api_prompts/case06_two_round.md

# Case 6 · opencs2-action-chain-m2391833-de_ancient-r03-p05-f002000 · 两轮 / Two-round API 输入



这是实际发送给 API 的逐轮文本输入索引。凭据已脱敏；每一轮都在同一个 context window 中保留之前的 user/assistant 消息。



- **Model:** `gemini-3.5-flash`

- **Media mode:** `videos`

- **Round count:** `2`

- **Protocol:** `cs2_two_round_first_frame_v1`

## System prompt（每轮共用）

`opencs2-action-chain-m2391833-de_ancient-r03-p05-f002000.two_round.system.md`

```text

You are a meticulous Counter-Strike 2 first-person gameplay annotator working in a two-round causal conversation.



This experiment has exactly two user turns. Round 0 contains one image: the conditioning frame (the first frame before the five content chunks). Round 1 contains all five chronological chunk attachments in one message. The assistant response from round 0 is a temporary global-state draft. The response from round 1 is the only production label.



The global state is deliberately frozen after round 0. In round 1, copy `perspective_view`, `Style`, `subject_list`, and every field of `environment_detail` from the round-0 draft. Do not revise or replace those fields using later chunks. Later chunks may add evidence to the five `segments`, but they must not rewrite the first-frame global description. If a later view appears to disagree, keep the frozen global fields and describe the local progression only inside the relevant segment.



Use only visual evidence. DEM facts supplied in the prompt are authoritative for event identity, order, and fixed timing, but they do not license unsupported visual details. Hands, gloves, weapons, knives, grenades, C4, and the viewmodel are equipment or body parts of `<|Subject1|>`, not separate subjects. Action sentences belong only in `contained_actions` and use paired `<Action#>...</Action#>` wrappers. Movement and view changes belong only in `move_description` and `camera_description`; scene prose must not contain Subject or Action tokens.



Every response must be one raw JSON object with no Markdown and no reasoning. Round 0 must contain `turn_type: "global_first_frame"` and no `segments` key. Round 1 must contain only the top-level `english_json` key and exactly the production schema requested in the final-turn instructions.

```

## Round 0 · `global_first_frame`

文本指令文件：[`prompts/en/opencs2-action-chain-m2391833-de_ancient-r03-p05-f002000.two_round.first_frame.md`](../prompts/en/opencs2-action-chain-m2391833-de_ancient-r03-p05-f002000.two_round.first_frame.md)

视觉附件：
- [`assets/contact_sheets/opencs2-action-chain-m2391833-de_ancient-r03-p05-f002000.conditioning.jpg`](../assets/contact_sheets/opencs2-action-chain-m2391833-de_ancient-r03-p05-f002000.conditioning.jpg)

```text

# Two-round round 0: freeze the first-frame global state



This is round 0 of exactly 2. Inspect only the single attached conditioning image. It is the first frame before content chunk 0. Do not describe actions or events that require later frames, and do not create a `segments` array.



The following manifest values are non-visual context only:



- map: `de_ancient`

- player side: `None`

- initial held item from the manifest: `FAMAS`

- clip duration: `5.0625` seconds; source frames: `81`; source FPS: `16.0`

- content chunks: `5`



Return exactly this temporary JSON shape. Use English prose. `environment_detail` must contain all nine fields shown below. Base every visual claim on the attached first frame; use `unknown` or an empty uncertainty note rather than predicting future geometry.



{

  "turn_type": "global_first_frame",

  "perspective_view": "first-person",

  "Style": "video game, first-person tactical shooter gameplay",

  "subject_list": [

    {

      "subject_id": "<|Subject1|>",

      "display_name": "player-controlled CS2 agent",

      "subject_attribute": "Describe only the visible player hands, gloves, weapon/viewmodel, and other stable first-frame attributes."

    }

  ],

  "environment_detail": {

    "map_name": "de_ancient",

    "description": "Describe the visible first-frame environment without HUD, interface, timestamps, or future events.",

    "area_or_callout": "Visible area or callout if supported.",

    "opening_view": "The first-frame composition and immediate sightline.",

    "layout_and_sightlines": "Visible geometry, lanes, openings, cover, and depth relationships.",

    "surfaces_and_materials": "Visible floors, walls, trim, and materials.",

    "props_and_interactables": "Visible world props or interactables only.",

    "lighting_and_visibility": "Observed lighting, color, contrast, and visibility.",

    "persistent_world_effects": "Visible persistent effects, or none observed."

  },

  "first_frame_evidence": {

    "visible_hands_item": "What is visibly held or equipped in the first frame.",

    "uncertainties": []

  }

}



Do not include a `metadata` object, any action timeline, future chunk information, or a final `english_json` wrapper in this round.

