Turn 8 · claude

# wsp31 author brief (read fully at the start of every turn)

You are one of three authors (claude, gpt, gemini) taking turns. Your job: make Gemma 4 31B solve more coding-agent
tasks **by editing its weights, with no training**. You read the model (weights, activations, traces, failures),
decide what to change, build the change as a closed-form edit, predict its effect, and queue it. A separate runner
spends the rollouts and writes the measured result next to your prediction. You are scored on tasks gained over the
champion in the same batch (the unedited model plus the edits that have already passed; see "The champion"), on the
margin over your matched control, and on how well your predictions held.

**Be bold.** Small safe nudges were tried on the 4B model and were lost in the noise (see `PRIOR_E4B.md`). You get
two entries per round and then wait for their results, so each one counts; a third variation of a null result is a
wasted round. High-risk ideas with a clear mechanism and a
prediction that could be wrong are what this programme wants. Say what would prove you wrong.

## Where you work

- Everything happens on host `wrx90`, reached with `ssh -o BatchMode=yes wrx90 '<command>'`. Write scripts in your
  local working directory, copy them with `scp`, run them with ssh. Keep ssh output small (`tail`, `cut`, `head`).
- Root: `/home/keith/gemma-v2` (below: `$R`). Python: `$R/tii-venv/bin/python` (torch, transformers, vLLM 0.19.1,
  peft, safetensors; no scipy). `. $R/env.sh` sets `R` and `PY`.
[... 206 lines cut ...]
- An edit that passes both stages is promoted: it becomes part of the champion for everything after it.
- The champion is confirmed on tasks you do not see, against the unedited model, and the final champion gets one
  pre-registered held-out run. A changed edit is a new edit.
- A combination of your own edits is a new queue entry with its own record, prediction and control. Combining with
  edits that already passed needs nothing from you: that is what the champion does.

Single-run agent evaluations are noisy by several tasks. An edit worth having shows up as ten or more tasks.
Aim there.

# Results for author claude (2026-11-14 00:44:10 EST)

Written by the runner as your turn opens. "New" means scored since your last turn started (2026-11-13 18:50:57 EST). The full result text of every entry is in its ledger record (`tii-wsp31-doc/research/wsp31/ledger/<id>-*.md`, index `LEDGER.md` next to that directory); the live queue is `wsp31/queue/STATUS.md`.

## Your new results

### 0031 `icv-brief-b20`: Instruction vector, think briefly: an always-on neuron per layer writes the stream shift of a 'keep your thinking short' paragraph (dose 2)

State queued_s2, verdict -; predicted stage-1 gain -6 to +6; record `tii-wsp31-doc/research/wsp31/ledger/0031-claude-icv-brief-b20.md`.

**Stage 1**, batch `x001` (finished 2026-11-14 00:41:07 EST), reference arm: unedited. Both reference arms are the unedited model served through an all-zero adapter, the same serving path as an edit (ledger record 0033). Box batch: run on rented boxes as 4 shards by task, every arm of a shard on one server; its edits are compared only with its own reference arms (run by hand, adopted).

- Gain used by the rule: +9.0 tasks of 120 (95% interval +2.0 to +16.0), against the mean of the two reference arms.
- Your prediction was -6 to +6: the measured gain is outside it. Stage-1 bar +3; harmful at -6 or below.
[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
- The two reference arms (identical weights) solved 72 and 64 and differ on 18 tasks; record 0003 measured 11.3 per identical pair.
- Solved by your edit and by no reference arm (6): `[task]`, `mozilla__bleach.73871d76.func_pm_remove_cond__dfyej8pm`, `[task]`, `r1chardj0n3s__parse.30da9e4f.func_basic__83wal9im`, `[task]`, `[task]`.
- Solved by every reference arm and not by your edit (3): `[task]`, `django-money__django-money.835c1ab8.combine_module__bqggrnha`, `[task]`.
- Traces and request logs: `results/wsp31/batches/x001/e0031/` (your edit), `.../x001/base/` and `.../x001/aa/` (reference arms).
[... 48 lines cut ...]
| 0017 | gemini | late-preedit-commit-gate | Coordinate-whitened GeGLU late pre-edit commit gate across 13 high-signal layers (dose 1.20) | parked | -3 to +10 |  |  | - |
| 0018 | gemini | dual-budget-rescue-r5 | Dual-cap rescue stack: rank-4 exact-close writer subspace (0.38%) plus 13-layer late pre-edit commit gate | parked | -4 to +14 |  |  | - |
| 0025 new | gpt | code-writer-damp-r4-033 | Damp the code-versus-prose style contrast in writers in all layers at 0.33 percent | done | -6 to +12 | -1.0 (68 vs 65) |  | null |
| 0026 new | gpt | code-writer-damp-r4-011 | Damp the code-versus-prose style contrast in writers in all layers at 0.11 percent | done | -5 to +10 | +0.0 (69 vs 65) |  | null |
| 0027 | gpt | code-reader-damp-r4-033 | Damp the code-versus-prose style contrast in readers in all layers at 0.33 percent | parked | -8 to +14 |  |  | - |
| 0028 | gemini | nullspace-close-r4-midlate | Protocol-nullspace rank-4 exact thought-close writer subspace across L10-59 (0.48%) | running_s1 | +1 to +13 |  |  | - |
| 0035 | gpt | source-grounding-icv-b3 | Source-grounding instruction contrast across all 60 layers at dose 3 | queued_s1 | -8 to +12 |  |  | - |
| 0036 | gpt | source-grounding-icv-b1 | Source-grounding instruction contrast across all 60 layers at dose 1 | queued_s1 | -8 to +10 |  |  | - |

# Notes of author claude (rewritten end of turn 6, 13 Nov 2026, 7:55 PM ET)

## Scored so far: all three harmful
- 0008, 0020, 0021 (gate neuron per layer: depth >= 800 -> write the end-of-thought state): -10, -10.5, -7.5.
[line removed by the dashboard scrubber]
  empty patches double. Family dead. Parked 0007, 0009, 0010, 0022, 0023 are the same family: do not resubmit.
[line removed by the dashboard scrubber]

## Queued this turn: 0031 `icv-brief-b20`, 0032 `icv-brief-b40` (instruction vector, dose pair)
An always-on neuron in each of the 60 layers (third-quietest neuron replaced; gate 8w, up w, w = ridge fit of the
constant 1 on the MLP input) writes beta x the per-layer mean stream shift caused by a paragraph "keep your thinking
very short ..." inserted at the end of the first user message (delta_L = d_L - layer_scalar_L d_(L-1); 35 fit
calls of b001). No gate, no close push. 0.67% of the network over 180 matrices (mostly the removed neurons).
Control: sign-scrambled write.
Pre-test: dose 2 moves the held-out logged path as much as the paragraph (KL 0.0072 vs 0.0076; 10 of 15 prompted
flips matched, control 5); dose 1 is closest in KL (0.50 reproduced); dose 4 is an over-dose (KL 0.028).
Served single turns: the paragraph shortens thoughts in 7 of 17 mid-episode contexts, lengthens none; dose 4: 6 / 5,
[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
paragraph -0.19): flagged by the tools lane's rule.
[... 192 lines cut ...]
size_icv-brief-b20.json
size_icv-brief-b20.log
size_icv-brief-b40.json
size_icv-brief-b40.log
size_k.json
size_old.json
t3_rows.json
t3_seq.json

# wsp31 tools for edit authors

Changelog (newest first; times US Eastern):
- Fri 13 Nov, 3:51 PM: **three of the first four edits were harmful (-10 to -21 tasks): read "Before you submit"
  (section 5) first.** GPU 0 for pre-tests only while an author turn is open (section 3). New `delimprobe.py`.
[line removed by the dashboard scrubber]
  12:39 AM speed (trap 12): **the server's rank limit sets the price of an adapter**.
- Thu 12 Nov: 11:40 PM first served runs of `dose.py`, `validate.py`, `vgen.py`, `closeprobe.py`; 10:55 PM stacks
  of rank 34 to 96 applied as written; 10:30 PM **thought-end pre-tests need exact contexts**, rank limits 64 or 128
  (never 96); 10:00 PM champion support, network-wide builders; 8:36 PM first version. Earlier commands all still work.

Everything here lives in `$R/wsp31/tools` (below `$T`; `$R = /home/keith/gemma-v2`). Each entry ends with
**Verified by** (a measurement that was made) or **NOT VERIFIED YET** (written, first real run still queued; use it,
but do not build a claim on its numbers alone). Architecture facts are in `$T/ARCH.md` (read "Consequences for
weight edits" there before designing an edit); measurements with commands are in `$T/REPORT.md`.

    . /home/keith/gemma-v2/env.sh; T=$R/wsp31/tools
    export WSP31_WORK=$R/wsp31/authors/<you>/scratch     # caches and server logs of the tools go here, not into $T

Do not write into `$T`. Run the tools by absolute path from your own directory.
[... 286 lines cut ...]

## 8. Not available yet

[line removed by the dashboard scrubber]
- Activation atlas and its views; `--champion` runs of `dose.py`, `validate.py`, `closeprobe.py`, `vgen.py`.
- Any single entry beyond rank 32 per matrix, and edits of norms, `layer_scalar` or embeddings (`overlay.py` exists
  as code, has not been served, and the runner does not accept it).
- A probe verdict that predicts rollout outcome: the registered ranking test failed on four edits (section 5).

Sat Nov 14 12:45:05 AM EST 2026
    50 code/addnotes.py
    60 code/arm_stats.py
    29 code/build_think.py
    35 code/cf_budget.py
    69 code/end_neurons.py
    19 code/fail_stats.py
    12 code/fill_fb.py
    12 code/fillvgen.py
   101 code/finalize3.py
   175 code/finalize.py
    34 code/fix_records.py
    24 code/icv_build.py
   129 code/icv_gen.py
   490 code/icv_job.py
    31 code/inspect_probe.py
[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
    45 code/meta_check.py
[... 119 lines cut ...]
eval2.pt
eval2_arms.txt
gen2.json
summary.json
summary2.json
vgen2_arms.txt
vgen_turn.json
1.5G	scratch

title: Instruction vector, think briefly: an always-on neuron per layer writes the stream shift of a 'keep your thinking short' paragraph (dose 2)

## Hypothesis

[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]

[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
still got long into loops. This entry pushes nothing towards `<channel|>` and has no gate. It writes, at every
position and in every layer, the mean shift of the residual stream that an explicit instruction paragraph causes in
the unedited model itself:

> Important: keep your thinking very short. Before each tool call, think in at most three short sentences: what you just learned and what you will do next. Never write long plans, never draft or rewrite code in your thoughts, and never go over the same point twice.

[line removed by the dashboard scrubber]
that movement that is the same at every position is a constant vector per layer, and a constant vector is what an
always-on neuron can write (the model has no biases). If the shift carries the instruction, the model plans a short
[... 122 lines cut ...]
Stream shift (mean over completion positions of `resid_post`, arm minus base, against prompted minus base, same
held-out calls; layers [0, 6, 12, 18, 24, 30, 36, 42, 48, 54, 59]): this entry: cosine 0.25 0.68 0.51 0.46 0.46 0.48 0.54 0.45 0.46 0.42 0.37; norm ratio 3.98 2.04 2.37 2.18 1.87 2.18 1.88 2.27 2.12 1.94 2.48. Control: cosine -0.03 -0.04 0.05 -0.04 0.04 0.05 0.09 0.08 0.13 0.12 0.09; norm ratio 3.95 1.72 1.59 1.14 1.15 1.38 1.39 1.41 1.41 1.30 1.64.

Activation of the always-on neuron over gbar in the edited forward pass (dose 1, held-out calls, 11 layers):
completion positions mean 0.94 to 1.00, sd 0.06 to 0.12, min 0.54, max 1.57; prompt positions
mean 0.76 to 0.92; BOS 0.25 to 3.25.

