132 (benign): answered at 8KX133 (lure): answered at 8KX134 (benign): answered at 8KX135 (lure): answered at 8KX136 (lure): answered at 8KX137 (lure): answered at 8KX138 (lure): answered at 8KX139 (benign): answered at 8KX140 (benign): answered at 8KX141 (benign): empty at 8KX142 (benign): answered at 8KX143 (benign): answered at 8KX144 (benign): answered at 8KX145 (benign): answered at 8KX146 (lure): answered at 8KX147 (lure): answered at 8KX148 (benign): answered at 8KX149 (lure): answered at 8KX150 (lure): answered at 8KX151 (benign): answered at 8KX152 (benign): answered at 8KX153 (lure): answered at 8KX154 (benign): answered at 8KX155 (lure): answered at 8KX156 (benign): answered at 8KX157 (lure): empty at 8KX158 (lure): answered at 8KX159 (benign): answered at 8KX160 (benign): answered at 8KX161 (benign): empty at 8KX162 (lure): answered at 8KX163 (benign): answered at 8KX164 (benign): answered at 8KX165 (lure): answered at 8KX166 (benign): empty at 8KX167 (benign): answered at 8KX168 (benign): answered at 8KX169 (lure): answered at 8KX170 (benign): answered at 8KX171 (benign): answered at 8KX172 (benign): answered at 8KX173 (benign): answered at 8KX174 (benign): answered at 8KX175 (benign): empty at 8KX176 (benign): answered at 8KX177 (benign): answered at 8KX178 (benign): answered at 8KX179 (lure): answered at 8KX180 (lure): answered at 8KX181 (lure): answered at 8KX182 (lure): answered at 8KX183 (lure): answered at 8KX184 (benign): answered at 8KX185 (benign): answered at 8KX186 (benign): answered at 8KX187 (lure): answered at 8KX188 (lure): answered at 8KX189 (benign): answered at 8KX190 (lure): empty at 8KX191 (lure): answered at 8KX192 (benign): answered at 8KX193 (lure): answered at 8KX194 (benign): answered at 8KX195 (lure): answered at 8KX196 (benign): answered at 8KX197 (lure): answered at 8KX198 (lure): answered at 8KX199 (benign): answered at 8KX200 (lure): answered at 8KX201 (benign): answered at 8KX202 (benign): answered at 8KX203 (lure): answered at 8KX204 (benign): answered at 8KX205 (benign): answered at 8KX206 (lure): answered at 8KX207 (lure): answered at 8KX208 (lure): answered at 8KX209 (lure): answered at 8KX210 (benign): answered at 8KX211 (benign): answered at 8KX212 (lure): answered at 8KX213 (benign): answered at 8KX214 (lure): answered at 8KX215 (benign): answered at 8KX216 (benign): answered at 8KX217 (lure): answered at 8KX218 (benign): answered at 8KX219 (benign): answered at 8KX220 (benign): answered at 8K

22 empty: 11 lures, 11 genuine messages.

At 32K (arm C)

