{
 "measurement": "Whether six free duck.ai models keep the same rendered JSON format a day later",
 "trial_date": "2026-08-22",
 "service": "duck.ai, free tier, no account, no payment, no CAPTCHA",
 "this_is_a_replication": {
  "of": "a twelve-run trial on 2026-08-21 that found an exact three-three split",
  "declared_before_re_running": true,
  "preregistration": ".measurements/model-trials/json-render-replication-preregistration.md",
  "why_the_first_one_was_never_published": "its evidence chain was not preserved: no screenshot was committed, so which model produced which output rested on notes. That is this measurement's whole subject, so it was held.",
  "prior_reproduced_for": [
   "GPT-5.6 Luna",
   "GPT-5.4 mini",
   "gpt-oss 120B",
   "Claude Haiku 4.5",
   "Mistral Small 4",
   "Gemma 4 31B"
  ],
  "prior_reproduced_count": "6 of 6",
  "THE_PRIOR_IS_NOT_PUBLISHED_AND_THIS_IS_NOT_CHECKABLE": "The 2026-08-21 trial's runs were never published, precisely because their evidence chain was not preserved. So a reader CANNOT verify '6 of 6 reproduced'; it rests on this operator's record of a trial that was withheld for being insufficiently evidenced. It is stated here and in the post as an unverifiable claim rather than as a receipt. Everything else in this corpus stands on the twelve runs published here, which do not need the prior at all: the three-three split and the per-model stability across two rounds are established by this trial alone."
 },
 "the_limit_that_shapes_everything": "duck.ai never exposes the model's raw bytes. The page gives only the rendering, so the observable is how duck.ai RENDERED the reply. A rendered code block is strong evidence the model emitted a markdown fence, because that is what duck.ai renders fences as. It is not the same as reading the bytes.",
 "method": {
  "prompt": "Return a JSON object describing the planet Mars with exactly these three keys: name, moons, diameter_km. Values: a string, an integer, an integer. Output only the JSON object and nothing else.",
  "rounds_per_model": 2,
  "new_chat_per_run": true,
  "observable": "presence of a rendered code-block element in the assistant message",
  "estimator": "none. Every cell is a pair of binary readings and nothing is averaged. Stability is reported only where both rounds agree, asserted in the assembler rather than eyeballed.",
  "content_recorded_not_graded": "Mars has two moons and published diameters differ by rounding, so grading content would measure something other than the instruction."
 },
 "controls": {
  "capture_versions_used": "v1 on the first two runs, v2 on the other ten. v1 walked up from a text leaf and is correct for a plain reply; it breaks on a code block because syntax highlighting splits the JSON into token-level spans. Both v1 runs are plain-text cells and their label and reply were cross-checked against their screenshots. The version is recorded per run.",
  "committed_screenshot_per_run": "12 of 12",
  "atomic_capture": "label, reply text and code-block presence are read in ONE page evaluation, because a previous trial found the page-text reader could return a stale 'Generating response' while the screenshot already showed the finished reply.",
  "label_tied_structurally_to_its_reply": "the label is taken by walking up from the reply to the message container that holds both, not by position on the page.",
  "picked_model_equals_reported_model": "12 of 12",
  "runs_captured_mid_stream": 0,
  "screenshots_whose_label_is_legible": "9 of 12",
  "screenshots_showing_no_model_name": [
   "/shots/json-render-2026-08-22/mistral-small-4-r0.jpg",
   "/shots/json-render-2026-08-22/mistral-small-4-r1.jpg",
   "/shots/json-render-2026-08-22/gpt-oss-120b-r1.jpg"
  ],
  "how_legibility_was_established": "By looking at all twelve header bands together in one montage, published at /shots/json-render-2026-08-22/header-bands.png, rather than by judging each run at capture time, which produced an inconsistent per-run record. The darkest pixel in the band right of the vendor icon is published per run beside it: the three unnamed screenshots are the three faintest, 243, 243 and 255 against 239 and below for every legible one, so the reading and the measurement agree."