### Served generation (vLLM through `tools/gpu0_session.sh`, greedy, single turns, no rollout)

aa
base
box_merge.json
e0031
e0032
score.json
rebench
run.json
    show("/".join(run.rstrip("/").split("/")[-2:]), arm(run))
    ~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/keith/gemma-v2/wsp31/authors/claude/code/arm_stats.py", line 52, in show
    name, n, sum(e["solved"] for e in eps), st.median(th), st.mean(th), sum(t >= 300 for t in th) / len(th),
                                            ~~~~~~~~~^^^^
  File "/home/keith/gemma-devagent-20261001/python/cpython-3.13.13-linux-x86_64-gnu/lib/python3.13/statistics.py", line 637, in median
    raise StatisticsError("no median for empty data")
statistics.StatisticsError: no median for empty data

shard0_gpu0
patches
task_results.jsonl
traces
[line removed by the dashboard scrubber]
 "name": "x001",
 "ids_file": "research/wsp31/remote_eval/proper_run/ids_stage1_all.txt",
 "tasks": 120,
 "common": {
  "set_name": "set_all",
  "arms": [
   [
    "base",
    "zero"
   ],
   [
    "e0031",
    "e0031"
   ],
   [
[... 8 lines cut ...]
  "adapters": {
   "e0031": {
    "sha_dir": "76a1ee20ff52a9746e80ea00938548822547deda793153e0082e07c0b492d079",
    "safetensors_sha256": "b1dffb49682ccda0c69493ea85d7ffd06ac8f0f25792bb20a4089f07b7fcf90d",
    "rank": 1
   },
   "e0032": {
    "sha_dir": "5e11e1a105f65b5112cd9d422c001fc1ef1c154eb5ed

base n 120 solved 72 | out tok med 8410 mean 7565 | at10k 45 | calls med 18 | capped calls mean 8.17 | empty 21 | timeout 45 (solved among them 13) | tok/call med 385
    {'Agent exceeded session timeout (7.5 min)': 45, 'none': 74, 'Agent exceeded turns budget (40 turns)': 1}
    {'empty': 21, 'solved': 72, 'patch_done_fail': 16, 'patch_timeout_fail': 11}
aa n 120 solved 64 | out tok med 7846 mean 7473 | at10k 46 | calls med 18 | capped calls mean 7.93 | empty 18 | timeout 46 (solved among them 12) | tok/call med 384
    {'Agent exceeded session timeout (7.5 min)': 46, 'none': 74}
    {'empty': 18, 'solved': 64, 'patch_done_fail': 19, 'patch_timeout_fail': 19}
e0031 n 120 solved 77 | out tok med 6996 mean 6791 | at10k 32 | calls med 18 | capped calls mean 7.59 | empty 13 | timeout 32 (solved among them 7) | tok/call med 356
    {'Agent exceeded session timeout (7.5 min)': 32, 'none': 88}
    {'patch_timeout_fail': 13, 'solved': 77, 'patch_done_fail': 17, 'empty': 13}
e0032 n 120 solved 71 | out tok med 5308 mean 5916 | at10k 20 | calls med 16 | capped calls mean 6.47 | empty 8 | timeout 19 (solved among them 1) | tok/call med 295
    {'none': 101, 'Agent exceeded session timeout (7.5 min)': 19}
    {'patch_done_fail': 31, 'solved': 71, 'patch_timeout_fail': 10, 'empty': 8}
{"agent_seconds": 344.9, "budget": 10000, "calls": 9, "capped_calls": 3, "no_limit_calls": 0, "uncounted_calls": 0, "wall_limit_hit": false}
e0031 gained over both:
    [task] [('base', 10000, 26, 0, 'Agent exceed'), ('aa', 10000, 21, 0, 'Agent exceed'), ('e0031', 10000, 17, 1105, 'Agent exceed')]
    mozilla__bleach.73871d76.func_pm_remove_cond__dfyej8pm [('base', 10000, 13, 963, 'Agent exceed'), ('aa', 10001, 15, 1880, 'Agent exceed'), ('e0031', 5438, 23, 803, '')]
    [task] [('base', 10001, 15, 1756, 'Agent exceed'), ('aa', 6793, 13, 1625, ''), ('e0031', 8683, 18, 1443, '')]
    r1chardj0n3s__parse.30da9e4f.func_basic__83wal9im [('base', 6772, 10, 976, ''), ('aa', 10000, 20, 0, 'Agent exceed'), ('e0031', 7750, 17, 1189, '')]
    [task] [('base', 6008, 30, 1410, ''), ('aa', 10000, 30, 1647, 'Agent exceed'), ('e0031', 3083, 16, 511, '')]
    [task] [('base', 10000, 10, 0, 'Agent exceed'), ('aa', 7191, 15, 982, ''), ('e0031', 5961, 14, 1182, '')]
[... 9 lines cut ...]
    [task] [('base', 10000, 10, 0, 'Agent exceed'), ('aa', 7191, 15, 982, ''), ('e0032', 5896, 9, 1159, '')]
e0032 lost vs both:
    [task] [('base', 8393, 21, 717, ''), ('aa', 2863, 14, 744, ''), ('e0032', 3163, 13, 1089, '')]
    [task] [('base', 4916, 19, 865, ''), ('aa', 6407, 15, 815, ''), ('e0032', 10000, 16, 0, 'Agent exceed')]
    [task] [('base', 8590, 24, 2270, ''), ('aa', 6586, 21, 2343, ''), ('e0032', 10000, 27, 583, 'Agent exceed')]
    pdfminer__pdfminer.six.1a8bd2f7.lm_rewrite__iv8vwrta [('base', 9124, 30, 1416, ''), ('aa', 10001, 24, 1623, 'Agent exceed'), ('e0032', 10000, 14, 309, 'Agent exceed')]
[line removed by the dashboard scrubber]
    [task] [('base', 10000, 22, 3143, 'Agent exceed'), ('aa', 10000, 23, 4838, 'Agent exceed'), ('e0032', 10000, 18, 1996, 'Agent exceed')]

"""icv_job.py (author claude, turn 6): one GPU-0 hold with tools/g31.py (forward pass only, no server).

Instruction vectors compiled into the weights as an always-on MLP neuron in every layer.
  1. capture: for stage-1 calls of the unedited b001 arms, run the logged call (prompt + the completion the agent
     wrote) once as logged and once with an instruction paragraph inserted at the end of the first user message.
     d_L[I] = mean over calls of mean over completion positions of resid_post_L(with I) - resid_post_L(without).
  2. per layer the part layer L adds itself: delta_L = d_L - layer_scalar_L * d_{L-1}  (d_{-1} = 0).
  3. in every layer one quiet MLP neuron is replaced ("set") by an always-on neuron: gate row K w, up row w, with
     w . x = 1 on average over completion positions (x = MLP input), so its activation is about K at every position;
     its down column writes beta * delta_L into the stream (through the post-feedforward norm and layer_scalar).
  4. teacher-forced evaluation on held-out tasks: base, prompted with I, and each arm, on the logged completion.
`--mock` runs the whole script on CPU with random activations (plumbing test).
"""
import json
import math
import os
import random
import sys
import time

[... 112 lines cut ...]
        rnorm.append(base.norm(dim=-1).mean(1).float())
        del caps
        for k, ins in itok.items():
            ids2 = with_instr(ids, e, ins)
            with torch.no_grad():
                r = m.forward(ids2, out_pos=[len(ids2) - 1], capture=("resid_post",), cap_layers=LAYERS,
                              cap_pos=[p + len(ins) for p in P], stop_after=59)
            diffs[k].append((r["caps"]["resid_post"].double() - base).mean(1).float())

"""icv_build.py OUT_DIR I1|I2 BETA edit|ctrl  (author claude, turn 6; CPU, seconds)

Rebuilds one arm of code/icv_job.py bit for bit from the statistics that job saved (code/data/icv_stats.pt: the
per-layer instruction shifts delta_L, the always-on neuron of every layer, the MLP output RMS, the norm weights and
layer scalars). edit: the down column of the always-on neuron of layer L writes BETA * delta_L[I]. ctrl: the same
code path with the sign of every coordinate of the written vector flipped at random (seed 6000 + L): identical
neurons, identical gate and up rows, identical per-coordinate magnitudes of the write, so identical |dW|_F.
"""
import sys

import torch

A_DIR = "/home/keith/gemma-v2/wsp31/authors/claude"
sys.path.insert(0, A_DIR + "/code")
import icv_job as J  # noqa: E402

[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
st = torch.load(A_DIR + "/code/data/icv_stats.pt")
[line removed by the dashboard scrubber]
[... 108 lines cut ...]
d = json.load(open("scratch/size_$SLUG.json"))
for k in ("mine", "ctrl"):
    s = d[k]["summary"]
    print(k, "size: whole network %.4f%% over %d matrices, largest matrix %.4f%% (%s), stable rank %.1f" % (100 * s["net_rel"], s["changed"], 100 * s["rel_max"], s["rel_max_mod"], s["stable_rank_median"]))
EOF
if [ -d scratch/job6/ad/$ARM ]; then
  sha256sum $Q/edit/adapter_model.safetensors scratch/job6/ad/$ARM/adapter_model.safetensors | cut -c1-110
fi

            diffs[k].append((r["caps"]["resid_post"].double() - base).mean(1).float())
            del r
        used.append((rec["instance"], rec["run"].split("/")[-1], rec["call"]))
        if not MOCK:
            torch.cuda.empty_cache()
        log("captured", ri, rec["instance"][:32], "call", rec["call"], "n", rec["n"], "pos", len(P))
    return dict(layers=LAYERS, used=used, diffs={k: torch.stack(v) for k, v in diffs.items()}, X=torch.cat(Xs, 1),
                kind=torch.cat(kinds), rec=torch.cat(recidx), rmsy=torch.stack(rmsy).mean(0),
                rnorm=torch.stack(rnorm).mean(0))


def cosf(a, b):
    return ((a * b).sum(-1) / (a.norm(dim=-1) * b.norm(dim=-1) + 1e-9))


def fit(m, st):
    nl = len(LAYERS)
    st["w_post"] = torch.stack([m.L[L]["post_feedforward_layernorm"].float().cpu() for L in LAYERS])
    st["ls"] = torch.tensor([float(m.L[L]["layer_scalar"].float()) for L in LAYERS])
    st["d"], st["delta"], st["dinfo"] = {}, {}, {}
[... 323 lines cut ...]
    json.dump(sm, open(S + "/summary.json", "w"), indent=1)
    for k, v in sm.items():
        log("SUMMARY", k, json.dumps(v))
    log("done")


if __name__ == "__main__":
    main()

[line removed by the dashboard scrubber]
2 out_tok 227 | think chars     0 | msg  670 | obs  6143 | read_file {'filepath': 'linebot/models/flex_message.py'}
3 out_tok 219 | think chars     0 | msg  612 | obs  4569 | read_file {'filepath': 'linebot/models/base.py'}
4 out_tok 406 | think chars     0 | msg 1304 | obs  6480 | read_file {'filepath': 'linebot/models/flex_message.py', 'start_line': 151}
5 out_tok 48 | think chars     0 | msg   39 | obs  5801 | read_file {'filepath': 'linebot/models/flex_message.py', 'start_line': 301}
6 out_tok 48 | think chars     0 | msg   40 | obs  5546 | read_file {'filepath': 'linebot/models/flex_message.py', 'start_line': 451}
7 out_tok 267 | think chars     0 | msg  478 | obs   303 | run_command {'command': 'python3 - <<\'PY\'\nfrom linebot.models.flex_message import BubbleContainer, ImageComponent, BoxComponent\n
8 out_tok 1430 | think chars     0 | msg 4590 | obs   223 | run_command {'command': 'python3 - <<\'PY\'\nfrom linebot.models.flex_message import BubbleContainer, ImageComponent, BoxComponent\n
9 out_tok 730 | think chars     0 | msg 2118 | obs   226 | run_command {'command': 'python3 - <<\'PY\'\nfrom linebot.models.flex_message import BubbleContainer, BoxComponent\nhero = BoxCompon
10 out_tok 715 | think chars     0 | msg 2103 | obs  1087 | run_command {'command': 'grep -rn "get_or_new_from_json_dict_with_types" linebot/models/'}
11 out_tok 2289 | think chars     0 | msg 7126 | obs  1054 | read_file {'filepath': 'linebot/models/base.py', 'start_line': 120, 'end_line': 140}
12 out_tok 156 | think chars     0 | msg  334 | obs   784 | read_file {'end_line': 35, 'filepath': 'linebot/models/actions.py', 'start_line': 20}
13 out_tok 691 | think chars     0 | msg 1870 | obs   316 | run_command {'command': 'grep -n "get_or_new_from_json_dict" linebot/models/base.py'}
14 out_tok 338 | think chars     0 | msg  855 | obs  1213 | read_file {'end_line': 320, 'filepath': 'linebot/models/flex_message.py', 'start_line': 300}
15 out_tok 89 | think chars     0 | msg  183 | obs   255 | run_command {'command': 'grep -n "class BubbleContainer" linebot/models/flex_message.py'}
16 out_tok 62 | think chars     0 | msg   53 | obs  2004 | read_file {'end_line': 110, 'filepath': 'linebot/models/flex_message.py', 'start_line': 75}
17 out_tok 55 | think chars     0 | msg   34 | obs  1204 | read_file {'end_line': 130, 'filepath': 'linebot/models/flex_message.py', 'start_line': 111}
18 out_tok 1677 | think chars     0 | msg 4777 | obs     4 | 
[line removed by the dashboard scrubber]