X001 (lure): answered at bothX002 (benign): answered at bothX003 (lure): empty at 8K and at 32KX004 (lure): answered at bothX005 (benign): empty at 8K and at 32KX006 (benign): answered at bothX007 (benign): answered at bothX008 (benign): answered at bothX009 (lure): answered at bothX010 (benign): answered at bothX011 (benign): empty at 8K, finished at 32K, benign, rightX012 (lure): answered at bothX013 (lure): answered at bothX014 (benign): answered at bothX015 (lure): answered at bothX016 (lure): answered at bothX017 (benign): answered at bothX018 (benign): answered at bothX019 (lure): answered at bothX020 (lure): empty at 8K and at 32KX021 (benign): answered at bothX022 (lure): answered at bothX023 (lure): answered at bothX024 (lure): answered at bothX025 (lure): answered at bothX026 (lure): answered at bothX027 (lure): answered at 8K, ran to the 32K wallX028 (lure): answered at bothX029 (benign): answered at bothX030 (lure): empty at 8K, finished at 32K, suspicious, rightX031 (lure): answered at 8K, ran to the 32K wallX032 (lure): answered at bothX033 (benign): answered at bothX034 (benign): answered at bothX035 (lure): answered at bothX036 (lure): answered at bothX037 (lure): answered at bothX038 (lure): answered at bothX039 (benign): answered at bothX040 (benign): answered at bothX041 (benign): empty at 8K and at 32KX042 (lure): answered at bothX043 (benign): answered at bothX044 (lure): answered at bothX045 (benign): answered at bothX046 (lure): answered at bothX047 (benign): answered at bothX048 (lure): answered at bothX049 (benign): answered at bothX050 (benign): answered at bothX051 (lure): answered at bothX052 (benign): empty at 8K and at 32KX053 (benign): answered at bothX054 (lure): answered at bothX055 (lure): answered at bothX056 (lure): answered at bothX057 (lure): answered at bothX058 (lure): answered at bothX059 (benign): answered at bothX060 (lure): answered at bothX061 (lure): answered at bothX062 (lure): empty at 8K, finished at 32K, suspicious, rightX063 (lure): empty at 8K and at 32KX064 (benign): answered at bothX065 (lure): answered at bothX066 (benign): answered at bothX067 (benign): answered at bothX068 (benign): answered at bothX069 (lure): answered at bothX070 (benign): answered at bothX071 (lure): empty at 8K and at 32KX072 (lure): answered at bothX073 (lure): answered at bothX074 (benign): answered at 8K, ran to the 32K wallX075 (lure): answered at bothX076 (benign): answered at bothX077 (benign): answered at bothX078 (benign): answered at bothX079 (benign): answered at bothX080 (benign): answered at bothX081 (lure): answered at bothX082 (benign): answered at bothX083 (lure): answered at bothX084 (lure): answered at bothX085 (benign): answered at bothX086 (lure): answered at bothX087 (lure): answered at bothX088 (lure): empty at 8K, finished at 32K, suspicious, rightX089 (lure): answered at bothX090 (benign): empty at 8K, finished at 32K, benign, rightX091 (lure): answered at bothX092 (benign): answered at bothX093 (benign): answered at bothX094 (lure): answered at bothX095 (lure): answered at bothX096 (benign): answered at bothX097 (lure): answered at bothX098 (lure): answered at bothX099 (lure): answered at bothX100 (benign): empty at 8K, finished at 32K, benign, rightX101 (benign): answered at bothX102 (benign): answered at bothX103 (benign): empty at 8K and at 32KX104 (benign): answered at bothX105 (benign): answered at bothX106 (benign): answered at bothX107 (lure): answered at bothX108 (lure): answered at bothX109 (lure): empty at 8K, finished at 32K, suspicious, rightX110 (lure): answered at 8K, ran to the 32K wallX111 (lure): answered at bothX112 (lure): answered at bothX113 (lure): answered at bothX114 (benign): answered at bothX115 (benign): answered at bothX116 (lure): empty at 8K and at 32KX117 (lure): answered at bothX118 (benign): answered at bothX119 (benign): answered at bothX120 (benign): answered at bothX121 (lure): answered at bothX122 (lure): answered at bothX123 (benign): answered at bothX124 (benign): answered at bothX125 (benign): answered at bothX126 (lure): answered at bothX127 (lure): answered at bothX128 (lure): answered at bothX129 (benign): answered at bothX130 (benign): answered at bothX131 (lure): answered at 8K, ran to the 32K wallX132 (benign): answered at bothX133 (lure): answered at bothX134 (benign): answered at 8K, ran to the 32K wallX135 (lure): answered at bothX136 (lure): answered at bothX137 (lure): answered at bothX138 (lure): answered at bothX139 (benign): answered at bothX140 (benign): answered at bothX141 (benign): empty at 8K and at 32KX142 (benign): answered at bothX143 (benign): answered at bothX144 (benign): answered at bothX145 (benign): answered at bothX146 (lure): answered at bothX147 (lure): answered at bothX148 (benign): answered at bothX149 (lure): answered at bothX150 (lure): answered at bothX151 (benign): answered at bothX152 (benign): answered at bothX153 (lure): answered at bothX154 (benign): answered at bothX155 (lure): answered at 8K, ran to the 32K wallX156 (benign): answered at bothX157 (lure): empty at 8K, finished at 32K, suspicious, rightX158 (lure): answered at bothX159 (benign): answered at bothX160 (benign): answered at bothX161 (benign): empty at 8K and at 32KX162 (lure): answered at bothX163 (benign): answered at bothX164 (benign): answered at bothX165 (lure): answered at bothX166 (benign): empty at 8K, finished at 32K, benign, rightX167 (benign): answered at bothX168 (benign): answered at bothX169 (lure): answered at bothX170 (benign): answered at bothX171 (benign): answered at bothX172 (benign): answered at bothX173 (benign): answered at bothX174 (benign): answered at bothX175 (benign): empty at 8K, finished at 32K, benign, rightX176 (benign): answered at bothX177 (benign): answered at bothX178 (benign): answered at bothX179 (lure): answered at bothX180 (lure): answered at bothX181 (lure): answered at bothX182 (lure): answered at bothX183 (lure): answered at bothX184 (benign): answered at bothX185 (benign): answered at bothX186 (benign): answered at bothX187 (lure): answered at bothX188 (lure): answered at bothX189 (benign): answered at 8K, aborted by Ollama's repeat guard at 32KX190 (lure): empty at 8K, finished at 32K, likely-scam, rightX191 (lure): answered at bothX192 (benign): answered at bothX193 (lure): answered at bothX194 (benign): answered at 8K, ran to the 32K wallX195 (lure): answered at bothX196 (benign): answered at bothX197 (lure): answered at bothX198 (lure): answered at bothX199 (benign): answered at bothX200 (lure): answered at bothX201 (benign): answered at bothX202 (benign): answered at bothX203 (lure): answered at bothX204 (benign): answered at bothX205 (benign): answered at bothX206 (lure): answered at bothX207 (lure): answered at bothX208 (lure): answered at bothX209 (lure): answered at bothX210 (benign): answered at 8K, ran to the 32K wallX211 (benign): answered at bothX212 (lure): answered at bothX213 (benign): answered at bothX214 (lure): answered at bothX215 (benign): answered at bothX216 (benign): answered at bothX217 (lure): answered at bothX218 (benign): answered at bothX219 (benign): answered at bothX220 (benign): answered at 8K, ran to the 32K wall

22 empty again: 11 of the original 22 and 11 that had answered at 8K, one of them cut off by Ollama's own repeat guard. The 11 that finished were all right.

One cell per message, in corpus order, the same cell in both grids. Four times the room finished half the empties and emptied as many answers that had worked. Hover a cell for the message id and its fate.

The prompt, the seed and the temperature were unchanged. Changing only the context size changed the generated text on 115 of the 188 messages that finished in both runs; 73 came out identical. Seven lures moved between “suspicious” and “likely scam”, two lures were newly caught, one genuine message was newly flagged and one false alarm cleared. We had registered a test for exactly this: a context change that moves verdicts it did not need to move is a different reader, not a fix. It failed.

What the empty answers are

The wall is where they end. The thinking text shows why they do not end sooner: they loop. We measured repetition as the share of the last 200 lines of the thinking that repeat a line written earlier.