 },
 "results": {
  "split": {
   "code_block": 3,
   "plain_text": 3,
   "code_block_models": [
    "Claude Haiku 4.5",
    "Mistral Small 4",
    "Gemma 4 31B"
   ],
   "plain_text_models": [
    "GPT-5.6 Luna",
    "GPT-5.4 mini",
    "gpt-oss 120B"
   ],
   "exact_three_three": true
  },
  "models_whose_grade_changed_between_rounds": null,
  "models_whose_exact_text_changed_between_rounds": [
   "gpt-oss 120B"
  ],
  "content": {
   "distinct_objects_whitespace_stripped": 1,
   "the_object": "{\"name\":\"Mars\",\"moons\":2,\"diameter_km\":6779}",
   "errors": 0
  },
  "cells": [
   {
    "model": "GPT-5.6 Luna",
    "vendor": "OpenAI",
    "round_0": "plain text",
    "round_1": "plain text",
    "grade_stable_across_rounds": true,
    "exact_text_identical_across_rounds": true,
    "reply_chars": [
     44,
     44
    ],
    "code_fence_element_present": [
     false,
     false
    ],
    "language_tag": [
     null,
     null
    ],
    "capture_version": [
     "v1",
     "v2"
    ],
    "banner": "Zero data retention for this chat. No AI training.",
    "screenshots": [
     "/shots/json-render-2026-08-22/gpt-5.6-luna-r0.jpg",
     "/shots/json-render-2026-08-22/gpt-5.6-luna-r1.jpg"
    ]
   },
   {
    "model": "GPT-5.4 mini",
    "vendor": "OpenAI",
    "round_0": "plain text",
    "round_1": "plain text",
    "grade_stable_across_rounds": true,
    "exact_text_identical_across_rounds": true,
    "reply_chars": [
     44,
     44
    ],
    "code_fence_element_present": [
     false,
     false
    ],
    "language_tag": [
     null,
     null
    ],
    "capture_version": [
     "v1",
     "v2"
    ],
    "banner": "Zero data retention for this chat. No AI training.",
    "screenshots": [
     "/shots/json-render-2026-08-22/gpt-5.4-mini-r0.jpg",
     "/shots/json-render-2026-08-22/gpt-5.4-mini-r1.jpg"
    ]
   },
   {
    "model": "gpt-oss 120B",
    "vendor": "OpenAI, open weights",
    "round_0": "plain text",
    "round_1": "plain text",
    "grade_stable_across_rounds": true,
    "exact_text_identical_across_rounds": false,
    "reply_chars": [
     44,
     51
    ],
    "code_fence_element_present": [
     false,
     false
    ],
    "language_tag": [
     null,
     null
    ],
    "capture_version": [
     "v2",
     "v2"
    ],
    "banner": "Zero provider visibility. No AI training.",
    "screenshots": [
     "/shots/json-render-2026-08-22/gpt-oss-120b-r0.jpg",
     "/shots/json-render-2026-08-22/gpt-oss-120b-r1.jpg"
    ]
   },
   {
    "model": "Claude Haiku 4.5",
    "vendor": "Anthropic",
    "round_0": "code block",
    "round_1": "code block",
    "grade_stable_across_rounds": true,
    "exact_text_identical_across_rounds": true,
    "reply_chars": [
     57,
     57
    ],
    "code_fence_element_present": [
     true,
     true
    ],
    "language_tag": [
     "json",
     "json"
    ],
    "capture_version": [
     "v2",
     "v2"
    ],
    "banner": "Limited data retention for this chat. No AI training.",
    "screenshots": [
     "/shots/json-render-2026-08-22/claude-haiku-4.5-r0.jpg",
     "/shots/json-render-2026-08-22/claude-haiku-4.5-r1.jpg"
    ]
   },
   {
    "model": "Mistral Small 4",
    "vendor": "Mistral",
    "round_0": "code block",
    "round_1": "code block",
    "grade_stable_across_rounds": true,
    "exact_text_identical_across_rounds": true,
    "reply_chars": [
     57,
     57
    ],
    "code_fence_element_present": [
     true,
     true
    ],
    "language_tag": [
     "json",
     "json"
    ],
    "capture_version": [
     "v2",
     "v2"
    ],
    "banner": "Zero data retention for this chat. No AI training.",
    "screenshots": [
     "/shots/json-render-2026-08-22/mistral-small-4-r0.jpg",
     "/shots/json-render-2026-08-22/mistral-small-4-r1.jpg"
    ]
   },
   {
    "model": "Gemma 4 31B",
    "vendor": "Google",
    "round_0": "code block",
    "round_1": "code block",
    "grade_stable_across_rounds": true,
    "exact_text_identical_across_rounds": true,
    "reply_chars": [
     57,
     57
    ],
    "code_fence_element_present": [
     true,
     true
    ],