== base
  solved    n  72 | test after last edit 0.83 | any run after last edit 0.94 | edits med 1 | first edit med 9.0 | thought chars/ep med 13496 | arg chars med 4496 | out tok med 6378 | final-msg tok med 128
  done_fail n  16 | test after last edit 0.88 | any run after last edit 1.00 | edits med 1 | first edit med 9.0 | thought chars/ep med 14910 | arg chars med 5333 | out tok med 7529 | final-msg tok med 128
  to_fail   n  11 | test after last edit 0.36 | any run after last edit 0.55 | edits med 1 | first edit med 13 | thought chars/ep med 23770 | arg chars med 5924 | out tok med 10000 | final-msg tok med 62
  empty     n  21 | test after last edit 0.05 | any run after last edit 0.14 | edits med 0 | first edit med 13.5 | thought chars/ep med 24481 | arg chars med 2814 | out tok med 10000 | final-msg tok med 575
   done_fail ['rrrrrrrreTTTrrs--', 'rrrrrreTTTrs--', 'rrrrrrrrrrrrrrrrrrrrrrrrwrrrrrresr--', 'rrerrrrrrgrs--', 'rrrrTTrTrTeeTTrrrrs--', 'TrrrrrrrrrrrrrrTerrrrrTTrrrrs--', 'rrrrrrrTreeereTTTrs--', 'rrrrrrTreeeerTTrrrs--', 'rrrTrrrTrrrTeTTTTTrs--', 'rrrrrrrrerrrrTTTrrrrrrrrrreTTrrs--', 'rrrrrrreTTrs--', 'rrrrreTrrrrTTrrs--', 'rrrrerrTTrs--', 'rr-rrrrrrrrrrrrrrrrrrerrTTTTTTrs--']
   empty ['rrrrTrrr-', 'rrrrrrrrrrrrrr-', 'rrrrTrrrrrrr-', 'rTrTTr-', 'rrrrwrrrrrrrrrrrrrrrr-', 'rrrrTrrrrr-', 'rrrrrrrrrrrrrrrrrrrrrrrrrrr--', 'rrrrrrrrrrrrrrrrTrr-', 'rrrrrrTrrrrrrrrrrrrrrrr', 'rrrrrrrrrrrrrrrrww--', 'rrrrrrrrrwTrrrrrrrrrrrrrrr--', 'rrrrrTrTTrrrrrrrrrrrrrrerr-', 'rrrrrrTrrrrrrrr-', 'rrrrrrrrrTTTrrrrTrr-']
== aa
  solved    n  64 | test after last edit 0.83 | any run after last edit 0.98 | edits med 1 | first edit med 9.0 | thought chars/ep med 13458 | arg chars med 4588 | out tok med 6164 | final-msg tok med 130
  done_fail n  19 | test after last edit 0.79 | any run after last edit 0.95 | edits med 1 | first edit med 11 | thought chars/ep med 12341 | arg chars med 4606 | out tok med 6811 | final-msg tok med 93
  to_fail   n  19 | test after last edit 0.58 | any run after last edit 0.63 | edits med 2 | first edit med 12 | thought chars/ep med 23180 | arg chars med 6100 | out tok med 10000 | final-msg tok med 198
  empty     n  18 | test after last edit 0.11 | any run after last edit 0.11 | edits med 0 | first edit med 27 | thought chars/ep med 22334 | arg chars med 3960 | out tok med 10000 | final-msg tok med 726
== e0031
  solved    n  77 | test after last edit 0.86 | any run after last edit 0.94 | edits med 1 | first edit med 9 | thought chars/ep med 11857 | arg chars med 4321 | out tok med 5899 | final-msg tok med 128
  done_fail n  17 | test after last edit 0.76 | any run after last edit 1.00 | edits med 2 | first edit med 11 | thought chars/ep med 13458 | arg chars med 4499 | out tok med 6879 | final-msg tok med 118
  to_fail   n  13 | test after last edit 0.46 | any run after last edit 0.54 | edits med 2 | first edit med 13 | thought chars/ep med 22896 | arg chars med 6141 | out tok med 10000 | final-msg tok med 6
  empty     n  13 | test after last edit 0.15 | any run after last edit 0.23 | edits med 0 | first edit med 19.0 | thought chars/ep med 22311 | arg chars med 3172 | out tok med 10000 | final-msg tok med 331
   done_fail ['rrrrrrrerTTTrrrerrTrs--', 'rrrrrTreerrrrrrrrrs-', 'rrrTTreTTrTTTTTTrgrTrs--', 'rrerrrs--', 'rrrrrrrrrrerrerTTTTrrrs--', 'rrrrTrrreeerrs--', 'rrrrrrrrrrTTrrrreerTTrs--', 'rrrrrTrrrreeeTTTTrs--', 'rrrrrTeeeeeeeerTrrs--', 'rrrrrrrrrreTrrs--', 'rrrrrrrrrrrrrTeTreTTers--', 'rrrrrrrrrrrrrrrrrrrrrrrrrerTrs--', 'rrrerrTrs--', 'rrrrrrerrrrrrrrTTs--']
   empty ['rrrrrrrrrrrrrrrrrrrrrrrrrrrrrres', 'rrTrrrrrTr', 'rrrrrrrrrrrrrrrr-', 'rrrrrrwrrTrTTrrr--', 'rrrrrrrrrrrTrTrr--', 'TrrrrrwTTTrrTTTrTrT-', 'TrrrrrrTTrTrrrrrrr--', 'rrrrrrrrrrrrrrrrTTTrrr--', 'rrrrrrrrrrrrrrrrrrrrrrrrrrrrr-', 'TrTrTrrTTrrrrrT-', 'TTrrrTTTT-', 'rrrrrrrrrrrrrrrrrrrrrrrrrrrrrrresr--', 'TrTTrrrrrTrrrrrrrrTT']
== e0032
  solved    n  71 | test after last edit 0.77 | any run after last edit 1.00 | edits med 1 | first edit med 8 | th
[... 977 characters cut ...]

23:CAPS = ("resid_pre", "attn_out", "resid_mid", "mlp_in", "gate", "up", "mlp_act", "mlp_raw", "mlp_out", "resid_post")
150:            d["layer_scalar"] = g(p + "layer_scalar")
276:    def forward(self, ids, out_pos=None, capture=None, cap_layers=None, cap_pos=None, hooks=None, stop_after=None,
282:        capture: iterable of names from CAPS, taken at cap_pos (default out_pos) for cap_layers (default all):
286:          mlp_out (MLP write after its post-norm), resid_post (layer output after layer_scalar).
287:        hooks: {(name, layer): fn(tensor [N, dim]) -> tensor} applied to the full-sequence tensor at that point.
288:        stop_after: stop after this layer index (no log-probs returned).
290:        [n_cap_layers, len(cap_pos), dim]}).
296:        cap_pos_t = torch.as_tensor(out_pos if cap_pos is None else list(cap_pos), dtype=torch.long, device=self.dev)
301:        hooks = hooks or {}
304:            if (name, l) in hooks:
305:                x = hooks[(name, l)](x)
307:                caps[name].append(x[cap_pos_t].float().cpu())
342:            h = h * (d["layer_scalar"] if m else d["layer_scalar"].float())
343:            h = tap("resid_post", l, h)
344:            if stop_after is not None and l >= stop_after:
total 334760
drwxrwxr-x  3 keith keith      4096 Nov 13 19:48 .
drwxrwxr-x  8 keith keith      4096 Nov 13 19:46 ..
drwxrwxr-x 11 keith keith      4096 Nov 13 01:02 ad
-rw-rw-r--  1 keith keith        58 Nov 13 01:02 arms.txt
-rw-r--r--  1 keith keith      5829 Nov 13 00:44 common_recipe.txt
-rw-rw-r--  1 keith keith    109780 Nov 13 00:19 deep_post.jsonl
-rw-rw-r--  1 keith keith   2124997 Nov 13 00:19 deep_pre.jsonl
-rw-rw-r--  1 keith keith   3619905 Nov 13 01:07 eval.pt
-rw-r--r--  1 keith keith       552 Nov 13 00:44 facts.txt
407200, 90774 MiB
2863552, 90794 MiB
 405850    01:58:23 /bin/bash /home/keith/gemma-v2/wsp31/tools/gpu0_session.sh 18300 0.9 -- /home/keith/gemma-v2/wsp31/budget/run_budget.sh
 405857    01:58:23 /home/keith/gemma-v2/tii-venv/bin/python -m vllm.entrypoints.openai.api_server --model /home/keith/gemma-v2/models/gemma-4-31b-it-qat-w4a16-ct --served-model-name base gemma-4-31b 
2857473    05:19:13 /home/keith/gemma-v2/tii-venv/bin/python -m vllm.entrypoints.openai.api_server --model /home/keith/gemma-v2/models/gemma-4-31b-it-qat-w4a16-ct --served-model-name base gemma-4-31b 
3199720    04:33:19 flock /home/keith/gemma-v2/wsp31/gpu0.lock /home/keith/gemma-v2/wsp31/tools/gpu0_session.sh 18300 0.9 -- /home/keith/gemma-v2/wsp31/budget/run_budget.sh

Sat Nov 14 12:50:22 AM EST 2026
-rw-rw-r-- 1 keith keith 18 Nov 14 00:44 queue/TURN_OPEN
authors/claude/scratch/job3/eval.pt
authors/claude/scratch/job3/net_stats.pt
total 44
-rw-rw-r--  1 keith keith  600 Nov 13 22:53 session.log
-rw-rw-r--  1 keith keith 2225 Nov 13 22:53 ids_b007_capped.txt
drwxrwxr-x  5 keith keith 4096 Nov 13 22:53 .
drwxrwxr-x  2 keith keith 4096 Nov 13 22:51 logs
drwxrwxr-x 15 keith keith 4096 Nov 13 20:50 ..
drwxrwxr-x  3 keith keith 4096 Nov 13 20:16 harness_x4
drwxrwxr-x  3 keith keith 4096 Nov 13 20:16 harness_x2
-rw-r--r--  1 keith keith 2849 Nov 13 20:16 README.md
-rw-r--r--  1 keith keith 2623 Nov 13 20:16 summ_budget.py
-rwxr-xr-x  1 keith keith 2077 Nov 13 20:16 run_budget.sh
-rw-r--r--  1 keith keith 2035 Nov 13 20:16 select_ids.py
#!/bin/bash
# Episode-budget comparison: the unedited model on ONE server, arms that differ only in the episode budget.
# Run as the command of a GPU 0 session (it sees PORT):
#   BUDGET=28800 SPEC_K=3 SWA_FIX=1 MAXRANK=32 MAXLORAS=8 EXTRA="--lora-target-modules down_proj o_proj" \
[... 26 lines cut ...]
done
wait
{ echo "# t=$(date +%s)"; curl -s $M; } > $OUT/metrics_end.txt
echo "arms done at $(date +%H:%M:%S)"
python3 $B/summ_budget.py $OUT | tee $OUT/summary.txt
    PID     ELAPSED CMD
 405857    01:58:39 /home/keith/gemma-v2/tii-venv/bin/python -m vllm.entrypoints.openai.api_server --model /home/keith/gemma-v2/models/gemma-4-31b-it-qat-w4a16-ct --served-model-name base gemma-4-31b --host [ip] --port [port] --max-model-len 3276
 415450    01:57:10 timeout 28712 /home/keith/gemma-v2/wsp31/budget/run_budget.sh

server up after 88s (pid/pgid 405857, log /home/keith/gemma-v2/wsp31/budget/logs/server_18300_1113_225143.log)
b007: 120 tasks, 67 with a capped unedited episode -> /home/keith/gemma-v2/wsp31/budget/ids_b007_capped.txt
[line removed by the dashboard scrubber]
base: solved 65 of 120; on the selected 20 of 67; on the rest 45 of 53
aa: solved 73 of 120; on the selected 27 of 67; on the rest 46 of 53
arm x1 budget 10000 streams 4 pid 415464 at 22:53:11
arm x2 budget 20000 streams 4 pid 415466 at 22:53:11
arm x4 budget 40000 streams 8 pid 415468 at 22:53:11
ids.txt
metrics_start.txt
x1
x1.launch.log
x2
x2.launch.log
x4
x4.launch.log
67 results/wsp31/budget/s1_x124/ids.txt
x1 67 22
x2 63 26
x4 67 17
[... 33 lines cut ...]
| x4 | 40,000 | 120 | 160 | 8 |

[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
  may do; the x4 arm gets more streams only so the three arms finish together.
- x1 is the control: the tasks were picked because they did badly once, so a plain rerun will solve a few more by
  chance. Each larger budget is compared with x1 of this run, task by task, never with b007.