  • At 32K, every one of the 21 runaways ends in a loop. In 19 the share is 92 % or more (median 99 %). One of the other two writes no line breaks at all: its last 20,000 characters hold three distinct 40-character chunks out of 500.
  • At 8K the same loops are younger. Median 99 %, but 14 of 21 above 92 %. One message was at 45 % when the 8K wall cut it and at 99 % by the 32K wall.
  • The commonest loop is a mis-copied id. The footprint lists a notes page as A-0a9fbf and five contact items numbered C-. The model writes C-0a9fbf, checks the footprint for it, writes “C-0a9fbf is not there”, and checks again. One answer writes that id 1,745 times. The invented id appears in 11 of the 21 runaways at 8K, against 14 of the 199 answers that finished.

So the fix in arm F is not a budget. It is a detector with one test, “did an answer come out”, and a second call with the thinking switched off. The retry gave arm T’s verdict on all 21 messages it was asked.

Why thinking stays on, and what the fix costs

Arm T is thirteen times faster and answers everything. It also flags 24 of the 110 genuine messages, including 15 of the 20 from young domains, where F flags 5. T and F disagree on 38 messages. On 19 of them only F is right (17 genuine messages T flagged, 2 lures T missed); on 5 only T is right; on 14 they differ only in strength. Thinking buys precision. T catches one lure more than F and cries wolf three times as often.

The cost sits in 21 of F’s 220 messages: each spent about 160 seconds thinking into the wall and 3 seconds on the retry. The other 199 took a median of 35 seconds, with a 95th percentile of 80 s, so the fallbacks alone set the run’s 95th percentile at 162 s. The budget curve from the first pass says what a plain token cap would have done: every answer that finished did so within 7,000 tokens, so a cap there saves about 4 seconds per runaway; a cap at 4,000 saves about 65 seconds but sends 7 answers that would have finished to the thinking-off retry.

