Name the benchmarked build: DeepSeek V4 Flash 0731

The label said "DeepSeek V4 Flash", which does not identify a fixed model.
OpenRouter also serves ~deepseek/deepseek-v4-flash-latest, an alias that
routes to whatever is newest in the family, so a published score under the
bare name comes to describe something we never ran. The label now matches
OpenRouter's own name for the build on file, and every run in the dataset
is deepseek-v4-flash-0731, so nothing about the measurement changes.

The longer label broke the build. angled_labels_overhanging() caught it in
the cost chart, which sorts by price and therefore puts the cheapest model
in column 0, where a leaning label has nowhere to sweep: 106.0px of spill
against 97.0px of clearance. AXIS_PAD_L goes 80 -> 98, which clears it by
8.5px and keeps the working slack every previous raise aimed for, at about
18px of plot width on the three charts that share the constant.

The geometry comment is rewritten with the new arithmetic, since it reasoned
in detail from numbers this change invalidates, and now records that a
rename can break the build, not just a wider cut. AGENTS.md picks up the new
value and a warning that pad_l=AXIS_PAD_L is a default argument bound at
import, so setting the module attribute at runtime measures the old pad and
silently proves nothing.

NO_PLAIN_Q4 in audit.py is keyed by label and moves with it.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: barretodavid

AGENTS.md

@@ -328,12 +328,18 @@ Two helpers keep charts consistent rather than gating them:
   rather than writing a regex: axis labels come in two shapes, a bare rotated `<text>`
   and a rotated `<g>` wrapping the label plus a rank-delta arrow, and a parser that knows
   only the first reports the MCP chart drawing four fewer models than it drew.
-- `LABEL_ANGLE = 55` and `AXIS_PAD_L = 80` are shared across charts deliberately, so
+- `LABEL_ANGLE = 55` and `AXIS_PAD_L = 98` are shared across charts deliberately, so
   they stay visually consistent. Change them in one place or not at all. `AXIS_PAD_L`
-  tracks `CHART_TOP_N`: more columns are narrower columns, which pulls column 0's centre
-  left and eats the clearance an angled label needs. The one documented exception is
-  `sci_bar_chart`, which keeps `pad_l = 64` because it alone draws in rank order, so its
-  first label is always the leader's and always short.
+  tracks `CHART_TOP_N` **and the longest label**: more columns are narrower columns, which
+  pulls column 0's centre left and eats the clearance an angled label needs, and a longer
+  name spills further into it. Renaming a model can therefore break the build, which is
+  what raised this from 80 (see the arithmetic above the constant). The one documented
+  exception is `sci_bar_chart`, which keeps `pad_l = 64` because it alone draws in rank
+  order, so its first label is always the leader's and always short.
+- **`pad_l=AXIS_PAD_L` is a default argument**, bound when the module is imported. Setting
+  `html_report.AXIS_PAD_L` at runtime changes nothing, so a dry run that tries to test a
+  new pad that way silently measures the old one. Patch `fn.__defaults__`, or just edit the
+  constant.
 
 For anything visual, run `bench.screenshot` and **look at the PNGs** before committing.
 Text and geometry checks have passed things that were plainly ugly on screen. Every

bench/audit.py

@@ -455,7 +455,7 @@ def _gguf(r):
 # at UD-Q4_K_XL and would have sent the class rule to the estimate while a repo
 # sat in the snapshot looking authoritative, when bartowski publishes the plain
 # file. A new name appearing here should be investigated, not added.
-NO_PLAIN_Q4 = {"Kimi K3", "DeepSeek V4 Flash"}
+NO_PLAIN_Q4 = {"Kimi K3", "DeepSeek V4 Flash 0731"}
 
 estimated = sorted(r["label"] for r in lb if r["open_weight"] and not r.get("vram_measured"))
 no_size = sorted({r["label"] for r in lb if r["open_weight"] and not _gguf(r).get("repo")}