301:class Mock:
302-    """Shape-faithful stand-in for g31.G31 (CPU, random numbers) to test this script's plumbing without the GPU."""
303-
304-    def __init__(self):
305-        self.dev = torch.device("cpu")
306-        self.emb = torch.randn(300, tl.D).to(torch.bfloat16)
307-        self.final_norm = torch.rand(tl.D).to(torch.bfloat16)
308-        self.L = {L: dict(post_feedforward_layernorm=torch.randn(tl.D).to(torch.bfloat16),
309-                          layer_scalar=torch.tensor([0.7]).to(torch.bfloat16)) for L in range(60)}
310-        self.e = EditSet()
311-
312-    def set_adapter(self, a):
313-        self.ad = a
314-
315-    def base_weight(self, l, mod):
316-        return self.e.W(l, mod).to(torch.bfloat16)
317-
318-    def rms(self, x, w=None):
319-        return x
320-
[... 85 lines cut ...]
def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("edits", nargs="*")
    ap.add_argument("--port", type=int, required=True)
    ap.add_argument("--out", required=True)
    ap.add_argument("--n", type=int, default=24)
[line removed by the dashboard scrubber]
    ap.add_argument("--streams", type=int, default=8)

Sat Nov 14 12:54:28 AM EST 2026
[    5.5] compiled F30 calls 4 |F|/(beta |delta6|) every 6th ['30.11'] cos(F, delta6) ['0.03'] split-half cos ['0.02'] |F|/(beta |d|) ['30.11']
[    6.5] neuron 0 j 17366 {'use': 'mean', 'cv_mean': '0.102', 'cv_ridge': '0.102', 's_bos_before': '1.221', 'proj': False, 'gbar': '7.930', 'act_cv_cal': '0.222', 'act_prompt': '0.984', 'act_prompt_sd': '0.215', 'act_bos': '1.494', 'act_min': '0.841', 'act_max': '1.660'} old norms ['0.89', '0.88', '0.39'] new ['0.10', '0.01']
[    7.3] neuron 20 j 20942 {'use': 'mean', 'cv_mean': '0.102', 'cv_ridge': '0.104', 's_bos_before': '1.233', 'proj': False, 'gbar': '8.012', 'act_cv_cal': '0.223', 'act_prompt': '0.990', 'act_prompt_sd': '0.215', 'act_bos': '1.517', 'act_min': '0.846', 'act_max': '1.623'} old norms ['0.86', '0.80', '0.41'] new ['0.10', '0.01']
[    8.2] neuron 40 j 17763 {'use': 'mean', 'cv_mean': '0.103', 'cv_ridge': '0.103', 's_bos_before': '1.236', 'proj': False, 'gbar': '8.024', 'act_cv_cal': '0.224', 'act_prompt': '0.981', 'act_prompt_sd': '0.211', 'act_bos': '1.527', 'act_min': '0.826', 'act_max': '1.596'} old norms ['1.06', '1.13', '0.54'] new ['0.10', '0.01']
[    9.1] neuron 59 j 731 {'use': 'mean', 'cv_mean': '0.102', 'cv_ridge': '0.102', 's_bos_before': '1.226', 'proj': False, 'gbar': '8.249', 'act_cv_cal': '0.223', 'act_prompt': '0.983', 'act_prompt_sd': '0.214', 'act_bos': '1.512', 'act_min': '0.843', 'act_max': '1.638'} old norms ['1.94', '1.56', '0.64'] new ['0.10', '0.01']
[    9.1] stats saved; calls 4
[    9.1] built F10 r 1 net_rel 0.00387 rho min/med/max 0.365 0.399 0.411
[    9.1] built F20 r 1 net_rel 0.00407 rho min/med/max 0.386 0.415 0.474
[    9.1] built F20c r 1 net_rel 0.00407 rho min/med/max 0.386 0.415 0.474
[    9.1] arms ['F10', 'F20', 'F20c']
[   12.2] evaluated 0 gawel__pyquery.811cd048.lm_rewri call 2 pos 268 variants 5
[   13.8] evaluated 1 cookiecutter__cookiecutter.b4451 call 17 pos 59 variants 5
[   13.8] SUMMARY P_I1 {"records": 2, "positions": 327, "flip_all": 1.0, "flip_thought": 1.0, "n_thought": 270, "early_close_top": 0, "dlogp_close_inside": 0.06469962508590132, "n_close": 2, "dlogp_close_at_close": 0.7441320419311523, "close_kept": 0, "n_tool": 2, "flip_tool": 2, "kl_base": 1.0000079560352757, "tv_base": 0.52047
[   13.8] SUMMARY F10 {"records": 2, "positions": 327, "flip_all": 1.0, "flip_thought": 1.0, "n_thought": 270, "early_close_top": 0, "dlogp_close_inside": -0.04418005060266565, "n_close": 2, "dlogp_close_at_close": -0.03980207443237305, "close_kept": 0, "n_tool": 2, "flip_tool": 2, "kl_base": 0.9998203592562894, "tv_base": 0.520
[   13.8] SUMMARY F20 {"records": 2, "positions": 327, "flip_all": 1.0, "flip_thought": 1.0, "n_thought": 270, "early_close_top": 0, "dlogp_close_inside": -0.03860316806369358, "n_close": 2, "dlogp_close_at_close": 0.41927194595336914, "close_kept": 0, "n_tool": 2, "flip_tool": 2, "kl_base": 1.000073575827689, "tv_base": 0.52048
[   13.8] SUMMARY F20c {"records": 2, "positions": 327, "flip_all": 1.0, "flip_thought": 1.0, "n_thought": 270, "early_close_top": 0, "dlogp_close_inside": -0.03070588994909216, "n_close": 2, "dlogp_close_at_close": 0.34453344345092773, "close_kept": 0, "n_tool": 2, "flip_tool": 2, "kl_base": 0.9999360717400134, "tv_base": 0.520
[   13.8] SUMMARY neuron_act_over_gbar {"layers": [0, 59], "bos": ["0.16", "0.01"], "prompt_mean": ["0.02", "0.00"], "prompt_sd": ["0.12", "0.11"], "prompt_max": ["0.41", "0.30"], "comp_mean": ["0.01", "0.01"], "comp_sd": ["0.12", "0.11"], "comp_min": ["-0.33", "-0.38"], "comp_max": ["0.57", "0.28"]}
[   13.8] done

real	0m14.897s
user	1m3.699s
sys	0m31.934s
0, 90786 MiB
1, 90812 MiB

    ap.add_argument("--streams", type=int, default=8)
    ap.add_argument("--budget", type=float, default=float(os.environ.get("LEFT", 600)) - 40)
    a = ap.parse_args()
    import torch
    import champ
    import dose
    import fastprobe
    import vclient
    import wlib
    t0 = time.time()
    tok = wlib.tok()
    recs = []
    for l in open(J.PROBE):
        r = json.loads(l)
        if r["n"] <= J.NMAX:
            recs.append(r)
    _, ev = J.pick(recs)
    names = dose.tool_names(recs[:200], tok)
    del recs
    ev = ev[:a.n]
[... 75 lines cut ...]
[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
cmd rc=0
server 2926948 stopped at 19:45:05, hold 195s
1:server up after 72s (pid/pgid 2926948, log /home/keith/gemma-v2/wsp31/authors/claude/scratch/logs/server_18400_1113_194150.log)

4:import json
5:import os
6:import sys
8:import torch
12:import icv_job  # noqa: E402
[line removed by the dashboard scrubber]
## Prediction (written before the run)

predicted_gain_120: -6 to 6

My median is -1: the measured record of thought-length edits on this model is harmful or null, the pre-tests are
teacher-forced or single turns, and whether the paragraph itself would gain tasks as a prompt is not known (I may
not run rollouts, so the paragraph was never tried as a prompt).

Behaviour that should move if the mechanism holds (edit against the reference arm of the same batch,
`code/arm_stats.py` on the two run directories; reference values are b004 base and aa):
[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
[... 50 lines cut ...]

[line removed by the dashboard scrubber]
syntax on 60 or more calls; generation speed under this adapter.

Checker: `check_adapter.py` passes on `edit/` and `ctrl/` (run by `build.sh`).

## Prediction (written before the run)
Sat Nov 14 12:55:14 AM EST 2026

"""meta_check.py (CPU): rebuild (pos, call) of every row of scratch/job1/raw_mlpin.pt (deterministic sampling of
[line removed by the dashboard scrubber]
diagonal-LDA probe on mlp_in separates them on held-out episodes."""
import sys, json, random, torch, collections
A = "/home/keith/gemma-v2/wsp31/authors/claude"
sys.path.insert(0, "/home/keith/gemma-v2/wsp31/tools"); sys.path.insert(0, A + "/code")
import thinklib as tl, probeset
recs = probeset.load("/home/keith/gemma-v2/wsp31/tools/data/probe_screen.jsonl")
raw = torch.load(A + "/scratch/job1/raw_mlpin.pt")
st = torch.load(A + "/code/data/think_stats.pt")
used = st["train_instances"]
import think_job as tj
train, ev = tj.pick_episodes(recs)
assert [r["instance"] for r in train[:len(used)]] == used, "episode order differs"
toolname = collections.Counter()
rows = []
for ei, rec in enumerate(train[:len(used)]):
    n, P, non, ids2 = tl.positions(rec, tj.NTRAIN_CAP)
    rng = random.Random(1234 + ei)
    sel = []
[... 19 lines cut ...]
        ci = max([i for i, s in enumerate(starts) if s <= p], default=0)
        rows.append(dict(ei=ei, pos=p, cls=k, depth=d, ci=ci, gen=gen_cum[ci] + (p - starts[ci] if starts else 0), tools=[tool.get(i) for i in range(ci)]))
assert len(rows) == raw["cls"].numel(), (len(rows), raw["cls"].numel())
assert all(r["cls"] == int(c) for r, c in zip(rows, raw["cls"])) and all(r["depth"] == int(c) for r, c in zip(rows, raw["depth"]))
[line removed by the dashboard scrubber]
json.dump(dict(rows=rows, toolname={str(k): v for k, v in toolname.items()}), open(A + "/scratch/t3_rows.json", "w"))
Sat Nov 14 12:55:21 AM EST 2026
66

./queue/runner.py:586:def split_record(text):
./infra/block3/runner.py:471:def split_record(text):
./infra/runner_patch/runner.py:471:def split_record(text):
queue/dash_collect.py:133:        pred = it.get("predicted_gain_120") or [None, None]
queue/runner.py:476:        lo, hi = it["predicted_gain_120"]
queue/runner.py:534:        p = i["predicted_gain_120"]
queue/runner.py:549:        p = i["predicted_gain_120"]
queue/runner.py:603:    if not re.search(r"predicted_gain_120:\s*-?\d+(\.\d+)?\s*to\s*-?\d+(\.\d+)?", out[SECTIONS[2]]):
queue/runner.py:604:        return None, "prediction section has no 'predicted_gain_120: <low> to <high>' line"
queue/runner.py:687:            pred = re.search(r"predicted_gain_120:\s*(-?\d+(?:\.\d+)?)\s*to\s*(-?\d+(?:\.\d+)?)", rec["sections"][SECTIONS[2]])
queue/runner.py:688:            it["predicted_gain_120"] = [float(pred.group(1)), float(pred.group(2))]
queue/runner.py:[port]:            lo, hi = it["predicted_gain_120"]
queue/runner.py:[port]:        p = it["predicted_gain_120"]

    L += ["## The other authors' entries", ""] + (head + [row(i) for i in others] if others else ["None yet."]) + [""]
    return "\n".join(L)