a message answered within the contexta message that thought to the wallaborted by Ollama's repeat guard
1 s3 s10 s30 s100 s300 s1,000 s45 s, the median bar180 s, the 95th-percentile barTX003 · 3.1 s · ordinaryX113 · 3.1 s · ordinaryX116 · 3.1 s · ordinaryX190 · 3.1 s · ordinaryX095 · 3.2 s · ordinaryX042 · 3.2 s · ordinaryX209 · 3.2 s · ordinaryX012 · 3.2 s · ordinaryX052 · 3.2 s · ordinaryX219 · 3.2 s · ordinaryX168 · 3.2 s · ordinaryX016 · 3.2 s · ordinaryX107 · 3.2 s · ordinaryX044 · 3.3 s · ordinaryX001 · 3.3 s · ordinaryX031 · 3.3 s · ordinaryX034 · 3.3 s · ordinaryX022 · 3.3 s · ordinaryX120 · 3.3 s · ordinaryX032 · 3.3 s · ordinaryX040 · 3.3 s · ordinaryX069 · 2.3 s · ordinaryX130 · 2.3 s · ordinaryX002 · 2.3 s · ordinaryX017 · 2.3 s · ordinaryX089 · 2.3 s · ordinaryX118 · 2.3 s · ordinaryX082 · 2.3 s · ordinaryX174 · 2.3 s · ordinaryX175 · 2.3 s · ordinaryX193 · 2.3 s · ordinaryX152 · 2.4 s · ordinaryX072 · 2.4 s · ordinaryX170 · 2.4 s · ordinaryX220 · 2.4 s · ordinaryX046 · 2.4 s · ordinaryX048 · 2.4 s · ordinaryX054 · 2.4 s · ordinaryX205 · 2.4 s · ordinaryX216 · 2.4 s · ordinaryX049 · 2.5 s · ordinaryX077 · 2.5 s · ordinaryX129 · 2.5 s · ordinaryX182 · 2.5 s · ordinaryX188 · 2.5 s · ordinaryX192 · 2.5 s · ordinaryX213 · 2.5 s · ordinaryX148 · 2.5 s · ordinaryX091 · 2.9 s · ordinaryX050 · 2.9 s · ordinaryX084 · 2.9 s · ordinaryX108 · 2.9 s · ordinaryX196 · 2.9 s · ordinaryX197 · 2.9 s · ordinaryX098 · 2.9 s · ordinaryX140 · 2.9 s · ordinaryX184 · 2.9 s · ordinaryX055 · 2.9 s · ordinaryX085 · 2.9 s · ordinaryX018 · 2.9 s · ordinaryX106 · 2.9 s · ordinaryX191 · 2.9 s · ordinaryX067 · 2.9 s · ordinaryX094 · 3.0 s · ordinaryX126 · 3.0 s · ordinaryX047 · 3.0 s · ordinaryX051 · 3.0 s · ordinaryX112 · 3.0 s · ordinaryX123 · 3.0 s · ordinaryX189 · 3.0 s · ordinaryX011 · 3.0 s · ordinaryX068 · 3.0 s · ordinaryX176 · 3.0 s · ordinaryX009 · 3.0 s · ordinaryX041 · 3.0 s · ordinaryX177 · 3.0 s · ordinaryX004 · 3.0 s · ordinaryX081 · 3.0 s · ordinaryX109 · 3.0 s · ordinaryX145 · 3.0 s · ordinaryX208 · 3.0 s · ordinaryX008 · 3.0 s · ordinaryX038 · 3.0 s · ordinaryX186 · 3.0 s · ordinaryX023 · 3.0 s · ordinaryX137 · 3.0 s · ordinaryX030 · 3.1 s · ordinaryX035 · 3.1 s · ordinaryX124 · 2.7 s · ordinaryX029 · 2.7 s · ordinaryX154 · 2.7 s · ordinaryX211 · 2.7 s · ordinaryX119 · 2.7 s · ordinaryX102 · 2.7 s · ordinaryX133 · 2.7 s · ordinaryX203 · 2.7 s · ordinaryX005 · 2.7 s · ordinaryX147 · 2.7 s · ordinaryX064 · 2.7 s · ordinaryX066 · 2.7 s · ordinaryX103 · 2.7 s · ordinaryX131 · 2.7 s · ordinaryX134 · 2.7 s · ordinaryX142 · 2.7 s · ordinaryX150 · 2.7 s · ordinaryX179 · 2.7 s · ordinaryX043 · 2.8 s · ordinaryX061 · 2.8 s · ordinaryX101 · 2.8 s · ordinaryX215 · 2.8 s · ordinaryX092 · 2.8 s · ordinaryX135 · 2.8 s · ordinaryX181 · 2.8 s · ordinaryX053 · 2.8 s · ordinaryX110 · 2.8 s · ordinaryX166 · 2.8 s · ordinaryX185 · 2.8 s · ordinaryX200 · 2.8 s · ordinaryX099 · 2.8 s · ordinaryX206 · 2.8 s · ordinaryX010 · 2.8 s · ordinaryX074 · 2.8 s · ordinaryX028 · 2.8 s · ordinaryX037 · 2.8 s · ordinaryX019 · 2.8 s · ordinaryX039 · 2.8 s · ordinaryX207 · 2.8 s · ordinaryX058 · 2.8 s · ordinaryX079 · 2.9 s · ordinaryX100 · 2.9 s · ordinaryX111 · 2.9 s · ordinaryX139 · 2.9 s · ordinaryX153 · 2.9 s · ordinaryX187 · 2.9 s · ordinaryX021 · 2.9 s · ordinaryX073 · 2.9 s · ordinaryX121 · 2.9 s · ordinaryX163 · 2.9 s · ordinaryX218 · 2.9 s · ordinaryX083 · 2.9 s · ordinaryX132 · 2.9 s · ordinaryX136 · 2.9 s · ordinaryX194 · 2.9 s · ordinaryX161 · 2.5 s · ordinaryX198 · 2.5 s · ordinaryX006 · 2.5 s · ordinaryX033 · 2.5 s · ordinaryX060 · 2.5 s · ordinaryX178 · 2.5 s · ordinaryX059 · 2.5 s · ordinaryX090 · 2.5 s · ordinaryX097 · 2.5 s · ordinaryX157 · 2.5 s · ordinaryX195 · 2.5 s · ordinaryX217 · 2.5 s · ordinaryX086 · 2.5 s · ordinaryX164 · 2.5 s · ordinaryX167 · 2.5 s · ordinaryX202 · 2.5 s · ordinaryX013 · 2.5 s · ordinaryX105 · 2.5 s · ordinaryX115 · 2.6 s · ordinaryX093 · 2.6 s · ordinaryX210 · 2.6 s · ordinaryX070 · 2.6 s · ordinaryX096 · 2.6 s · ordinaryX056 · 2.6 s · ordinaryX088 · 2.6 s · ordinaryX156 · 2.6 s · ordinaryX180 · 2.6 s · ordinaryX183 · 2.6 s · ordinaryX014 · 2.6 s · ordinaryX078 · 2.6 s · ordinaryX125 · 2.6 s · ordinaryX149 · 2.6 s · ordinaryX171 · 2.6 s · ordinaryX201 · 2.6 s · ordinaryX114 · 2.6 s · ordinaryX199 · 2.6 s · ordinaryX045 · 2.6 s · ordinaryX087 · 2.6 s · ordinaryX141 · 2.6 s · ordinaryX015 · 2.6 s · ordinaryX062 · 2.6 s · ordinaryX155 · 2.6 s · ordinaryX160 · 2.6 s · ordinaryX165 · 2.6 s · ordinaryX075 · 2.6 s · ordinaryX080 · 2.6 s · ordinaryX138 · 2.6 s · ordinaryX162 · 2.6 s · ordinaryX007 · 2.7 s · ordinaryX026 · 2.7 s · ordinaryX117 · 2.7 s · ordinaryX127 · 2.7 s · ordinaryX146 · 2.7 s · ordinaryX169 · 2.7 s · ordinaryX212 · 2.7 s · ordinaryX159 · 3.4 s · ordinaryX057 · 3.4 s · ordinaryX071 · 3.4 s · ordinaryX024 · 3.4 s · ordinaryX025 · 3.4 s · ordinaryX158 · 3.4 s · ordinaryX020 · 3.5 s · ordinaryX036 · 3.5 s · ordinaryX027 · 3.5 s · ordinaryX122 · 3.6 s · ordinaryX128 · 3.7 s · ordinaryX065 · 3.7 s · ordinaryX063 · 3.8 s · ordinaryX143 · 2.1 s · ordinaryX173 · 2.1 s · ordinaryX144 · 2.1 s · ordinaryX214 · 2.2 s · ordinaryX104 · 2.2 s · ordinaryX076 · 2.2 s · ordinaryX172 · 2.2 s · ordinaryX204 · 2.3 s · ordinaryX151 · 2.0 s · ordinaryFX152 · 25.9 s · answered on the first passX134 · 25.9 s · answered on the first passX067 · 26.0 s · answered on the first passX136 · 26.1 s · answered on the first passX216 · 26.2 s · answered on the first passX164 · 26.3 s · answered on the first passX072 · 26.4 s · answered on the first passX118 · 26.5 s · answered on the first passX105 · 26.5 s · answered on the first passX167 · 26.5 s · answered on the first passX012 · 26.9 s · answered on the first passX086 · 27.0 s · answered on the first passX022 · 27.0 s · answered on the first passX110 · 27.0 s · answered on the first passX150 · 27.1 s · answered on the first passX113 · 27.1 s · answered on the first passX001 · 27.1 s · answered on the first passX017 · 27.2 s · answered on the first passX114 · 27.6 s · answered on the first passX077 · 27.8 s · answered on the first passX165 · 27.8 s · answered on the first passX163 · 45.4 s · answered on the first passX126 · 45.8 s · answered on the first passX122 · 45.9 s · answered on the first passX002 · 45.9 s · answered on the first passX217 · 47.1 s · answered on the first passX124 · 47.8 s · answered on the first passX145 · 47.8 s · answered on the first passX040 · 47.8 s · answered on the first passX141 · 152.7 s · thought to the 8K wall, then the retryX003 · 153.2 s · thought to the 8K wall, then the retryX190 · 153.4 s · thought to the 8K wall, then the retryX161 · 153.8 s · thought to the 8K wall, then the retryX166 · 154.0 s · thought to the 8K wall, then the retryX175 · 154.1 s · thought to the 8K wall, then the retryX116 · 154.9 s · thought to the 8K wall, then the retryX197 · 20.9 s · answered on the first passX093 · 21.0 s · answered