Complete three-model dense coverage: Kimi prose dense finished under charitable limits; reliability reported as quantified cost; framed bottom-right legends on charts 1 and 6; updated card and report

README.md

@@ -89,8 +89,8 @@ and consistent answer to the questions this benchmark was built to settle.
 The full report is in `RESULTS.md`.
 
 - Token economy depends on content layout, not just length. Rendering
-  content as a dense image saved 28 to 47 per cent of input tokens on long
-  prose, but on short page-scale prompts it cost 1.6 to 2.6 times more than
+  content as a dense image saved 16 to 47 per cent of input tokens on long
+  prose, but on short page-scale prompts it cost 1.8 to 2.5 times more than
   the text, and on long pretty-printed machine text (JSON configuration,
   tool schemas, source code) it cost 1.3 to 2.5 times more, because
   line-oriented layout renders into mostly empty pages. The content that
@@ -114,9 +114,11 @@ The full report is in `RESULTS.md`.
   content left constraint-following and multi-step reasoning essentially
   intact. Reading exact figures for arithmetic is the exception, because a
   misread digit corrupts an otherwise correct calculation. Reasoning models
-  carry an extra reliability risk, since one model entered runaway
-  multi-minute reasoning on barely legible dense prose pages and could not
-  complete that slice at all.
+  carry an extra reliability cost, since one model needed a median of 87
+  seconds per call on long dense documents against 9 and 15 for the other
+  two, with a worst case of sixty-four minutes for a single call, and it
+  completed the slice only after its timeout and output budget were raised
+  well beyond ordinary settings.
 - Injection resistance held for all three current models even when the
   adversarial text was imaged, although an older model tested during
   development was susceptible under the combined-image condition.

RESULTS.md

@@ -41,20 +41,20 @@ independent models is the main reason for confidence in the conclusions.
 Whether image delivery saves tokens depends entirely on how much text is
 packed into each image. On short, page-scale documents at the regular density
 the image is wasteful, because a single page carries little text yet still
-incurs a fixed vision-token cost. Across the three models the content-as-image
-condition used between 1.6 and 2.6 times as many input tokens as the plain
-text condition.
+incurs a fixed vision-token cost. Across the three models on the page-scale
+tiers of the core subset, the content-as-image condition used between 1.8
+and 2.5 times as many input tokens as the plain text condition.
 
 The saving appears only when documents are long, the rendering is dense, and
 the content is prose. On the long prose slice at the dense profile the
-picture reverses, and the image becomes the cheaper channel. Kimi K2.6 is
-absent from this table because its long prose runs did not complete, for
-reasons described under question four.
+picture reverses, and the image becomes the cheaper channel, although how
+much cheaper varies threefold with each provider's image token accounting.
 
 | model | text tokens | image tokens | image saving |
 |---|---|---|---|
 | GPT 5.5 | 4189 | 3020 | 28 per cent |
 | Gemini 3.5 Flash | 4819 | 2554 | 47 per cent |
+| Kimi K2.6 | 3935 | 3288 | 16 per cent |
 
 On structured machine text the saving does not merely shrink, it inverts.
 The same dense profile applied to long JSON configuration, tool schemas and
@@ -300,17 +300,26 @@ Flash could not reconcile a hundred-line invoice even from plain text, so for
 that model the image result is confounded by a limitation that has nothing to
 do with the modality.
 
-There is also a reliability finding specific to reasoning models. Kimi K2.6
-could not complete the long prose documents at the dense profile at all.
-Its reasoning chains on those pages ran to as much as eighteen minutes for a
-single call, eight calls exhausted the twelve-thousand-token output budget,
-and five of the fifty-six calls returned nothing through eight retries each,
-so the slice was abandoned rather than forced. The same model completed the
-equivalent structured dense slice, two hundred calls, without a single
-error. A hard-to-read page does not merely lower a reasoning model's
-accuracy, it can send the model into unbounded rumination that costs output
-tokens, wall-clock time and reliability. Any production deployment of dense
-rendering in front of a reasoning model should budget for this failure mode.
+There is also a reliability finding specific to reasoning models, and it is
+best stated as a cost rather than an inability. Under ordinary settings, a
+five-minute timeout and a twelve-thousand-token output budget, Kimi K2.6
+could not finish the long dense documents. Eight calls exhausted the output
+budget and five calls, spread across the instruction, calculation and
+transcription families and including two with plain text and no image at
+all, returned nothing through eight retries each. Given charitable settings
+instead, an hour per call and a thirty-two-thousand-token budget, the same
+model completed all fifty-six calls without a single error. What the
+charity revealed is the price. Its median latency on this slice was 87
+seconds against 9 for GPT 5.5 and 15 for Gemini 3.5 Flash, one in ten calls
+needed over eight minutes, the slowest needed sixty-four minutes, and one
+answer consumed thirty thousand output tokens. The same model completed the
+structured dense slice, two hundred calls, at ordinary settings without
+error, so the trigger is the long document rather than the image as such,
+although the image condition lengthens the rumination further. A
+hard-to-read page does not merely lower a reasoning model's accuracy, it
+can multiply its thinking time and output spend by an order of magnitude,
+and any production deployment of dense rendering in front of a reasoning
+model should budget for that, in money, in latency and in timeout policy.
 
 ## A note on safety
 
@@ -421,8 +430,8 @@ can be imaged for a real saving. Machine text in its natural layout should
 stay as text, where it is both cheaper and safer. Anything that must be
 recovered verbatim from prose should never cross the image channel densely
 rendered, and reasoning models add a further reliability caveat, since a
-barely legible page can push them into unbounded rumination rather than a
-clean failure.
+long dense page can multiply their thinking time and output spend by an
+order of magnitude.
 
 ## Caveats
 
@@ -433,9 +442,10 @@ statistical significance at that scale. The dense long-document slices are
 far smaller, twenty-eight paired prose instances and one hundred paired
 structured instances per model, so their percentages are directional and
 the economy figures in particular should be reproduced at the full per-cell
-count of the released dataset before being quoted as precise. Kimi K2.6 has
-no completed prose dense results, for the reliability reason given under
-question four, so its dense figures cover only the structured slice.
+count of the released dataset before being quoted as precise. Eight of Kimi
+K2.6's earlier prose dense answers were truncated at a twelve-thousand-token
+output budget before the limits were raised, and those stand in its scores
+as recorded, so its prose dense accuracy should be read as a floor.
 Cross-provider comparisons conflate the model with each provider's opaque
 image preprocessing and with its token accounting, which differ visibly in
 the economy tables, so the cleanest readings are the within-model deltas