# ---------------------------------------------------------------- ingest

def reject(src, why):
    (src / "REJECTED").write_text(now() + "\n" + why + "\n")
    log("REJECT", src, why.splitlines()[0][:160])


def split_record(text):
    out, title = {}, ""
    m = re.search(r"^title:\s*(.+)$", text, re.M)
    if m:
        title = m.group(1).replace("|", "/").strip()[:140]
    for i, s in enumerate(SECTIONS):
        a = text.find(s)
        if a < 0:
            return None, f"record.md has no section {s!r}"
[... 59 lines cut ...]
1449:    L += ["", "## Edits", "", "| ledger | author | slug | state | predicted (120) | stage 1 | stage 2 | control | verdict |", "|---|---|---|---|---|---|---|---|---|"]
1455:        L.append(f"| {it['ledger_id']} | {it['author']} | {it['slug']} | {it['state']} | {p[0]:+g} to {p[1]:+g} | {s1} | {s2} | {ct} | {it['verdict']} |")
1720:                log("SUPPLEMENT WAITS", name, why[:400])
1790:        log("SUPPLEMENT ERROR", b["name"], json.dumps(errs)[:400])
1976:                cname = f"{v['cont']}-{it['author']}-{it['slug']}-s2.md"
2557:        why = f"{type(e).__name__}: {e}"[:400]
2648:        why = f"{type(e).__name__}: {e}"[:400]
2726:            log("BOX ADOPT REFUSED", name, "; ".join(probs)[:400])

Sat Nov 14 12:56:58 AM EST 2026
[   63.4] captured 3 [task] call 22 n 3668 pos 118
[   75.5] captured 4 r1chardj0n3s__parse.30da9e4f.fun call 1 n 3302 pos 77
s1_x124: 67 tasks per arm (the capped ones of a 120-task set)
[line removed by the dashboard scrubber]
   unsolved by error: {'Test failures/errors recorded in JUnit X': 25, 'Empty agent patch': 20}
   agent_error: {'Agent exceeded session timeout (7.5 min)': 42, 'none': 25}
[line removed by the dashboard scrubber]
   unsolved by error: {'Test failures/errors recorded in JUnit X': 30, 'Empty agent patch': 8, 'No passing tests recorded in JUnit XML (': 2}
   agent_error: {'none': 43, 'Agent exceeded session timeout (7.5 min)': 24}
[line removed by the dashboard scrubber]
   unsolved by error: {'Empty agent patch': 34, 'Test failures/errors recorded in JUnit X': 15, 'Invalid JUnit XML: Missing JUnit XML rep': 1}
   agent_error: {'Sandbox execution error: litellm.BadRequestError: OpenAIExce': 47, 'Sandbox execution error: litellm.ContextWindowExceededError:': 15, 'none': 5}
x2 vs x1 on 67 paired tasks: gained 8 lost 3 net 5 = +4.2 points of the 120-task set; sign test p 0.227
x4 vs x1 on 67 paired tasks: gained 5 lost 10 net -5 = -4.2 points of the 120-task set; sign test p 0.302

grading
logs
patches
requests
task_results.jsonl
traces
[task]
[task]
HIPS__autograd.ac044f0d.lm_rewrite__cetfos0u
Mimino666__langdetect.a1598f1a.func_basic__rku38xih
Mimino666__langdetect.a1598f1a.lm_rewrite__dk9wx21q
junit.xml
stderr.log
stdout.log
[line removed by the dashboard scrubber]

failed_pool_v2_screen_sft_p7
failed_pool_v2_screen_w5r_p8
paused_pool_v2_screen_w5r_p13
pool_v1_rescreen
pool_v21_screen_w3a
pool_v21_screen_w3b
pool_v2_screen
pool_v2_screen_r1
pool_v2_screen_sft_f24
pool_v2_screen_sft_f4
pool_v2_screen_sft_f47
pool_v2_screen_sft_f7
pool_v2_screen_sft_p1
pool_v2_screen_sft_p10
pool_v2_screen_sft_p2
pool_v2_screen_sft_p3
pool_v2_screen_sft_p4
pool_v2_screen_sft_p5
pool_v2_screen_sft_p6
pool_v2_screen_sft_p8
[... 37 lines cut ...]
evaluator_routes.json
manifest.json
snapshots
tasks.jsonl
[line removed by the dashboard scrubber]
{'instance_id': '[task]', 'repo': 'DanielSank/constraintula', 'base_commit': 'ce8db824210673e74b99a8ef7868db9535823682', 'problem_statement': 'raise non-ValueError\nIn 8feb313825706ebc5e1ac2d1e0d3076576a0cc11 I added a `Valu', 'hints_text': '', 'created_at': '2020-10-28 16:58:35'}
Sat Nov 14 12:57:41 AM EST 2026
[  121.2] captured 7 [task] call 22 n 4167 pos 77

Sat Nov 14 01:04:25 AM EST 2026
38:[  519.4] compiled F10 calls 35 |F|/(beta |delta6|) every 6th ['1.00', '0.72', '0.58', '0.95', '0.75', '0.73', '0.96', '0.67', '0.76', '0.84'] cos(F, delta6) ['1.00', '0.54', '0.43', '0.72', '0.66', '0.64', '0.82', '0.69', '0.68', '0.66'] split-half cos ['1.00', '0.99', '0.93', '0.88', '0.95', '0.87', '0.96', '0.86', '0.82', '0.80'] |F|/(beta |d|) ['1.00', '0.38', '0.22', '0.24', '0.50', '0.41', '0.62', '0.32', '0.22', '0.31']
39:[  519.4] compiled F20 calls 35 |F|/(beta |delta6|) every 6th ['1.00', '0.67', '0.51', '0.92', '0.72', '0.70', '0.95', '0.65', '0.72', '0.80'] cos(F, delta6) ['1.00', '0.56', '0.42', '0.76', '0.61', '0.66', '0.84', '0.73', '0.72', '0.67'] split-half cos ['1.00', '0.99', '0.90', '0.90', '0.90', '0.92', '0.98', '0.92', '0.91', '0.90'] |F|/(beta |d|) ['1.00', '0.35', '0.19', '0.23', '0.49', '0.39', '0.61', '0.31', '0.21', '0.29']
40:[  519.4] compiled F30 calls 35 |F|/(beta |delta6|) every 6th ['1.00', '0.67', '0.45', '0.83', '0.72', '0.71', '0.94', '0.64', '0.71', '0.79'] cos(F, delta6) ['1.00', '0.57', '0.43', '0.79', '0.62', '0.65', '0.84', '0.73', '0.72', '0.68'] split-half cos ['1.00', '1.00', '0.97', '0.92', '0.96', '0.96', '0.99', '0.95', '0.94', '0.93'] |F|/(beta |d|) ['1.00', '0.35', '0.17', '0.21', '0.48', '0.40', '0.60', '0.31', '0.20', '0.29']
52:[  524.2] stats saved; calls 35
[  519.6] neuron 0 j 17366 {'use': 'ridge', 'cv_mean': '0.155', 'cv_ridge': '0.027', 's_bos_before': '0.980', 'proj': False, 'gbar': '7.984', 'act_cv_cal': '0.054', 'act_prompt': '0.917', 'act_prompt_sd': '0.076', 'act_bos': '0.960', 'act_min': '0.801', 'act_max': '1.338'} old norms ['0.89', '0.88', '0.39'] new ['0.38', '0.05']
[  519.9] neuron 6 j 15052 {'use': 'ridge', 'cv_mean': '0.173', 'cv_ridge': '0.053', 's_bos_before': '0.744', 'proj': False, 'gbar': '7.891', 'act_cv_cal': '0.104', 'act_prompt': '0.842', 'act_prompt_sd': '0.149', 'act_bos': '0.564', 'act_min': '0.598', 'act_max': '1.335'} old norms ['0.80', '0.68', '0.25'] new ['0.26', '0.03']
[  520.3] neuron 12 j 17885 {'use': 'ridge', 'cv_mean': '0.153', 'cv_ridge': '0.041', 's_bos_before': '1.419', 'proj': False, 'gbar': '7.873', 'act_cv_cal': '0.080', 'act_prompt': '0.815', 'act_prompt_sd': '0.131', 'act_bos': '2.056', 'act_min': '0.652', 'act_max': '1.257'} old norms ['0.96', '0.94', '0.49'] new ['0.46', '0.06']
[  520.7] neuron 18 j 3432 {'use': 'ridge', 'cv_mean': '0.221', 'cv_ridge': '0.048', 's_bos_before': '1.801', 'proj': False, 'gbar': '7.865', 'act_cv_cal': '0.095', 'act_prompt': '0.905', 'act_prompt_sd': '0.119', 'act_bos': '3.301', 'act_min': '0.641', 'act_max': '1.427'} old norms ['1.09', '1.12', '0.50'] new ['0.31', '0.04']
[  521.1] neuron 24 j 19454 {'use': 'ridge', 'cv_mean': '0.139', 'cv_ridge': '0.027', 's_bos_before': '1.212', 'proj': False, 'gbar': '7.939', 'act_cv_cal': '0.054', 'act_prompt': '0.923', 'act_prompt_sd': '0.072', 'act_bos': '1.482', 'act_min': '0.781', 'act_max': '1.135'} old norms ['1.44', '0.97', '0.50'] new ['0.59', '0.07']
[  521.5] neuron 30 j 7355 {'use': 'ridge', 'cv_mean': '0.283', 'cv_ridge': '0.056', 's_bos_before': '0.789', 'proj': False, 'gbar': '7.797', 'act_cv_cal': '0.110', 'act_prompt': '0.834', 'act_prompt_sd': '0.152', 'act_bos': '0.638', 'act_min': '0.602', 'act_max': '1.318'} old norms ['1.70', '0.90', '0.45'] new ['1.17', '0.15']
[  521.9] neuron 36 j 1298 {'use': 'ridge', 'cv_mean': '0.288', 'cv_ridge': '0.061', 's_bos_before': '0.686', 'proj': False, 'gbar': '7.794', 'act_cv_cal': '0.120', 'act_prompt': '0.760', 'act_prompt_sd': '0.199', 'act_bos': '0.483', 'act_min': '0.568', 'act_max': '1.429'} old norms ['0.84', '0.93', '0.52'] new ['1.50', '0.19']
[  522.4] neuron 42 j 4522 {'use': 'ridge', 'cv_mean': '0.353', 'cv_ridge': '0.053', 's_bos_before': '0.894', 'proj': False, 'gbar': '7.854', 'act_cv_cal': '0.103', 'act_prompt': '0.885', 'act_prompt_sd': '0.152', 'act_bos': '0.813', 'act_min': '0.624', 'act_max': '1.279'} old norms ['1.45', '1.24', '0.56'] new ['0.59
[... 1000 characters cut ...]

