CONFIRMED BY THE USER FOR THIS RUN

1. Fixed caption and visual context
The canonical `caption` is supplied from `refined_caption` by removing only the
source <p>, </p>, and [SEG:n] markup, collapsing whitespace, and trimming its ends.
Its wording and punctuation are otherwise preserved. Return this exact string.
All source fields are retained outside your output. Do not treat the source SEG
markers, token indices, or old caption-to-object links as gold text annotations;
those old links are deliberately not supplied as answers.
You see ten clean sampled RGB frames in temporal order; only ONE reference frame
has supplied masks. Frame indices 00--09 are sample indices, not native frame IDs.
The clean frames are context for disambiguation and visual fact checking. Grounded
mask links refer only to the supplied masks at the stated reference frame.
The reference RGB, colored ID mask overlays, per-ID identity crops, and labeled
binary-mask sheets identify the available masks. All supplied source masks are
shown; no generated or propagated masks are provided. Match by the actual images,
not just by the source category hint. A supplied person mask does not automatically
provide a separate hand, face, clothing, or other part mask.

2. Missing-mask policy explicitly chosen by the user
Annotate the full caption according to the complete guidelines, including eligible
entities for which no matching source mask is supplied. Do not omit their text,
rewrite the caption to remove them, or attach their text to a different entity's
mask. Keep all mentions/coreference groups. Set `mask_id` to null, choose the
appropriate grounding_status, and include a brief reason. These groups remain
incomplete spatial annotations; they are not complete image--text correspondences
and must not be called final one-to-one training examples. If an entity is visible
only in a context frame, or is not verifiable in any supplied image, report that
visibility explicitly; never invent a reference-frame mask or assert an unobserved
object is present. The caption remains unchanged even when its content is suspect.

3. Exact spans and inconsistent examples explicitly reviewed by the user
The source document above remains unchanged. In its Example 7, `a cloudy sky` is
actually at [38,50), whereas [37,50) includes a preceding space. All actual indices
for this run are computed locally against the supplied canonical caption, and
every resulting caption[start:end] must equal the exact mention text.
Return exact mention substrings and their 1-based occurrence numbers, not guessed
numeric character offsets. Include every repeated occurrence independently.
In section 3.1.1, `rightmost bus` is not a contiguous substring of
`the rightmost and leftmost bus`; likewise `wooden chair` is not a contiguous
substring of `the wooden and metal chair`. When a construction cannot be expressed
as exact continuous spans under the original rules, report a rule_conflict issue
and put the unresolved text in that issue. Preserve every other valid annotation;
do not fabricate a substring, silently join words, rewrite the text, or choose an
unapproved reinterpretation. The affected sample requires manual review and is
excluded from the formal completed set until the issue is resolved.
The document contains stale section references to section 2.8; its complete
counter/container/proper-name rules are present in sections 2.6.1--2.6.3. No section
has been deleted from the supplied original document.

4. Group-to-mask completeness
Keep the document's text grouping decisions (including plural groups, stuff,
body parts, possessives, selectors, and shared anchors). Each resolved group has
one supplied `mask_id`. If a collective group would require merging several
separate source masks, leave mask_id null, use requires_mask_union, and list the
source IDs in candidate_mask_ids for review. Do not silently split a text group,
merge source masks, or relabel several distinct instances as one confirmed mask.
If an identity is ambiguous, use mask_id null and ambiguous, with candidates where
available. A possibly incorrect/partial source mask can be marked mask_quality_issue.
Do not force a match to maximize coverage. Separate missing_mask from ambiguous
identity and from not_visible_in_reference. Repeated mentions of one entity stay
in the same group, but do not merge distinct entities just because their categories
match or an old source caption would have linked them to the same SEG ID.

5. Output transport only -- not a replacement for the guidelines
Return exactly one JSON object, no Markdown or explanatory text outside the JSON:
{
  "sample_id": "the supplied sample ID",
  "caption": "the exact supplied canonical caption",
  "groups": [
    {
      "group_id": "g1",
      "entity_label": "brief identity description",
      "entity_kind": "thing | stuff | body_part | collective",
      "mentions": [{"text": "exact substring", "occurrence": 1}],
      "mask_id": "m0 or null",
      "candidate_mask_ids": [],
      "grounding_status": "matched | missing_mask | not_visible_in_reference | ambiguous | requires_mask_union | mask_quality_issue",
      "visibility": "reference_visible | context_only | not_observed | uncertain",
      "confidence": "high | medium | low",
      "reason": "brief evidence or reason for the missing/uncertain link"
    }
  ],
  "issues": [{"type": "rule_conflict | caption_visual_mismatch | uncertain_coreference | other", "text": "exact relevant text or empty string", "description": "brief explanation"}],
  "needs_review": true,
  "review_reasons": []
}
Use actual JSON null (not the string "null") for an unresolved mask_id. Enumerated
values above are alternatives; choose one value, not the whole pipe-separated list.
Use only supplied mask IDs for links/candidates. `matched` requires one non-null
mask ID, reference_visible, and clear same-entity visual evidence. Missing spatial
labels are not background/negative evidence. Set needs_review=true for every
unresolved mask/group, uncertainty, source-mask problem, or rule/caption issue.
Before returning, check full caption coverage, all repeated references, exact
occurrences, body-part separation, possessive overlap, no overlap inside each group,
and the complete original rule set. A structural JSON success does not by itself
certify semantic accuracy or human review.