    "language_tag": [
     "json",
     "json"
    ],
    "capture_version": [
     "v2",
     "v2"
    ],
    "banner": "Zero provider visibility. No AI training.",
    "screenshots": [
     "/shots/json-render-2026-08-22/gemma-4-31b-r0.jpg",
     "/shots/json-render-2026-08-22/gemma-4-31b-r1.jpg"
    ]
   }
  ],
  "the_second_axis_and_why_it_is_NOT_a_clean_correlation": {
   "what_round_0_alone_suggested": "every plain-text reply was compact at 44 characters and every code-block reply was pretty-printed at 57, a perfect correlation.",
   "what_round_1_did_to_it": "gpt-oss 120B returned PRETTY-PRINTED multi-line JSON at 51 characters with NO code fence, so fencing and pretty-printing came apart.",
   "reading": "the GRADE is stable for all six models across both rounds; the exact SHAPE is not, for one of them. Reporting the round-0 correlation as a finding would have been an artefact of a single round."
  }
 },
 "known_limits": {
  "the_bytes_were_never_seen": "Any claim about what a parser would do to these models' raw output is outside what this observed.",
  "n_is_two_per_model": "Two rounds on one day. Enough to show each model's rendered grade was stable across two attempts and no more. No rate is claimed and no percentage appears anywhere.",
  "one_prompt_one_tiny_schema": "Three flat keys, no nesting, no arrays. A larger schema is untested.",
  "free_tier_one_ui": "The same models through their vendors' own APIs, or with a structured-output mode on, are a different setting and nothing here transfers.",
  "one_format_instruction": "It says nothing about reasoning, accuracy, or any of the reasons you would actually choose between these models.",
  "three_screenshots_do_not_name_their_model": "GENERATED, not typed, because the typed version of this limit named the wrong models. gpt-oss-120b-r1.jpg, mistral-small-4-r0.jpg, mistral-small-4-r1.jpg render the vendor icon and no model name into the published image. For those runs the model identity rests solely on the atomic DOM capture. One of them, gpt-oss-120b-r1.jpg, is the run carrying the shape deviation, so the weakest link in the evidence chain sits under a headline observation and is named here for that reason.",
  "copy_control_not_tested": "Every code block carries a copy control. Whether it strips the fence was not tested."
 },
 "runs": [
  {
   "round": 0,
   "picked": "GPT-5.6 Luna",
   "label_from_dom": "GPT-5.6 Luna",
   "reply": "{\"name\":\"Mars\",\"moons\":2,\"diameter_km\":6779}",
   "inCode": false,
   "pre_in_container": 0,
   "code_in_container": 0,
   "pre_on_page": 0,
   "reply_chars": 44,
   "still_generating": false,
   "captured_at": "2026-08-22T02:52:17.870Z",
   "screenshot": "/shots/json-render-2026-08-22/gpt-5.6-luna-r0.jpg",
   "rendered": "plain text",
   "capture_version": "v1",
   "capture_note": "v1 walks up from a leaf containing \"diameter_km\". Correct for a PLAIN reply, and it breaks on a code block because syntax highlighting splits the JSON into token-level spans. Both v1 runs are plain-text cells and their label and reply were cross-checked against their screenshots.",
   "banner": "Zero data retention for this chat. No AI training.",
   "banner_provenance": "Not captured by the harness on this run; the field was added from run 3 onward. Read off this run's COMMITTED screenshot, shots-2026-08-22/gpt-5.6-luna-r0.jpg, which shows the line in full. Recorded as read-from-artefact rather than harness-captured.",
   "screenshot_label_legible": true,
   "screenshot_label_method": "Read by eye from a montage of all twelve header bands, published at /shots/json-render-2026-08-22/header-bands.png, with the darkest pixel in the band recorded beside it. Not a per-run judgement made at capture time, which was inconsistent.",
   "label_cross_check": "DOM label matches the label visible in the screenshot",
   "screenshot_label_min_ink_0_255": 0,
   "screenshot_label_ink_note": "Darkest pixel in the 260x26 band right of the vendor icon, so the icon cannot contribute. 255 means the band is pure white, i.e. no label was rendered into the published image at all."