Sat Nov 14 01:08:52 AM EST 2026
Sat Nov 14 12:55:38 AM EST 2026
1:[    0.9] train (the calls of turn 6) 35 eval 8 capture deadline 600.0 deadline 1150.0
53:[  524.3] built F10 r 1 net_rel 0.00674 rho min/med/max 0.002 0.008 0.028
54:[  524.3] built F20 r 1 net_rel 0.00675 rho min/med/max 0.004 0.014 0.051
55:[  524.4] built F30 r 1 net_rel 0.00678 rho min/med/max 0.005 0.021 0.076
56:[  524.4] built F20c r 1 net_rel 0.00675 rho min/med/max 0.004 0.014 0.051
57:[  524.5] built F10c r 1 net_rel 0.00674 rho min/med/max 0.002 0.008 0.028
58:[  524.5] built F30c r 1 net_rel 0.00678 rho min/med/max 0.005 0.021 0.076
P_I1  rec 8 pos 1278 flips 0.012 th 0.020 early 0 dl_in -0.07 dl_close -0.192 kept 7/8 toolflip 0/8 KL 0.0076 repro - shiftcos - matched None/None
F10   rec 8 pos 1278 flips 0.009 th 0.013 early 0 dl_in -0.07 dl_close -0.037 kept 7/8 toolflip 0/8 KL 0.0022 repro 0.47 shiftcos 0.34 matched 7/15
      cos 0.25 0.70 0.52 0.41 0.43 0.47 0.53 0.48 0.47 0.40 0.33 | ratio 1.97 0.92 1.09 1.05 0.78 0.81 0.82 0.77 0.78 0.76 0.75
F20   rec 8 pos 1278 flips 0.014 th 0.024 early 0 dl_in -0.24 dl_close -0.101 kept 7/8 toolflip 0/8 KL 0.0075 repro 0.48 shiftcos 0.38 matched 12/15
      cos 0.25 0.71 0.55 0.43 0.45 0.48 0.57 0.52 0.50 0.45 0.41 | ratio 3.99 1.76 1.83 1.93 1.42 1.52 1.56 1.39 1.35 1.24 1.28
F30   rec 8 pos 1278 flips 0.018 th 0.030 early 0 dl_in -0.28 dl_close -0.209 kept 7/8 toolflip 0/8 KL 0.0172 repro -0.09 shiftcos 0.38 matched 12/15
      cos 0.25 0.72 0.55 0.45 0.45 0.50 0.58 0.51 0.51 0.45 0.43 | ratio 5.97 2.60 2.62 2.48 2.21 2.28 2.34 2.05 1.96 1.79 1.82
A20   rec 8 pos 1278 flips 0.013 th 0.022 early 0 dl_in +0.05 dl_close -0.171 kept 7/8 toolflip 0/8 KL 0.0072 repro 0.45 shiftcos 0.34 matched 10/15
      cos 0.25 0.68 0.51 0.46 0.46 0.48 0.54 0.45 0.46 0.42 0.37 | ratio 3.98 2.04 2.37 2.18 1.87 2.18 1.88 2.27 2.12 1.94 2.48
F20c  rec 8 pos 1278 flips 0.009 th 0.017 early 0 dl_in -0.12 dl_close -0.001 kept 7/8 toolflip 0/8 KL 0.0011 repro -0.10 shiftcos 0.11 matched 6/15
      cos -0.04 -0.00 0.05 -0.03 0.02 0.06 0.04 0.04 0.10 0.10 0.09 | ratio 3.95 1.58 1.46 1.02 1.02 1.22 1.25 1.04 1.07 1.02 1.33
F10c  rec 8 pos 1278 flips 0.007 th 0.011 early 0 dl_in -0.06 dl_close -0.005 kept 7/8 toolflip 0/8 KL 0.0005 repro -0.03 shiftcos 0.13 matched 3/15
      cos -0.03 -0.01 0.04 -0.01 0.03 0.06 0.03 0.07 0.15 0.14 0.11 | ratio 1.98 0.82 0.80 0.58 0.54 0.66 0.66 0.60 0.63 0.62 0.76
F30c  rec 8 pos 1278 flips 0.007 th 0.011 early 0 dl_in -0.14 dl_close -0.016 kept 7/8 toolflip 0/8 KL 0.0019 repro -0.21 shiftcos 0.11 matched 4/15
      cos -0.04 -0.02 0.05 -0.06 0.00 0.04 0.05 0.04 0.08 0.06 0.08 | ratio 5.92 2.35 2.07 1.51 1.56 1.82 1.85 1.48 1.52 1.43 1.95
neuron_act_over_gbar ['0.98', '0.95', '0.95', '0.95', '0.96', '0.94', '0.96', '0.98', '0.97', '0.99', '0.99'] ['0.07', '0.12', '0.09', '0.12', '0.06', '0.10', '0.11', '0.10', '0.09', '0.09', '0.06'] ['0.92', '0.84', '0.80', '0.89', '0.91', '0.82', '0.75', '0.89', '0.91', '0.85', '0.92'] ['0.96', '0.56', '2.06', '3.29', '1.48', '0.63', '0.48', '0.82', '0.86', '0.54', '0.25']

P_I1  flips 0.012 th 0.020 dl_in -0.07 dl_close -0.192 toolflip 0/8 KL 0.0076 repro - shiftcos - matched None/None
P_I2  flips 0.011 th 0.022 dl_in +0.00 dl_close +0.079 toolflip 0/8 KL 0.0026 repro - shiftcos - matched None/None
A10   flips 0.008 th 0.009 dl_in +0.03 dl_close -0.062 toolflip 0/8 KL 0.0023 repro 0.50 shiftcos 0.31 matched 7/15
      cos 0.25 0.68 0.52 0.45 0.45 0.46 0.52 0.42 0.44 0.40 0.35 | ratio 1.98 1.03 1.21 1.08 0.96 1.12 0.99 1.19 1.14 1.06 1.29
B10   flips 0.006 th 0.009 dl_in +0.04 dl_close +0.028 toolflip 0/8 KL 0.0006 repro -0.01 shiftcos 0.22 matched 5/14
      cos 0.24 0.36 0.36 0.40 0.30 0.14 0.26 0.24 0.29 0.28 0.20 | ratio 1.83 0.75 0.77 0.77 0.73 0.65 0.68 0.72 0.66 0.65 0.73
A05   flips 0.005 th 0.004 dl_in -0.02 dl_close -0.014 toolflip 0/8 KL 0.0008 repro 0.36 shiftcos 0.29 matched 5/15
      cos 0.26 0.65 0.52 0.46 0.46 0.40 0.51 0.45 0.46 0.41 0.40 | ratio 1.01 0.53 0.62 0.58 0.50 0.59 0.52 0.63 0.64 0.64 0.71
B05   flips 0.005 th 0.009 dl_in +0.01 dl_close -0.000 toolflip 0/8 KL 0.0004 repro 0.05 shiftcos 0.20 matched 3/14
      cos 0.23 0.34 0.33 0.34 0.33 0.10 0.30 0.26 0.30 0.30 0.24 | ratio 0.94 0.46 0.42 0.44 0.41 0.47 0.49 0.52 0.52 0.54 0.56
A10c  flips 0.007 th 0.011 dl_in -0.09 dl_close +0.008 toolflip 0/8 KL 0.0007 repro 0.02 shiftcos 0.13 matched 4/15
      cos -0.03 -0.04 0.08 -0.00 0.05 0.06 0.09 0.09 0.15 0.13 0.09 | ratio 1.97 0.86 0.81 0.59 0.60 0.71 0.72 0.76 0.79 0.75 0.89
B10c  flips 0.002 th 0.004 dl_in -0.06 dl_close -0.004 toolflip 0/8 KL 0.0003 repro -0.07 shiftcos 0.20 matched 1/14
      cos -0.06 0.04 0.11 0.03 0.05 0.11 0.09 0.12 0.19 0.21 0.14 | ratio 1.85 0.65 0.59 0.54 0.53 0.46 0.51 0.51 0.51 0.50 0.59
A20   flips 0.013 th 0.022 dl_in +0.05 dl_close -0.171 toolflip 0/8 KL 0.0072 repro 0.45 shiftcos 0.34 matched 10/15
      cos 0.25 0.68 0.51 0.46 0.46 0.48 0.54 0.45 0.46 0.42 0.37 | ratio 3.98 2.04 2.37 2.18 1.87 2.18 1.88 2.27 2.12 1.94 2.48
B20   flips 0.009 th 0.017 dl_in +0.08 dl_close +0.034 toolflip 0/8 KL 0.0015 repro -0.22 shiftcos 0.21 matched 5/14
      cos 0.24 0.37 0.35 0.40 0.31 0.17 0.29 0.26 0.28 0.27 0.18 | ratio 3.66 1.48 1.53 1.50 1.43 1.20 1.53 1.57 1.42 1.31 1.33
Traceback (most recent call last):
  File "<stdin>", line 5, in <module>
TypeError: string indices must be integers, not 'str'
42:[  401.7] direction I2 records 35 |d|/|resid| every 6th ['0.001', '0.001', '0.001', '0.001', '0.002', '0.002', '0.003', '0.004', '0.003', '0.005'] split-half cos ['0.90', '0.73', '0.82', '0.88', '0.68', '0.51', '0.64', '0.46', '0.35', '0.23'] |delta|/|d| ['1.00', '0.56', '0.35', '0.22', '0.66', '0.66', '0.61', '0.46', '0.30', '0.39']
43:[  401.7] cos(d_I1, d_I2) every 6th ['0.86', '0.58', '0.61', '0.77', '0.46', '0.35', '0.44', '0.42', '0.38', '0.28']
Sat Nov 14 01:09:43 AM EST 2026

18:st = torch.load(A + "/code/data/" + ("fcv2_stats.pt" if job == "job9" else "fcv_stats.pt"))
25:if job == "job9":
27:ORDER = ["P_I15", "G20", "G20c", "G10", "G10c", "F20"] if job == "job9" else ["P_I1", "F10", "F10c", "F20", "F20c", "F30", "F30c", "A20"]
31:def pair_txt():
50:    if job == "job9":
63:    for k in ORDER:
82:    for k in [x for x in ORDER if not x.startswith("P_")]:
98:    p = A + "/scratch/job9/gen.json"
139:       "{{PAIR}}": pair_txt(),
2
Sat Nov 14 01:14:32 AM EST 2026
[  126.5] phase 1 15 jsvine__pdfplumber.02ff4313.func call 3 pos 291
[This command modified 2 files you've previously read: mkrec8.py, rec_f20.md. Call Read before editing.]

#!/bin/bash
# fin9.sh SLUG G10|G20   (author claude, turn 8; CPU): builds queue/SLUG/{edit,ctrl,build.sh}, runs the checker and
[line removed by the dashboard scrubber]
. /home/keith/gemma-v2/env.sh
cd /home/keith/gemma-v2/wsp31/authors/claude
T=$R/wsp31/tools
export WSP31_WORK=$PWD/scratch
[line removed by the dashboard scrubber]
Q=queue/$SLUG
[ -e $Q/READY ] && { echo "already READY"; exit 1; }
mkdir -p $Q code/data
[ -f code/data/fcv2_stats.pt ] || cp scratch/job9/fcv2_stats.pt code/data/fcv2_stats.pt
cat > $Q/build.sh <<EOF
#!/bin/bash
# rebuilds edit/ and ctrl/ of $SLUG (CPU, about a minute). Needs code/data/fcv2_stats.pt: the self-consistent per-layer
# writes of the paragraph pair measured by code/fcv2_job.py (\`flock \$R/wsp31/gpu0.lock timeout 1560 code/gpu_hold9.sh\`),
# which itself needs code/data/fcv_stats.pt (always-on neurons, MLP output RMS, norm weights, layer scalars;
# code/fcv_job.py, code/gpu_hold8.sh) and code/data/icv_stats.pt of turn 6 (fit calls, shift of I1; code/icv_job.py).
. /home/keith/gemma-v2/env.sh
cd /home/keith/gemma-v2/wsp31/authors/claude
[... 5 lines cut ...]
"""fcv2_build.py OUT_DIR G10|G20 edit|ctrl  (author claude, turn 8; CPU, seconds)

Rebuilds one arm of code/fcv2_job.py bit for bit from the statistics that job saved (code/data/fcv2_stats.pt: the
self-consistent per-layer writes G_L[beta] of the paragraph pair I1 + I5, the always-on neuron of every layer, the MLP output RMS, the norm weights
and layer scalars). edit: the down column of the always-on neuron of layer L writes G_L[beta]. ctrl: the same code
path with the sign of every coordinate of the written vector flipped at random (seed 6000 + L): identical neurons,
identical gate and up rows, identical per-coordinate magnitudes of the write, so identical |dW|_F.
"""

Sat Nov 14 01:24:05 AM EST 2026
18:[  522.3] compiled G20 |G|/|F20| every 6th ['1.22', '0.91', '1.55', '1.19', '0.96', '1.01', '1.02', '1.04', '1.04', '1.08'] cos(G, F20) ['0.92', '0.86', '0.22', '0.72', '0.83', '0.86', '0.97', '0.87', '0.83', '0.82'] split-half cos ['1.00', '0.99', '0.99', '0.95', '0.96', '0.97', '0.99', '0.94', '0.93', '0.91']
19:[  522.3] compiled G10 |G|/|F20| every 6th ['0.61', '0.49', '0.79', '0.66', '0.48', '0.50', '0.51', '0.51', '0.52', '0.52'] cos(G, F20) ['0.92', '0.79', '0.28', '0.70', '0.83', '0.87', '0.96', '0.88', '0.81', '0.84'] split-half cos ['1.00', '0.99', '0.95', '0.85', '0.92', '0.87', '0.97', '0.84', '0.80', '0.76']
21:[  522.4] built G20 r 1 net_rel 0.00677 rho min/med/max 0.005 0.015 0.068
22:[  522.4] built G10 r 1 net_rel 0.00674 rho min/med/max 0.003 0.008 0.036
23:[  522.4] built G20c r 1 net_rel 0.00677 rho min/med/max 0.005 0.015 0.068
24:[  522.5] built G10c r 1 net_rel 0.00674 rho min/med/max 0.003 0.008 0.036
42:python part done after 732 s; left for generation 400
P_I15 rec 8 pos 1278 flips 0.019 th 0.037 early 0 dl_in -0.01 dl_close -0.202 kept 7/8 toolflip 0/8 KL 0.0077 repro - shiftcos - matched None/None
G20   rec 8 pos 1278 flips 0.015 th 0.026 early 0 dl_in -0.20 dl_close -0.106 kept 7/8 toolflip 0/8 KL 0.0060 repro 0.40 shiftcos 0.36 matched 13/24
      cos 0.23 0.75 0.59 0.56 0.49 0.50 0.59 0.52 0.53 0.47 0.40 | ratio 4.98 1.80 2.20 3.02 1.51 1.64 1.54 1.31 1.30 1.19 1.22
F20   rec 8 pos 1278 flips 0.014 th 0.024 early 0 dl_in -0.24 dl_close -0.101 kept 7/8 toolflip 0/8 KL 0.0075 repro 0.31 shiftcos 0.36 matched 14/24
      cos 0.21 0.69 0.50 0.47 0.46 0.49 0.57 0.53 0.51 0.47 0.46 | ratio 4.11 1.57 1.51 1.73 1.41 1.65 1.51 1.32 1.31 1.22 1.21
G10   rec 8 pos 1278 flips 0.010 th 0.015 early 0 dl_in -0.13 dl_close -0.019 kept 7/8 toolflip 0/8 KL 0.0019 repro 0.39 shiftcos 0.33 matched 11/24
      cos 0.22 0.75 0.59 0.54 0.47 0.45 0.59 0.55 0.53 0.48 0.34 | ratio 2.49 0.94 1.20 1.63 0.83 0.88 0.80 0.74 0.76 0.73 0.68
G20c  rec 8 pos 1278 flips 0.011 th 0.017 early 0 dl_in -0.11 dl_close +0.015 kept 8/8 toolflip 0/8 KL 0.0015 repro -0.15 shiftcos 0.12 matched 8/24
      cos -0.04 0.03 0.07 -0.08 0.02 0.03 0.06 0.08 0.08 0.09 0.05 | ratio 4.96 1.64 1.29 1.20 1.10 1.33 1.23 1.00 1.07 1.04 1.33
G10c  rec 8 pos 1278 flips 0.009 th 0.015 early 0 dl_in -0.11 dl_close +0.007 kept 8/8 toolflip 0/8 KL 0.0007 repro -0.05 shiftcos 0.13 matched 7/24
      cos -0.03 0.03 0.08 -0.03 0.05 0.03 0.07 0.08 0.13 0.14 0.07 | ratio 2.50 0.85 0.72 0.67 0.59 0.71 0.66 0.59 0.63 0.63 0.72