bench/html_report.py

@@ -180,17 +180,25 @@ def line_chart(x_labels, vals, win_i, color, w=380, h=200, y_min=0, y_max=100,
 #
 # assert_output_is_portable enforces this geometry, so a future name long enough
 # to break it fails the build instead of rendering clipped.
-# Raised twice as the cut moved, both times for the same label. At the top
-# twelve, pad 64 left "DeepSeek V4 Flash (off)" (152px) reaching 87.1px against
-# 87px of clearance and the build guard caught it at 0.1px; 72 bought an 8px
-# margin. The top fifteen narrows every column from 54 to 43px, which pulls
-# column 0's centre 5.4px left and spends most of that margin again, so 80. A
-# pixel of pad is worth 0.97px of clearance, not 1: the plot loses the width, so
-# the columns narrow slightly too. The same worst case now clears by 13.9px, and
-# the angle stays put, being already steep enough that every extra degree costs
+#
+# Raised three times, every time for the same label, because the cost chart sorts
+# by price and this model is the cheapest, so it holds column 0 where a leaning
+# label has nowhere to sweep. A pixel of pad is worth 0.97px of clearance, not 1:
+# the plot loses the width, so the columns narrow slightly too.
+#   top twelve, pad 64:  "DeepSeek V4 Flash (off)" 152px spills 87.1 vs 87.0, and
+#                        the guard caught it at 0.1px; 72 bought an 8px margin
+#   top fifteen, pad 80: columns narrow 54 -> 43px, pulling column 0's centre
+#                        5.4px left and spending most of that margin again
+#   dated label, pad 98: "DeepSeek V4 Flash 0731 (off)" is 184.8px and spills
+#                        106.0 against 114.5 of clearance, so it clears by 8.5px
+# The date is in the label because the undated name does not identify a fixed
+# model (see the registry entry), and 98 is chosen over the bare minimum of 93 to
+# keep the ~8px of working slack every raise above has aimed for.
+#
+# The angle stays put: it is already steep enough that every extra degree costs
 # more label depth than it buys.
 LABEL_ANGLE = 55
-AXIS_PAD_L = 80
+AXIS_PAD_L = 98
 _LABEL_CH = 11 * 0.60      # px per character at font-size 11, measured exact
                            # against getComputedTextLength (scratchpad/measure_labels.py)
 

bench/sci.py

@@ -327,12 +327,18 @@ def run_cost(run):
                "deepseek/deepseek-v4-pro@minimal", "deepseek/deepseek-v4-pro@disabled"],
      "label": "DeepSeek V4-Pro", "lab": "DeepSeek", "open_weight": True},
     # the 0731 release, not the preview: DeepSeek calls it the official one, and
-    # the two expose different ladders (preview xhigh/high, release max/high/low)
+    # the two expose different ladders (preview xhigh/high, release max/high/low).
+    #
+    # The label carries the date because the bare name does not identify anything
+    # stable. OpenRouter also serves ~deepseek/deepseek-v4-flash-latest, an alias
+    # that routes to whatever is newest in the family, so a published
+    # "DeepSeek V4 Flash" score silently comes to describe a model we never ran.
+    # This label matches OpenRouter's own name for the build we measured.
     {"specs": ["deepseek/deepseek-v4-flash-0731@max",
                "deepseek/deepseek-v4-flash-0731@high",
                "deepseek/deepseek-v4-flash-0731@low",
                "deepseek/deepseek-v4-flash-0731@disabled"],
-     "label": "DeepSeek V4 Flash", "lab": "DeepSeek", "open_weight": True},
+     "label": "DeepSeek V4 Flash 0731", "lab": "DeepSeek", "open_weight": True},
     {"specs": ["tencent/hy3", "tencent/hy3@max", "tencent/hy3@xhigh", "tencent/hy3@high",
                "tencent/hy3@medium", "tencent/hy3@low", "tencent/hy3@minimal",
                "tencent/hy3@disabled"],

bench/snapshot_meta.py

@@ -71,8 +71,8 @@ def gguf_sizes(repo, expect_base):
     `files` is the whole ladder under the exact names the repo publishes, because
     "the best quant that fits a 64 GB machine" can only be answered from the
     ladder that exists: these repos ship 13 to 38 files each, from TQ1_0 to BF16,
-    and the five-name subset we used to record made DeepSeek V4 Flash look like
-    it topped out at IQ4_XS when the repo goes to UD-Q8_K_XL.
+    and the five-name subset we used to record made DeepSeek V4 Flash 0731 look
+    like it topped out at IQ4_XS when the repo goes to UD-Q8_K_XL.
 
     The five canonical keys stay alongside it, derived from `files` with the same
     plain-over-UD preference as before, so the class rule and the open-weights