  },
  {
   "round": 0,
   "picked": "GPT-5.4 mini",
   "label_from_dom": "GPT-5.4 mini",
   "reply": "{\"name\":\"Mars\",\"moons\":2,\"diameter_km\":6779}",
   "inCode": false,
   "pre_in_container": 0,
   "code_in_container": 0,
   "pre_on_page": 0,
   "reply_chars": 44,
   "still_generating": false,
   "captured_at": "2026-08-22T02:53:33.077Z",
   "screenshot": "/shots/json-render-2026-08-22/gpt-5.4-mini-r0.jpg",
   "rendered": "plain text",
   "capture_version": "v1",
   "capture_note": "v1 walks up from a leaf containing \"diameter_km\". Correct for a PLAIN reply, and it breaks on a code block because syntax highlighting splits the JSON into token-level spans. Both v1 runs are plain-text cells and their label and reply were cross-checked against their screenshots.",
   "banner": "Zero data retention for this chat. No AI training.",
   "banner_provenance": "Not captured by the harness on this run; the field was added from run 3 onward. Read off this run's COMMITTED screenshot, shots-2026-08-22/gpt-5.4-mini-r0.jpg, which shows the line in full. Recorded as read-from-artefact rather than harness-captured.",
   "screenshot_label_legible": true,
   "screenshot_label_method": "Read by eye from a montage of all twelve header bands, published at /shots/json-render-2026-08-22/header-bands.png, with the darkest pixel in the band recorded beside it. Not a per-run judgement made at capture time, which was inconsistent.",
   "label_cross_check": "DOM label matches the label visible in the screenshot",
   "screenshot_label_min_ink_0_255": 195,
   "screenshot_label_ink_note": "Darkest pixel in the 260x26 band right of the vendor icon, so the icon cannot contribute. 255 means the band is pure white, i.e. no label was rendered into the published image at all."
  },
  {
   "round": 0,
   "picked": "Claude Haiku 4.5",
   "label_from_dom": "Claude Haiku 4.5",
   "reply": "{\n  \"name\": \"Mars\",\n  \"moons\": 2,\n  \"diameter_km\": 6779\n}",
   "inCode": true,
   "pre_in_message": 1,
   "code_in_message": 1,
   "lang_tag": "json",
   "reply_chars": 57,
   "still_generating": false,
   "captured_at": "2026-08-22T02:55:08.716Z",
   "screenshot": "/shots/json-render-2026-08-22/claude-haiku-4.5-r0.jpg",
   "rendered": "code block",
   "capture_version": "v2",
   "banner": "Limited data retention for this chat. No AI training.",
   "screenshot_label_legible": true,
   "screenshot_label_method": "Read by eye from a montage of all twelve header bands, published at /shots/json-render-2026-08-22/header-bands.png, with the darkest pixel in the band recorded beside it. Not a per-run judgement made at capture time, which was inconsistent.",
   "label_cross_check": "DOM label matches the label visible in the screenshot",
   "screenshot_label_min_ink_0_255": 196,
   "screenshot_label_ink_note": "Darkest pixel in the 260x26 band right of the vendor icon, so the icon cannot contribute. 255 means the band is pure white, i.e. no label was rendered into the published image at all."