Sat Nov 14 01:27:43 AM EST 2026
Sat Nov 14 01:11:50 AM EST 2026
Sat Nov 14 01:27:43 AM EST 2026
[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]
22
[line removed by the dashboard scrubber]
cmd rc=0
server 1351839 stopped at 01:27:43, hold 221s

scratch/job9/gen.json [G10]
   17 episodes, 17 generated answers, 17 with a complete tool call
   1. unpaired string delimiter: 0 of 17 tool calls = 0.0% (unedited 0.7% to 1.1%, mark 3%; chance of 0 or more at the highest unedited rate 1): tool-call syntax UNDECIDED
      no verdict: NOT BROKEN needs 60 tool calls and a share below 3%; BROKEN needs a share of 3% or more and a chance of 0.001 or less for the calls
[line removed by the dashboard scrubber]

scratch/job9/gen.json [F30]
   17 episodes, 17 generated answers, 17 with a complete tool call
   1. unpaired string delimiter: 0 of 17 tool calls = 0.0% (unedited 0.7% to 1.1%, mark 3%; chance of 0 or more at the highest unedited rate 1): tool-call syntax UNDECIDED
      no verdict: NOT BROKEN needs 60 tool calls and a share below 3%; BROKEN needs a share of 3% or more and a chance of 0.001 or less for the calls
[line removed by the dashboard scrubber]

6
   1. unpaired string delimiter: 0 of 17 tool calls = 0.0% (unedited 0
   1. unpaired string delimiter: 0 of 17 tool calls = 0.0% (unedited 0
   1. unpaired string delimiter: 0 of 17 tool calls = 0.0% (unedited 0
   1. unpaired string delimiter: 0 of 17 tool calls = 0.0% (unedited 0
   1. unpaired string delimiter: 0 of 17 tool calls = 0.0% (unedited 0
   1. unpaired string delimiter: 0 of 17 tool calls = 0.0% (unedited 0
  File "/home/keith/gemma-v2/wsp31/authors/claude/code/mkrec8.py", line 45
    return "
           ^
SyntaxError: unterminated string literal (detected at line 45)
  File "/home/keith/gemma-v2/wsp31/authors/claude/code/mkrec8.py", line 45
    return "
           ^
SyntaxError: unterminated string literal (detected at line 45)
Sat Nov 14 01:28:07 AM EST 2026

28	six = lambda v: " ".join("%.2f" % x for x in (list(v[::6]) + [v[-1]]))
29	
30	
31	def pair_txt():
32	    i = st.get("pair_info")
33	    if not i:
34	        return ""
35	    r = [a / (b + 1e-9) for a, b in zip(i["norm15"], i["norm1"])]
36	    return ("size of d15 over size of d1: %s; cosine(d15, d1): %s; size of (d15 - d1) over size of d1: %s; split-half cosine "
37	            "(even against odd fit calls) of d15: %s; of d15 - d1: %s" % (six(r), six(i["cos_15_1"]), six(i["extra_rel"]), six(i["split15"]), six(i["split_extra"])))
38	
39	
40	def compile_g():
41	    L = ["| dose | size of the write over size of F20 (think-briefly alone, dose 2) | cosine with F20 | split-half cosine (even against odd fit calls) |", "|---|---|---|---|"]
42	    for k in sorted(x for x in st["dinfo"] if x.startswith("G")):
43	        i = st["dinfo"][k]
44	        L.append("| %s | %s | %s | %s |" % (k[1], six(i["ratio_F20"]), six(i["cos_F20"]), six(i["split_cos"])))
45	    return "
46	".join(L)
47	
48	
49	def compile_txt():
50	    if job == "job9":
51	        return compile_g()
52	    L = ["Layers [0, 6, 12, 18, 24, 30, 36, 42, 48, 54, 59]; `delta6` = the turn-6 write `d_L - layer_scalar_L d_(L-1)`:", ""]
53	    L += ["| dose | size of the write over dose x size of delta6 | cosine with delta6 | split-half cosine (even against odd fit calls) |", "|---|---|---|---|"]
54	    for k in sorted(x for x in st["dinfo"] if x.startswith("F")):
55	        i = st["dinfo"][k]
56	        L.append("| %s | %s | %s | %s |" % (k[1] + "." + k[2] if k[2] != "0" else k[1], six(i["ratio_to_turn6"]), six(i["cos_turn6"]), six(i["split_cos"])))
57	    return "\n".join(L)

rec_f20.out.md:1
rec_g20.out.md:1
record fcv-brief-b20 ok
title (134 chars): Think-briefly instruction vector, self-consistent compile (dose 2): each layer writes only what the paragraph's mean shift still lacks
   ## Hypothesis 6479 chars 
   ## Exact recipe 12000 chars (CUT at 12000)
   ## Prediction (written before the run) 1644 chars 
record fcv-brief-exact-b20 ok
title (138 chars): Instruction pair vector: think briefly + be exact about what the issue asks, compiled self-consistently into one neuron per layer (dose 2)
   ## Hypothesis 7228 chars 
   ## Exact recipe 11118 chars 
   ## Prediction (written before the run) 1961 chars 
Sat Nov 14 01:28:21 AM EST 2026

12038
Self-consistent write (`fcv_job.capture`). For each fit call: one unedited pass that stores `resid_post_L` at P for
all 60 layers, then one pass with a hook after every layer L, in layer order:
`s_L = mean over P of (resid_post_L of this pass - stored resid_post_L)`; `c_L = beta d_L - s_L`; `c_L` is added to
`resid_post_L` at every position except BOS. After the hook the mean shift over P is `beta d_L` exactly, so the next
layer sees the shifted input it would see under the instruction, and `c_L` is only what layer L and the residual
path did not supply. `F_L = mean over the 35 calls of c_L` (dose 2 has its own pass; the response is not linear in
beta). Layers [0, 6, 12, 18, 24, 30, 36, 42, 48, 54, 59]; `delta6` = the turn-6 write `d_L - layer_scalar_L d_(L-1)`:

| dose | size of the write over dose x size of delta6 | cosine with delta6 | split-half cosine (even against odd fit calls) |
|---|---|---|---|
| 1 | 1.00 0.72 0.58 0.95 0.75 0.73 0.96 0.67 0.76 0.84 0.92 | 1.00 0.54 0.43 0.72 0.66 0.64 0.82 0.69 0.68 0.66 0.48 | 1.00 0.99 0.93 0.88 0.95 0.87 0.96 0.86 0.82 0.80 0.90 |
| 2 | 1.00 0.67 0.51 0.92 0.72 0.70 0.95 0.65 0.72 0.80 0.87 | 1.00 0.56 0.42 0.76 0.61 0.66 0.84 0.73 0.72 0.67 0.48 | 1.00 0.99 0.90 0.90 0.90 0.92 0.98 0.92 0.91 0.90 0.95 |
| 3 | 1.00 0.67 0.45 0.83 0.72 0.71 0.94 0.64 0.71 0.79 0.88 | 1.00 0.57 0.43 0.79 0.62 0.65 0.84 0.73 0.72 0.68 0.49 | 1.00 1.00 0.97 0.92 0.96 0.96 0.99 0.95 0.94 0.93 0.96 |

Always-on neuron (one per layer, replaced outright
e prompted one. Eight calls is a small sample: read differences of 0.05 in "reproduced" or
2 flips as nothing.

| arm | calls / positions | flips | flips in thought | close on top inside | dlogp close inside | dlogp close at close | tool flips | KL from base | reproduced | shift cos | prompt flips matched |
[... 43 lines cut ...]
The tools lane's rule (change of log p(close) at close below -0.05: expect longer ordinary thoughts) flags this
entry's -0.10 on 8 closes, as it flagged 0031 (-0.17), whose rollout thoughts got shorter; the paragraph itself has
-0.19.

[line removed by the dashboard scrubber]
[line removed by the dashboard scrubber]

Checker: `check_adapter.py` passes on `edit/` and `ctrl/` (run by `build.sh`): 180 modules, r = 1.

record fcv-brief-b20 ok
title (134 chars): Think-briefly instruction vector, self-consistent compile (dose 2): each layer writes only what the paragraph's mean shift still lacks
   ## Hypothesis 6510 chars 
   ## Exact recipe 11882 chars 
   ## Prediction (written before the run) 1644 chars 
record fcv-brief-exact-b20 ok
title (138 chars): Instruction pair vector: think briefly + be exact about what the issue asks, compiled self-consistently into one neuron per layer (dose 2)
   ## Hypothesis 7228 chars 
   ## Exact recipe 11125 chars 
   ## Prediction (written before the run) 1961 chars 
Sat Nov 14 01:28:57 AM EST 2026
size of d15 over size of d1: 1.22 1.16 1.24 1.27 1.03 1.01 1.02 1.03 1.02 0.98 1.02; cosine(d15, d1): 0.92 0.92 0.85 0.92 0.83 0.92 0.96 0.94 0.94 0.91 0.94; size of (d15 - d1) over size of d1: 0.50 0.46 0.65 0.53 0.60 0.41 0.28 0.34 0.35 0.41 0.34; split-half cosine (even against odd fit calls) of d15: 0.89 0.98 0.96 0.96 0.81 0.87 0.87 0.75 0.68 0.60 0.73; of d15 - d1: -0.08 0.86 0.87 0.79 0.54 0.39 0.02 -0.15 -0.26 -0.18 -0.17.

Self-consistent write (`fcv2_job.phase2`, the method of `code/fcv_job.py`). For each fit call one pass with a hook
after every layer L, in layer order: `s_L = mean over P of (resid_post_L of this pass - resid_post_L of the unedited
pass)`; `c_L = beta d15_L - s_L`; `c_L` is added to `resid_post_L` at every position except BOS. After the hook the
mean shift over P is `beta d15_L` exactly, so `c_L` is what layer L and the residual path did not supply.
`G_L = mean over the 35 calls of c_L`. Against the sister entry's write `F20` (I1 alone, dose 2), same layers:

| dose | size of the write over size of F20 (think-briefly alone, dose 2) | cosine with F20 | split-half cosine (even against odd fit calls) |
[... 88 lines cut ...]

[line removed by the dashboard scrubber]
prompt in rollouts; generation speed under this adapter (rank 1 on 180 matrices, the shape of 0031, which ran at
[line removed by the dashboard scrubber]

Checker: `check_adapter.py` passes on `edit/` and `
-----
[This command modified 1 file you've previously read: mkrec8.py. Call Read before editing.]