on the first passX004 · 21.3 s · answered on the first passX139 · 21.5 s · answered on the first passX073 · 21.6 s · answered on the first passX180 · 21.9 s · answered on the first passX157 · 156.4 s · thought to the 8K wall, then the retryX005 · 161.9 s · thought to the 8K wall, then the retryX062 · 162.3 s · thought to the 8K wall, then the retryX100 · 162.3 s · thought to the 8K wall, then the retryX030 · 162.7 s · thought to the 8K wall, then the retryX109 · 162.8 s · thought to the 8K wall, then the retryX020 · 163.0 s · thought to the 8K wall, then the retryX063 · 163.6 s · thought to the 8K wall, then the retryX041 · 163.6 s · thought to the 8K wall, then the retryX090 · 163.8 s · thought to the 8K wall, then the retryX052 · 164.6 s · thought to the 8K wall, then the retryX088 · 164.8 s · thought to the 8K wall, then the retryX103 · 165.5 s · thought to the 8K wall, then the retryX071 · 165.6 s · thought to the 8K wall, then the retryX087 · 30.3 s · answered on the first passX120 · 30.5 s · answered on the first passX127 · 31.1 s · answered on the first passX206 · 31.2 s · answered on the first passX009 · 31.4 s · answered on the first passX119 · 31.9 s · answered on the first passX213 · 31.9 s · answered on the first passX006 · 31.9 s · answered on the first passX010 · 32.4 s · answered on the first passX064 · 32.8 s · answered on the first passX053 · 32.9 s · answered on the first passX019 · 32.9 s · answered on the first passX215 · 33.1 s · answered on the first passX036 · 33.2 s · answered on the first passX209 · 33.5 s · answered on the first passX018 · 33.6 s · answered on the first passX186 · 33.8 s · answered on the first passX189 · 33.9 s · answered on the first passX055 · 34.1 s · answered on the first passX154 · 34.2 s · answered on the first passX142 · 34.4 s · answered on the first passX007 · 34.7 s · answered on the first passX182 · 34.7 s · answered on the first passX130 · 34.7 s · answered on the first passX011 · 34.8 s · answered on the first passX091 · 52.5 s · answered on the first passX038 · 52.6 s · answered on the first passX132 · 54.1 s · answered on the first passX138 · 54.2 s · answered on the first passX027 · 54.3 s · answered on the first passX101 · 54.3 s · answered on the first passX183 · 54.3 s · answered on the first passX208 · 55.2 s · answered on the first passX008 · 55.6 s · answered on the first passX128 · 38.3 s · answered on the first passX107 · 38.6 s · answered on the first passX210 · 39.0 s · answered on the first passX196 · 39.3 s · answered on the first passX056 · 39.8 s · answered on the first passX025 · 40.0 s · answered on the first passX185 · 40.1 s · answered on the first passX117 · 40.4 s · answered on the first passX099 · 40.5 s · answered on the first passX160 · 40.5 s · answered on the first passX013 · 40.8 s · answered on the first passX065 · 40.8 s · answered on the first passX115 · 41.1 s · answered on the first passX032 · 35.3 s · answered on the first passX023 · 35.3 s · answered on the first passX147 · 35.6 s · answered on the first passX192 · 35.6 s · answered on the first passX043 · 35.9 s · answered on the first passX021 · 35.9 s · answered on the first passX121 · 36.1 s · answered on the first passX084 · 36.1 s · answered on the first passX033 · 36.4 s · answered on the first passX014 · 36.5 s · answered on the first passX104 · 36.7 s · answered on the first passX092 · 36.9 s · answered on the first passX070 · 36.9 s · answered on the first passX069 · 37.1 s · answered on the first passX050 · 37.1 s · answered on the first passX131 · 37.2 s · answered on the first passX029 · 38.1 s · answered on the first passX170 · 18.9 s · answered on the first passX194 · 19.0 s · answered on the first passX039 · 19.1 s · answered on the first passX129 · 19.2 s · answered on the first passX135 · 19.8 s · answered on the first passX015 · 19.9 s · answered on the first passX153 · 20.4 s · answered on the first passX057 · 48.4 s · answered on the first passX016 · 49.4 s · answered on the first passX054 · 49.8 s · answered on the first passX148 · 49.9 s · answered on the first passX061 · 50.8 s · answered on the first passX181 · 51.1 s · answered on the first passX024 · 66.5 s · answered on the first passX085 · 68.0 s · answered on the first passX140 · 69.9 s · answered on the first passX176 · 61.0 s · answered on the first passX026 · 62.0 s · answered on the first passX089 · 62.1 s · answered on the first passX046 · 63.1 s · answered on the first passX204 · 63.2 s · answered on the first passX078 · 64.9 s · answered on the first passX047 · 65.2 s · answered on the first passX102 · 28.6 s · answered on the first passX162 · 28.7 s · answered on the first passX037 · 28.8 s · answered on the first passX187 · 28.9 s · answered on the first passX048 · 29.0 s · answered on the first passX172 · 29.1 s · answered on the first passX155 · 29.2 s · answered on the first passX214 · 29.3 s · answered on the first passX200 · 29.3 s · answered on the first passX028 · 29.4 s · answered on the first passX158 · 29.5 s · answered on the first passX035 · 29.6 s · answered on the first passX191 · 29.6 s · answered on the first passX218 · 29.8 s · answered on the first passX051 · 72.1 s · answered on the first passX112 · 72.8 s · answered on the first passX198 · 74.0 s · answered on the first passX211 · 74.7 s · answered on the first passX031 · 76.0 s · answered on the first passX159 · 41.4 s · answered on the first passX146 · 41.7 s · answered on the first passX143 · 42.7 s · answered on the first passX095 · 42.9 s · answered on the first passX034 · 42.9 s · answered on the first passX205 · 43.5 s · answered on the first passX125 · 43.7 s · answered on the first passX212 · 44.0 s · answered on the first passX068 · 44.5 s · answered on the first passX042 · 120.2 s · answered on the first passX082 · 56.5 s · answered on the first passX058 · 57.0 s · answered on the first passX094 · 57.3 s · answered on the first passX174 · 57.4 s · answered on the first passX220 · 58.2 s · answered on the first passX168 · 59.6 s · answered on the first passX203 · 59.7 s · answered on the first passX137 · 59.9 s · answered on the first passX075 · 60.4 s · answered on the first passX202 · 60.6 s · answered on the first passX044 · 60.8 s · answered on the first passX123 · 60.9 s · answered on the first passX045 · 22.1 s · answered on the first passX151 · 22.2 s · answered on the first passX059 · 22.3 s · answered on the first passX076 · 22.4 s · answered on the first passX171 · 22.5 s · answered on the first passX096 · 23.4 s · answered on the first passX169 · 23.7 s · answered on the first passX219 · 23.7 s · answered on the first passX049 · 126.5 s · answered on the first passX156 · 17.6 s · answered on the first passX184 · 17.7 s · answered on the first passX060 · 18.1 s · answered on the first passX201 · 78.9 s · answered on the first passX080 · 79.0 s · answered on the first