  },
  {
   "round": 0,
   "picked": "Mistral Small 4",
   "label_from_dom": "Mistral Small 4",
   "reply": "{\n  \"name\": \"Mars\",\n  \"moons\": 2,\n  \"diameter_km\": 6779\n}",
   "inCode": true,
   "pre_in_message": 1,
   "code_in_message": 1,
   "lang_tag": "json",
   "reply_chars": 57,
   "still_generating": false,
   "captured_at": "2026-08-22T02:56:37.356Z",
   "screenshot": "/shots/json-render-2026-08-22/mistral-small-4-r0.jpg",
   "rendered": "code block",
   "capture_version": "v2",
   "banner": "Zero data retention for this chat. No AI training.",
   "screenshot_label_legible": false,
   "screenshot_label_method": "Read by eye from a montage of all twelve header bands, published at /shots/json-render-2026-08-22/header-bands.png, with the darkest pixel in the band recorded beside it. Not a per-run judgement made at capture time, which was inconsistent.",
   "label_cross_check": "NOT CHECKABLE: the published screenshot shows the vendor icon and no model name, so there is nothing to cross-check against. For this run the model identity rests solely on the atomic DOM capture, which read the label and the reply in one evaluation from the container holding both.",
   "screenshot_label_min_ink_0_255": 243,
   "screenshot_label_ink_note": "Darkest pixel in the 260x26 band right of the vendor icon, so the icon cannot contribute. 255 means the band is pure white, i.e. no label was rendered into the published image at all."
  },
  {
   "round": 0,
   "picked": "gpt-oss 120B",
   "label_from_dom": "gpt-oss 120B",
   "reply": "{\"name\":\"Mars\",\"moons\":2,\"diameter_km\":6779}",
   "inCode": false,
   "pre_in_message": 0,
   "code_in_message": 0,
   "lang_tag": null,
   "reply_chars": 44,
   "still_generating": false,
   "captured_at": "2026-08-22T02:58:07.731Z",
   "screenshot": "/shots/json-render-2026-08-22/gpt-oss-120b-r0.jpg",
   "rendered": "plain text",
   "capture_version": "v2",
   "banner": "Zero provider visibility. No AI training.",
   "note": "On 2026-08-21 this model interposed an \"Enable web search for this chat?\" card before answering. It did NOT do so on this run.",
   "screenshot_label_legible": true,
   "screenshot_label_method": "Read by eye from a montage of all twelve header bands, published at /shots/json-render-2026-08-22/header-bands.png, with the darkest pixel in the band recorded beside it. Not a per-run judgement made at capture time, which was inconsistent.",
   "label_cross_check": "DOM label matches the label visible in the screenshot",
   "screenshot_label_min_ink_0_255": 215,
   "screenshot_label_ink_note": "Darkest pixel in the 260x26 band right of the vendor icon, so the icon cannot contribute. 255 means the band is pure white, i.e. no label was rendered into the published image at all."
  },
  {
   "round": 0,
   "picked": "Gemma 4 31B",
   "label_from_dom": "Gemma 4 31B",
   "reply": "{\n  \"name\": \"Mars\",\n  \"moons\": 2,\n  \"diameter_km\": 6779\n}",
   "inCode": true,
   "pre_in_message": 1,
   "code_in_message": 1,
   "lang_tag": "json",
   "reply_chars": 57,
   "still_generating": false,
   "captured_at": "2026-08-22T02:59:11.989Z",
   "screenshot": "/shots/json-render-2026-08-22/gemma-4-31b-r0.jpg",
   "rendered": "code block",
   "capture_version": "v2",
   "banner": "Zero provider visibility. No AI training.",
   "screenshot_label_legible": true,
   "screenshot_label_method": "Read by eye from a montage of all twelve header bands, published at /shots/json-render-2026-08-22/header-bands.png, with the darkest pixel in the band recorded beside it. Not a per-run judgement made at capture time, which was inconsistent.",
   "label_cross_check": "DOM label matches the label visible in the screenshot",
   "screenshot_label_min_ink_0_255": 198,
   "screenshot_label_ink_note": "Darkest pixel in the 260x26 band right of the vendor icon, so the icon cannot contribute. 255 means the band is pure white, i.e. no label was rendered into the published image at all."