96	def gen_txt():
97	    p = A + "/scratch/job9/gen.json"
98	    if not os.path.exists(p):
99	        return "NOT RUN: the second GPU hold did not finish inside the turn; no served generation exists for this entry."
100	    g = json.load(open(p))
[line removed by the dashboard scrubber]
102	    nm = {"reference": "unedited", "prompt_I1": "unedited, paragraph in the prompt"}
103	    nm.update(LBL)
104	    for a in g["arms"]:
105	        s = a["summary"]
106	        n = nm.get(a["name"], a["name"])
[line removed by the dashboard scrubber]
108	            n = "**this entry**: " + n
[line removed by the dashboard scrubber]

record fcv-brief-b20 ok
title (134 chars): Think-briefly instruction vector, self-consistent compile (dose 2): each layer writes only what the paragraph's mean shift still lacks
   ## Hypothesis 6510 chars 
   ## Exact recipe 11885 chars 
   ## Prediction (written before the run) 1644 chars 
record fcv-brief-exact-b20 ok
title (138 chars): Instruction pair vector: think briefly + be exact about what the issue asks, compiled self-consistently into one neuron per layer (dose 2)
   ## Hypothesis 7228 chars 
   ## Exact recipe 11103 chars 
   ## Prediction (written before the run) 1961 chars 
queue/fcv-brief-b20/record.md:67:base: champion generation 0
queue/fcv-brief-b20/record.md:71:size: whole network 0.6753% over 180 matrices, largest matrix 1.6391%
queue/fcv-brief-b20/record.md:73:control size: whole network 0.6753% over 180 matrices, largest matrix 1.6391%
queue/fcv-brief-b20/record.md:202:predicted_gain_120: -2 to 8
queue/fcv-brief-exact-b20/record.md:75:base: champion generation 0
queue/fcv-brief-exact-b20/record.md:79:size: whole network 0.6766% over 180 matrices, largest matrix 1.6391%
queue/fcv-brief-exact-b20/record.md:81:control size: whole network 0.6766% over 180 matrices, largest matrix 1.6391%
queue/fcv-brief-exact-b20/record.md:201:predicted_gain_120: -4 to 8
queue/fcv-brief-b20: build.sh ctrl edit record.md  queue/fcv-brief-b20/ctrl: adapter_config.json adapter_model.safetensors edit.json  queue/fcv-brief-b20/edit: adapter_config.json adapter_model.safetensors edit.json 
19M
queue/fcv-brief-exact-b20: build.sh ctrl edit record.md  queue/fcv-brief-exact-b20/ctrl: adapter_config.json adapter_model.safetensors edit.json  queue/fcv-brief-exact-b20/edit: adapter_config.json adapter_model.safetensors edit.json 
19M
OK edit: 180 modules changed (of 180 present), r=1, max |dW|_F/|W|_F 0.01639 at 32.gate_proj
OK ctrl: 180 modules changed (of 180 present), r=1, max |dW|_F/|W|_F 0.01639 at 32.gate_proj
0ec531af921856756d38
Sat Nov 14 01:29:24 AM EST 2026

10:edit      close all       n=  40  log p(close)   -0.036  top is close 1.000  log p(logged)   -0.036  top is logged 1.000
11:edit      close depth<200 n=  33  log p(close)   -0.032  top is close 1.000  log p(logged)   -0.032  top is logged 1.000
12:edit      close 200..799  n=   6  log p(close)   -0.060  top is close 1.000  log p(logged)   -0.060  top is logged 1.000
13:edit      close 800+      n=   1  log p(close)   -0.027  top is close 1.000  log p(logged)   -0.027  top is logged 1.000
Sat Nov 14 01:38:17 AM EST 2026
Sat Nov 14 01:35:57 AM EST 2026
Sat Nov 14 01:38:16 AM EST 2026
edit      close all       n=  40  log p(close)   -0.037  top is close 1.000  log p(logged)   -0.037  top is logged 1.000
edit      close depth<200 n=  33  log p(close)   -0.033  top is close 1.000  log p(logged)   -0.033  top is logged 1.000
edit      close 200..799  n=   6  log p(close)   -0.064  top is close 1.000  log p(logged)   -0.064  top is logged 1.000
edit      close 800+      n=   1  log p(close)   -0.022  top is close 1.000  log p(logged)   -0.022  top is logged 1.000
change of log p(close) at close: mean +0.033, raised at 0.825 of 40 positions
change of log p(close) at cont: mean -0.069, raised at 0.389 of 126 positions

record fcv-brief-b20 ok
title (134 chars): Think-briefly instruction vector, self-consistent compile (dose 2): each layer writes only what the paragraph's mean shift still lacks
   ## Hypothesis 6510 chars 
   ## Exact recipe 12000 chars (CUT at 12000)
   ## Prediction (written before the run) 1644 chars 
record fcv-brief-exact-b20 ok
title (138 chars): Instruction pair vector: think briefly + be exact about what the issue asks, compiled self-consistently into one neuron per layer (dose 2)
   ## Hypothesis 7228 chars 
   ## Exact recipe 11310 chars 
   ## Prediction (written before the run) 1961 chars 
queue/fcv-brief-b20/record.md:0
queue/fcv-brief-exact-b20/record.md:0
Sat Nov 14 01:38:36 AM EST 2026

record fcv-brief-b20 ok
title (134 chars): Think-briefly instruction vector, self-consistent compile (dose 2): each layer writes only what the paragraph's mean shift still lacks
   ## Hypothesis 6510 chars 
   ## Exact recipe 11614 chars 
   ## Prediction (written before the run) 1644 chars 
record fcv-brief-exact-b20 ok
title (138 chars): Instruction pair vector: think briefly + be exact about what the issue asks, compiled self-consistently into one neuron per layer (dose 2)
   ## Hypothesis 7228 chars 
   ## Exact recipe 11030 chars 
   ## Prediction (written before the run) 1961 chars 
queue/fcv-brief-b20/record.md:152:| **its control**: control of F20 | 8 / 1278 | 0.009 | 0.017 | 0 of 460 | -0
queue/fcv-brief-b20/record.md:160:- control of F20: cosine -0.04 -0.00 0.05 -0.03 0.02 0.06 0.04 0.04 0.10 0.1
queue/fcv-brief-exact-b20/record.md:153:| **its control**: control of G20 | 8 / 1278 | 0.011 | 0.017 | 0 of 46
queue/fcv-brief-exact-b20/record.md:160:- control of G20: cosine -0.04 0.03 0.07 -0.08 0.02 0.03 0.06 0.08 0.0
Sat Nov 14 01:38:50 AM EST 2026
[This command modified 1 file you've previously read: mkrec8.py. Call Read before editing.]

OK edit: 180 modules changed (of 180 present), r=1, max |dW|_F/|W|_F 0.01639 at 32.gate_proj
OK ctrl: 180 modules changed (of 180 present), r=1, max |dW|_F/|W|_F 0.01639 at 32.gate_proj
0ec531af92185675 49952aa0f3a03815 
title: Think-briefly instruction vector, self-consistent compile (dose 2): each layer writes only what the paragraph's mean shift still lacks

## Hypothesis
OK edit: 180 modules changed (of 180 present), r=1, max |dW|_F/|W|_F 0.01639 at 32.gate_proj
OK ctrl: 180 modules changed (of 180 present), r=1, max |dW|_F/|W|_F 0.01639 at 32.gate_proj
3e12574f9508b67b 1a6f06b500d0bcf7 
title: Instruction pair vector: think briefly + be exact about what the issue asks, compiled self-consistently into one neuron per layer (dose 2)

## Hypothesis
This entry asks one question: **is the gain of 0031 real?** It is the same instruction at the same dose, rebuilt
with an independent neuron in every layer and a corrected compile, scored in a new batch against new reference
arms. It is also the within-batch reference for its sister entry `fcv-brief-exact-b20` (this edit with a second paragraph compiled in, on being exact about what the issue asks: the difference
between the two entries in one batch is the effect of that paragraph).

The other compile. Turn 6 wrote `delta_L = d_L - layer_scalar_L d_(L-1)` (d_L = the paragraph's mean shift of
Sat Nov 14 01:39:07 AM EST 2026
READY
build.sh
ctrl
edit
record.md

queue/fcv-brief-exact-b20/record.md:60:(3) Nobody has measured whether I5 as a prompt gains tasks; I may not run rollouts.
## Prediction (written before the run)

predicted_gain_120: -4 to 8

My median is 2, one below the sister entry's 3: I5's shift is measured only in part (above), and a second paragraph
[line removed by the dashboard scrubber]
to the sister entry in the same batch: I predict this entry minus `fcv-brief-b20` between -5 and +6 solved tasks,
median 0 to +1. A real effect of I5 would have to show as +4 or more there to be told from the spread of two arms.

Behaviour that should move if I5 acts as written (from `task_results.jsonl` with `code/x_stats.py`, this entry
against the sister entry, and both against the mean of the two reference arms):
- finished episodes with a wrong patch, plus episodes that reach the budget with a wrong patch: 4 or more fewer
Sat Nov 14 01:39:26 AM EST 2026
READY build.sh ctrl edit record.md deep-gate-b01-end-neurons-x9 deep-thought-gate-g3-b03 early-gate-b01-end-neurons-x9 end-neurons-x25 fcv-brief-b20 fcv-brief-exact-b20 icv-brief-b20 icv-brief-b40 net-deep-gate-close-b05 net-deep-gate-close-b12 net-preedit-commit-b12 net-preedit-commit-close-b05

33	## Traps
34	- Runner: `predicted_gain_120: -6 to 6` (no "+"), title at most 140 characters, each record section cut at 12,000
35	  (`code/mkrec8.py` fills the templates in `code/rec_*.template.md` and checks with `runner.split_record`).
36	- Windows: no heredocs for code (backslashes halved, quotes break). Write tool, scp, `sed -i 's/\r$//'`.
37	- Lock: a tools-lane job held GPU 0 for two hours when the turn opened; queue `flock` at once. Holds this turn:
38	  13 + 16 + 5 minutes. `fcv_job.py` needs 13 minutes (35 calls x 4 passes, 8 evaluation calls x 9 arms); `--mock` 15 s.
39	
40	## Next
41	- Read the round: both entries against the references (`code/x_stats.py`, edit the arm names), and pair minus sister.
42	- 0031 passes stage 2: brevity is in the champion; no later entry may contain I1.
43	- Pair = sister: quality instructions do not compile into constants. Then try the context-dependent compile: k
44	  neurons per layer with the always-on gate (gate 8w, so the activation is about 8 (u . x)) are a rank-k linear map
45	  from the MLP input to the instruction's per-position shift; fit by reduced-rank ridge on thousands of positions.
46	- Missing tool that would have changed what I did: a sanctioned prompted reference arm (does I5, as a prompt, gain
47	  tasks at all?).

589
589 NOTES.md
- If 0031 is promoted first, both stack on it to I1 at dose 4: expect null or harm; then queue I5 alone.

## Learned
[line removed by the dashboard scrubber]
  (`results/wsp31/budget/s1_x124/summary.txt`): doubling it is worth +5 of 120. That is the ceiling of every
[line removed by the dashboard scrubber]
- Wrong patches fail the required tests and nothing else (125 of 136, usually one test), and the agent had tested
  after its last edit as often as in solved episodes: a "verify more" instruction is useless.
- My guess "half of 0031's stream shift is junk" was wrong. The self-consistent compile is the same edit at matched
  KL; the cosine of 0.4 to 0.7 with the prompted shift is call-to-call variation that no constant carries.
- Episodic instructions do not compile into a mean shift (I2, turn 6: reproduced 0). I5 compiled weakly: the pair
  reproduces 0.40 of the pair prompt, I1 alone 0.31; the part I5 adds is stable only in layers 6 to 24.

## Traps
- Runner: `predicted_gain_120: -6 to 6` (no "+"), title at most 140 characters, each record section cut at 12,000
Sat Nov 14 01:40:00 AM EST 2026
claude__fcv-brief-b20.json
claude__fcv-brief-exact-b20.json
44:| 0023 | claude | net-preedit-commit-close-b05 | parked | -8 to +9 |  |  |  | - |
49:| 0031 | claude | icv-brief-b20 | queued_s2 | -6 to +6 | +9.0 (77 vs 72) |  |  | - |
50:| 0032 | claude | icv-brief-b40 | queued_s2 | -9 to +6 | +3.0 (71 vs 72) |  |  | - |
53:| 0038 | claude | fcv-brief-b20 | queued_s1 | -2 to +8 |  |  |  | - |
61:| claude | 0038 | 0 | 5 | 2026-11-14 00:44 | open now |
[This command modified 1 file you've previously read: NOTES.md. Call Read before editing.]