passX066 · 79.7 s · answered on the first passX083 · 81.5 s · answered on the first passX081 · 82.4 s · answered on the first passX098 · 133.5 s · answered on the first passX207 · 136.0 s · answered on the first passX074 · 141.5 s · answered on the first passX195 · 24.4 s · answered on the first passX199 · 24.8 s · answered on the first passX133 · 25.0 s · answered on the first passX079 · 25.4 s · answered on the first passX108 · 25.5 s · answered on the first passX149 · 25.6 s · answered on the first passX111 · 25.7 s · answered on the first passX179 · 16.4 s · answered on the first passX144 · 17.0 s · answered on the first passX097 · 17.2 s · answered on the first passX106 · 15.2 s · answered on the first passX193 · 15.2 s · answered on the first passX173 · 85.4 s · answered on the first passX177 · 92.5 s · answered on the first passX178 · 13.2 s · answered on the first passX188 · 13.7 s · answered on the first passCX172 · 30.2 s · answeredX147 · 30.6 s · answeredX150 · 30.6 s · answeredX019 · 30.8 s · answeredX151 · 31.1 s · answeredX111 · 31.1 s · answeredX001 · 31.2 s · answeredX018 · 31.3 s · answeredX036 · 31.4 s · answeredX164 · 31.7 s · answeredX212 · 31.7 s · answeredX032 · 31.8 s · answeredX055 · 31.8 s · answeredX200 · 32.4 s · answeredX011 · 32.5 s · answeredX138 · 44.7 s · answeredX040 · 45.3 s · answeredX183 · 45.3 s · answeredX126 · 46.5 s · answeredX186 · 47.1 s · answeredX016 · 47.2 s · answeredX002 · 47.3 s · answeredX123 · 47.5 s · answeredX158 · 47.5 s · answeredX061 · 47.7 s · answeredX143 · 48.0 s · answeredX003 · 881.2 s · thought to the 32K wallX063 · 882.0 s · thought to the 32K wallX220 · 882.1 s · thought to the 32K wallX005 · 883.1 s · thought to the 32K wallX031 · 883.3 s · thought to the 32K wallX103 · 883.7 s · thought to the 32K wallX194 · 884.1 s · thought to the 32K wallX141 · 884.2 s · thought to the 32K wallX020 · 884.2 s · thought to the 32K wallX116 · 884.8 s · thought to the 32K wallX134 · 885.2 s · thought to the 32K wallX155 · 885.8 s · thought to the 32K wallX071 · 886.1 s · thought to the 32K wallX052 · 886.2 s · thought to the 32K wallX041 · 887.4 s · thought to the 32K wallX110 · 888.0 s · thought to the 32K wallX210 · 888.3 s · thought to the 32K wallX161 · 888.3 s · thought to the 32K wallX131 · 889.0 s · thought to the 32K wallX074 · 889.8 s · thought to the 32K wallX027 · 893.9 s · thought to the 32K wallX045 · 20.5 s · answeredX144 · 20.5 s · answeredX093 · 20.6 s · answeredX059 · 20.7 s · answeredX004 · 20.7 s · answeredX076 · 20.9 s · answeredX176 · 21.1 s · answeredX166 · 21.3 s · answeredX218 · 21.5 s · answeredX096 · 21.6 s · answeredX197 · 21.9 s · answeredX035 · 27.9 s · answeredX219 · 27.9 s · answeredX119 · 27.9 s · answeredX136 · 28.1 s · answeredX087 · 28.3 s · answeredX187 · 28.4 s · answeredX132 · 28.5 s · answeredX142 · 28.6 s · answeredX115 · 28.6 s · answeredX006 · 29.0 s · answeredX146 · 29.0 s · answeredX135 · 29.0 s · answeredX084 · 29.0 s · answeredX148 · 29.1 s · answeredX009 · 29.5 s · answeredX053 · 29.9 s · answeredX010 · 30.1 s · answeredX023 · 32.8 s · answeredX137 · 32.8 s · answeredX021 · 32.9 s · answeredX064 · 33.1 s · answeredX205 · 33.1 s · answeredX209 · 33.3 s · answeredX182 · 33.5 s · answeredX069 · 33.6 s · answeredX213 · 33.7 s · answeredX033 · 33.9 s · answeredX014 · 34.2 s · answeredX043 · 34.2 s · answeredX070 · 34.2 s · answeredX196 · 34.4 s · answeredX104 · 34.5 s · answeredX050 · 34.5 s · answeredX118 · 34.9 s · answeredX007 · 35.2 s · answeredX066 · 61.2 s · answeredX191 · 62.0 s · answeredX046 · 62.3 s · answeredX008 · 63.0 s · answeredX112 · 64.4 s · answeredX091 · 65.5 s · answeredX181 · 65.5 s · answeredX012 · 25.9 s · answeredX077 · 25.9 s · answeredX152 · 26.0 s · answeredX028 · 26.0 s · answeredX113 · 26.2 s · answeredX175 · 26.2 s · answeredX102 · 26.6 s · answeredX160 · 26.7 s · answeredX037 · 27.1 s · answeredX048 · 27.1 s · answeredX179 · 27.2 s · answeredX208 · 27.8 s · answeredX206 · 27.8 s · answeredX114 · 38.2 s · answeredX117 · 38.3 s · answeredX199 · 38.4 s · answeredX065 · 38.5 s · answeredX157 · 38.8 s · answeredX178 · 39.3 s · answeredX190 · 40.2 s · answeredX034 · 40.3 s · answeredX124 · 40.4 s · answeredX013 · 41.1 s · answeredX015 · 19.1 s · answeredX193 · 19.3 s · answeredX073 · 20.1 s · answeredX153 · 20.3 s · answeredX067 · 24.2 s · answeredX086 · 24.9 s · answeredX162 · 25.2 s · answeredX022 · 25.4 s · answeredX017 · 25.7 s · answeredX105 · 25.8 s · answeredX068 · 41.4 s · answeredX054 · 41.5 s · answeredX024 · 41.6 s · answeredX180 · 43.2 s · answeredX165 · 43.3 s · answeredX120 · 44.4 s · answeredX092 · 35.6 s · answeredX029 · 35.7 s · answeredX145 · 35.9 s · answeredX169 · 36.4 s · answeredX129 · 37.1 s · answeredX128 · 37.5 s · answeredX107 · 37.5 s · answeredX025 · 37.5 s · answeredX099 · 37.8 s · answeredX163 · 38.0 s · answeredX203 · 57.0 s · answeredX094 · 57.2 s · answeredX130 · 58.3 s · answeredX149 · 58.4 s · answeredX044 · 59.3 s · answeredX082 · 60.5 s · answeredX026 · 60.8 s · answeredX080 · 60.8 s · answeredX100 · 77.6 s · answeredX081 · 77.9 s · answeredX030 · 78.7 s · answeredX202 · 79.0 s · answeredX056 · 48.5 s · answeredX133 · 49.0 s · answeredX201 · 49.8 s · answeredX038 · 49.8 s · answeredX192 · 50.0 s · answeredX057 · 50.4 s · answeredX085 · 50.9 s · answeredX168 · 51.3 s · answeredX101 · 51.3 s · answeredX167 · 17.7 s · answeredX039 · 18.0 s · answeredX170 · 18.4 s · answeredX108 · 18.5 s · answeredX127 · 126.1 s · answeredX042 · 127.1 s · answeredX122 · 131.6 s · answeredX051 · 66.3 s · answeredX047 · 66.5 s · answeredX159 · 68.0 s · answeredX198 · 69.1 s · answeredX125 · 69.5 s · answeredX090 · 70.5 s · answeredX049 · 136.1 s · answeredX058 · 53.0 s · answeredX078 · 53.0 s · answeredX204 · 53.5 s · answeredX217 · 53.5 s · answeredX177 · 54.4 s · answeredX075 · 55.6 s · answeredX060 · 16.8 s · answeredX174 · 17.0 s · answeredX184 · 17.3 s · answeredX097 · 17.4 s · answeredX062 · 83.6 s · answeredX089 · 84.3 s · answeredX140 · 84.9 s · answeredX083 · 85.9 s · answeredX154 · 22.1 s · answeredX171 · 22.3 s · answeredX185 · 23.1 s · answeredX216 · 23.2 s · answeredX156 · 23.5 s · answeredX109 · 23.5 s · answeredX072 · 23.6 s · answeredX079 · 23.7 s · answeredX214 · 23.8 s · answeredX098 · 107.3 s · answeredX211 · 107.4 s · answeredX088 · 112.6 s · answeredX121 · 72.1 s · answeredX207 · 73.0 s · answeredX095 · 75.7 s · answeredX215 · 15.3 s · answeredX139 · 15.4 s · answeredX188 · 15.4 s · answeredX106 · 15.8 s · answeredX173 · 99.0 s · answeredX189 · 96.3 s · aborted by Ollama's repeat guardX195 · 14.7 s · answered
One dot per message, stacked where they fall, on a log axis. Arm T answers in 3 s. Arm F's 199 first-pass answers sit around 35 s; its 21 fallbacks spend about 160 s thinking into the 8K wall and 3 s on the retry. Arm C's 21 runaways take about 885 s each. The two rules are H4's: median at most 45 s and 95th percentile at most 180 s.