  },
  {
   "round": 1,
   "picked": "GPT-5.6 Luna",
   "label_from_dom": "GPT-5.6 Luna",
   "reply": "{\"name\":\"Mars\",\"moons\":2,\"diameter_km\":6779}",
   "inCode": false,
   "pre_in_message": 0,
   "code_in_message": 0,
   "lang_tag": null,
   "reply_chars": 44,
   "still_generating": false,
   "captured_at": "2026-08-22T03:00:32.882Z",
   "screenshot": "/shots/json-render-2026-08-22/gpt-5.6-luna-r1.jpg",
   "rendered": "plain text",
   "capture_version": "v2",
   "banner": "Zero data retention for this chat. No AI training.",
   "screenshot_label_legible": true,
   "screenshot_label_method": "Read by eye from a montage of all twelve header bands, published at /shots/json-render-2026-08-22/header-bands.png, with the darkest pixel in the band recorded beside it. Not a per-run judgement made at capture time, which was inconsistent.",
   "label_cross_check": "DOM label matches the label visible in the screenshot",
   "screenshot_label_min_ink_0_255": 214,
   "screenshot_label_ink_note": "Darkest pixel in the 260x26 band right of the vendor icon, so the icon cannot contribute. 255 means the band is pure white, i.e. no label was rendered into the published image at all."
  },
  {
   "round": 1,
   "picked": "GPT-5.4 mini",
   "label_from_dom": "GPT-5.4 mini",
   "reply": "{\"name\":\"Mars\",\"moons\":2,\"diameter_km\":6779}",
   "inCode": false,
   "pre_in_message": 0,
   "code_in_message": 0,
   "lang_tag": null,
   "reply_chars": 44,
   "still_generating": false,
   "captured_at": "2026-08-22T03:01:42.872Z",
   "screenshot": "/shots/json-render-2026-08-22/gpt-5.4-mini-r1.jpg",
   "rendered": "plain text",
   "capture_version": "v2",
   "banner": "Zero data retention for this chat. No AI training.",
   "screenshot_label_legible": true,
   "screenshot_label_method": "Read by eye from a montage of all twelve header bands, published at /shots/json-render-2026-08-22/header-bands.png, with the darkest pixel in the band recorded beside it. Not a per-run judgement made at capture time, which was inconsistent.",
   "label_cross_check": "DOM label matches the label visible in the screenshot",
   "screenshot_label_min_ink_0_255": 193,
   "screenshot_label_ink_note": "Darkest pixel in the 260x26 band right of the vendor icon, so the icon cannot contribute. 255 means the band is pure white, i.e. no label was rendered into the published image at all."
  },
  {
   "round": 1,
   "picked": "Claude Haiku 4.5",
   "label_from_dom": "Claude Haiku 4.5",
   "reply": "{\n  \"name\": \"Mars\",\n  \"moons\": 2,\n  \"diameter_km\": 6779\n}",
   "inCode": true,
   "pre_in_message": 1,
   "code_in_message": 1,
   "lang_tag": "json",
   "reply_chars": 57,
   "still_generating": false,
   "captured_at": "2026-08-22T03:02:44.218Z",
   "screenshot": "/shots/json-render-2026-08-22/claude-haiku-4.5-r1.jpg",
   "rendered": "code block",
   "capture_version": "v2",
   "banner": "Limited data retention for this chat. No AI training.",
   "screenshot_label_legible": true,
   "screenshot_label_method": "Read by eye from a montage of all twelve header bands, published at /shots/json-render-2026-08-22/header-bands.png, with the darkest pixel in the band recorded beside it. Not a per-run judgement made at capture time, which was inconsistent.",
   "label_cross_check": "DOM label matches the label visible in the screenshot",
   "screenshot_label_min_ink_0_255": 195,
   "screenshot_label_ink_note": "Darkest pixel in the 260x26 band right of the vendor icon, so the icon cannot contribute. 255 means the band is pure white, i.e. no label was rendered into the published image at all."