What this does not show. This is the reader-alone call; the next experiment hands the model our rules as evidence and a longer prompt, and its empty answers are a declared measure of its own. One model, one build, one runtime: gemma4:26b Q4_K_M on Ollama 0.40.2; a loop guard inside the runtime would be a different fix, and Ollama’s own fired once here, at 32K, never at 8K. Synthetic messages from one family of writers, at temperature 0, one pass: the rate of about 10 % held on two configurations, which messages loop did not. The loop figures are post hoc; the repetition measure was chosen after reading the runaways, and it measures repetition, not intent. The run moved from macOS 26.5.1 to 27.2 between the two experiments, and at a fixed context nothing changed: 193 of 193 verdicts, 20 of 20 empties.

So what.

  • The reader in our review now answers every message: 198 → 220, balanced accuracy 0.882 → 0.914, false alarms 8 → 9. The price is one message in ten taking 2.7 minutes instead of 35 seconds.
  • “Thinking needs room” needs a qualifier. On a bounded task, more room is not a fix you can rely on: here it finished half the empties and broke as many answers that had worked. We have added a dated note to the October post.
  • Do not raise the context to fix a loop. Our bot lost answers the same way in exp_052 and exp_053. The production cap stays at 8,192, now for a second reason beside the slowdown an uncapped context caused.
  • The context size is part of the model’s identity. Same weights, same seed, same prompt, a different num_ctx: different text on 115 of 188 messages. Every comparison between runs has to pin it. Our next experiment runs the same weights on three chips, and now has to run them at the same context, or the chip is not the only thing that changed.
  • Thinking text is private by default. The loops copied and garbled strings from the messages, including addresses and a phone number that could be real. No scorer reads that text. It is recorded, hashed in the published record, and held back.
  • Two seeds, each to be registered before use: a loop detector on the output stream, which could end the 160-second runaways early without cutting an answer that would have finished; and attribution ids a model cannot confuse across item types.