  },
  {
   "round": 1,
   "picked": "Mistral Small 4",
   "label_from_dom": "Mistral Small 4",
   "reply": "{\n  \"name\": \"Mars\",\n  \"moons\": 2,\n  \"diameter_km\": 6779\n}",
   "inCode": true,
   "pre_in_message": 1,
   "code_in_message": 1,
   "lang_tag": "json",
   "reply_chars": 57,
   "still_generating": false,
   "captured_at": "2026-08-22T03:03:44.458Z",
   "screenshot": "/shots/json-render-2026-08-22/mistral-small-4-r1.jpg",
   "rendered": "code block",
   "capture_version": "v2",
   "banner": "Zero data retention for this chat. No AI training.",
   "screenshot_label_legible": false,
   "screenshot_label_method": "Read by eye from a montage of all twelve header bands, published at /shots/json-render-2026-08-22/header-bands.png, with the darkest pixel in the band recorded beside it. Not a per-run judgement made at capture time, which was inconsistent.",
   "label_cross_check": "NOT CHECKABLE: the published screenshot shows the vendor icon and no model name, so there is nothing to cross-check against. For this run the model identity rests solely on the atomic DOM capture, which read the label and the reply in one evaluation from the container holding both.",
   "screenshot_label_min_ink_0_255": 243,
   "screenshot_label_ink_note": "Darkest pixel in the 260x26 band right of the vendor icon, so the icon cannot contribute. 255 means the band is pure white, i.e. no label was rendered into the published image at all."
  },
  {
   "round": 1,
   "picked": "gpt-oss 120B",
   "label_from_dom": "gpt-oss 120B",
   "reply": "{\n\"name\": \"Mars\",\n\"moons\": 2,\n\"diameter_km\": 6779\n}",
   "inCode": false,
   "pre_in_message": 0,
   "code_in_message": 0,
   "lang_tag": null,
   "reply_chars": 51,
   "still_generating": false,
   "captured_at": "2026-08-22T03:04:55.392Z",
   "screenshot": "/shots/json-render-2026-08-22/gpt-oss-120b-r1.jpg",
   "rendered": "plain text",
   "capture_version": "v2",
   "banner": "Zero provider visibility. No AI training.",
   "DEVIATION": "Round 0 for this model returned COMPACT one-line JSON at 44 chars. This round returned PRETTY-PRINTED multi-line JSON at 51 chars, still with NO code fence. The GRADE (plain text) is unchanged; the SHAPE is not. This refutes the perfect correlation between fencing and pretty-printing that round 0 alone suggested, and it independently reproduces the earlier trial's note that this model's exact shape is unstable while its grade is not.",
   "screenshot_label_legible": false,
   "screenshot_label_method": "Read by eye from a montage of all twelve header bands, published at /shots/json-render-2026-08-22/header-bands.png, with the darkest pixel in the band recorded beside it. Not a per-run judgement made at capture time, which was inconsistent.",
   "label_cross_check": "NOT CHECKABLE: the published screenshot shows the vendor icon and no model name, so there is nothing to cross-check against. For this run the model identity rests solely on the atomic DOM capture, which read the label and the reply in one evaluation from the container holding both.",
   "screenshot_label_min_ink_0_255": 255,
   "screenshot_label_ink_note": "Darkest pixel in the 260x26 band right of the vendor icon, so the icon cannot contribute. 255 means the band is pure white, i.e. no label was rendered into the published image at all."
  },
  {
   "round": 1,
   "picked": "Gemma 4 31B",
   "label_from_dom": "Gemma 4 31B",
   "reply": "{\n  \"name\": \"Mars\",\n  \"moons\": 2,\n  \"diameter_km\": 6779\n}",
   "inCode": true,
   "pre_in_message": 1,
   "code_in_message": 1,
   "lang_tag": "json",
   "reply_chars": 57,
   "still_generating": false,
   "captured_at": "2026-08-22T03:06:04.553Z",
   "screenshot": "/shots/json-render-2026-08-22/gemma-4-31b-r1.jpg",
   "rendered": "code block",
   "capture_version": "v2",
   "banner": "Zero provider visibility. No AI training.",
   "screenshot_label_legible": true,
   "screenshot_label_method": "Read by eye from a montage of all twelve header bands, published at /shots/json-render-2026-08-22/header-bands.png, with the darkest pixel in the band recorded beside it. Not a per-run judgement made at capture time, which was inconsistent.",
   "label_cross_check": "DOM label matches the label visible in the screenshot",
   "screenshot_label_min_ink_0_255": 239,
   "screenshot_label_ink_note": "Darkest pixel in the 260x26 band right of the vendor icon, so the icon cannot contribute. 255 means the band is pure white, i.e. no label was rendered into the published image at all."
  }
 ]
}