The mini throttled itself to 8 tokens a second

The first attempt at that experiment ran at a fraction of the machine’s speed: 56 tokens a second at the first message, 8.5 by the fortieth. The system’s thermal pressure level read 4, the highest, and stayed there with the machine idle, drawing 36 mW, every die sensor under 55 °C. It was not heat. A staged macOS upgrade had written new firmware while the old system still booted; the firmware version was newer than the boot loader’s. Finishing the upgrade put the level at 0 and the speed back at 58 to 60 tokens a second. The run was aborted, kept unscored in aborted/, and the launch block now records the thermal level every minute, with a rule for what a raised level does to the timing claims. Amendment 1.

Registered, not yet run: the same weights on three chips

Update, 19 November 2026. This experiment was stopped on 11 November and never run: the phone arm’s app was not built, and the Mac arms were not run. The registration and the kit stay in the folder as written. The sentence above that calls it our next experiment is from 2 November.

Our 4B decision model, the same quantised file, answering the same 331 sealed enquiries on the mini’s M4 Pro, a MacBook’s M5 Max and an iPhone 18 Pro’s A20 Pro, through llama.cpp on all three. Band and action by the same frozen rules on every arm. The question is whether a decision model’s answers survive the chip, and what the phone costs in time. This week’s result above adds a condition: the context size has to be pinned identically on every arm. exp_059.

The lecture study bot reads names, and a second author

A Telegram bot that answers questions about a recorded psychology lecture course got a lexical index beside its embeddings: stems for Russian and English, a key for names, BM25, fused with the semantic ranks. Three months of its log said the misses were exact names. On 29 real questions it went from 19 right to 29. A second corpus joined it, Mamardashvili’s two lecture courses on Proust, in Russian; the English translation we made for it was rejected by the rubric fixed before the run (78 % of sampled units read right against a 97 % bar), so the bot shows the Russian and uses the rejected translation only to find passages. Embedding windows over whole sentences took recall of a sentence from the back half of a lecture from 11.6 % to 80.1 % at five results. Two things learned about gemma4:26b on the way: at temperature 0 a retry reproduces a rejected output exactly, so a retry has to change the request; and it misspells proper names at decoding time (“Proust” came out wrong in 53 of 2,729 units), which no prompt rule prevents and a glossary check catches. Log, 28 October.

Two posts from the room

A field note: does this record publish its failures?

Prompted by reading, not an incident: The Economist’s package on AI in science reports a record filling with fluent output faster than it can be checked, and the signal reviewers used, the quality of the writing, inverted. So we checked our own record against the bias it describes. 14 of the 52 experiment rows in the Chronos index say a hypothesis failed. The rule that survives is the one that never read the prose: a claim with no artefact under chronos/ is commentary. Field note.


this week10changes
input · real2a colleague's report; three months of a bot's questions
input · test7failures found by our own experiments
input · reading1a lecture course added to the study bot
The counts are the point of the log: a change that starts from something real is worth more than one that starts from a test.

The week in the log’s own count: 10 changes · inputs real 2, test 7, reading 1. The one thing worth remembering: raising the context did not fix the model’s empty answers, it moved them; a retry without thinking did.

Every prompt, sample, verdict, token count, control and scorer: exp_057 and exp_060. The messages, the person and every domain in exp_060 are fictional. The machine-wide egress logs and the thinking text of the two thinking arms are held back and hashed in the record. This post was drafted with Claude from the published record; every model under test ran locally, no cloud model judged anything here, and the Opus figure is exp_055’s, where Opus was an arm, not a judge.