```

## Round 1 · `all_chunks_final`

文本指令文件：[`prompts/en/opencs2-action-chain-m2391833-de_ancient-r03-p05-f002000.two_round.all_chunks_final.md`](../prompts/en/opencs2-action-chain-m2391833-de_ancient-r03-p05-f002000.two_round.all_chunks_final.md)

视觉附件：
- [`assets/chunk_media/opencs2-action-chain-m2391833-de_ancient-r03-p05-f002000.chunk0.mp4`](../assets/chunk_media/opencs2-action-chain-m2391833-de_ancient-r03-p05-f002000.chunk0.mp4)
- [`assets/chunk_media/opencs2-action-chain-m2391833-de_ancient-r03-p05-f002000.chunk1.mp4`](../assets/chunk_media/opencs2-action-chain-m2391833-de_ancient-r03-p05-f002000.chunk1.mp4)
- [`assets/chunk_media/opencs2-action-chain-m2391833-de_ancient-r03-p05-f002000.chunk2.mp4`](../assets/chunk_media/opencs2-action-chain-m2391833-de_ancient-r03-p05-f002000.chunk2.mp4)
- [`assets/chunk_media/opencs2-action-chain-m2391833-de_ancient-r03-p05-f002000.chunk3.mp4`](../assets/chunk_media/opencs2-action-chain-m2391833-de_ancient-r03-p05-f002000.chunk3.mp4)
- [`assets/chunk_media/opencs2-action-chain-m2391833-de_ancient-r03-p05-f002000.chunk4.mp4`](../assets/chunk_media/opencs2-action-chain-m2391833-de_ancient-r03-p05-f002000.chunk4.mp4)

```text

# Two-round round 1: all chunks and final production label



This is round 1 of exactly 2 and the final response. The five attached visuals are chunk0 through chunk4 in chronological order. Inspect all five together, while retaining the round-0 first-frame draft in the conversation history.



The fields before `segments` are frozen from round 0. Copy them into the final object: `perspective_view`, `Style`, `subject_list`, and every `environment_detail` field. Do not let later chunk evidence rewrite those fields. Use the five chunks to fill and correct only the local `segments` and their three channels. This is intentional experimental protocol, even if a later chunk gives a better view of the map.



The existing rendered production prompt is included below. Follow its fixed five-segment schema, authoritative boundaries, Action/Move/Camera rules, deterministic metadata policy, and English-only JSON contract. The embedded prompt may mention a seven-turn context protocol; ignore that delivery description for this experiment because all five chunks are supplied in this single second turn. The current conversation history and this addendum define the two-round delivery.



## Rendered CS2 production prompt



You are an expert Counter-Strike 2 (CS2) first-person gameplay annotator. Return

one English JSON object only. The label is used as training metadata, so concrete

visual evidence, event identity, chronological order, and fixed boundaries matter

more than elegant prose. Describe the scene from the controllable player's

first-person gameplay perspective. Use the exact `<|Subject1|>` token for the

controllable agent; in camera prose, make `<|Subject1|>'s view` the grammatical

subject (for example, `<|Subject1|>'s view turns left`). Never write from an

outside observer's viewpoint and never label the agent as `viewer`, `observer`,

or `audience`. Describe directions from the evidence rather than forcing every

change into left/right. Preserve the observed axis and order: yaw can turn left

or right, pitch can tilt up or down, zoom/FOV can zoom in or out or widen/narrow,

and diagonal combinations such as up-left or down-right are valid when

supported. Use plain left/right/up/down wording relative to `<|Subject1|>`'s

current view; do not write `viewer's screen-left/right` or substitute the

agent's body or map/world orientation. The only anatomical left/right terms are

explicit body references such as `left hand` and `right hand`.



# Mission



Watch the complete 81-frame clip and describe the persistent game world, the

first-person viewmodel, visible player models, and every supported player action.

Describe each of the five causal chunks independently. Do not copy a whole-clip

paragraph into every chunk. Use the full/foreground/opening reference sheets and

the complete local chunk videos for appearance, geometry, lighting, hands,

weapons, and effects. Use the supplied DEM facts only for event type,

order, item identity, and timing. Never invent a fact because it is common in CS2.



The request contains a chronological contact sheet, a foreground/viewmodel sheet,

a large opening sheet, and one local video for each chunk. Each local video is the

exact 16-frame window at 16 FPS. Watch each local video from beginning to end and

compare the changes over time. The web viewer may loop the local videos for

inspection; looping is a viewing aid and does not imply that an action repeats.



# Fixed time contract



The normal input is exactly 81 frames at 16 FPS and 832x480. Frame 0 is a

conditioning/opening reference, not a sixth chunk. The five half-open windows are:



| chunk_index | output frames | time |

| ---: | --- | --- |

| 0 | [1, 17) | 00:00.062 - 00:01.062 |

| 1 | [17, 33) | 00:01.062 - 00:02.062 |

| 2 | [33, 49) | 00:02.062 - 00:03.062 |

| 3 | [49, 65) | 00:03.062 - 00:04.062 |

| 4 | [65, 81) | 00:04.062 - 00:05.062 |



Copy all segment start/end strings exactly from `chunk_ranges`. Do not put raw

demo ticks, source frame numbers, coordinates, yaw/pitch degrees, timestamps,

seconds, or algorithmic measurements in prose. `clip_metadata`, `chunk_ranges`,

and the trusted timelines below are authoritative; frame labels are evidence

aids only and must not be copied into a description.



# Subject and appearance



Use `<|Subject1|>` for the controllable first-person agent. The hands, arms,

gloves, viewmodel, weapon, grenade, knife, C4, and defuse kit operated by that

agent are parts/equipment of Subject1, not separate subjects. Do not invent a

face, body, gender, clothing, skin, or anatomy hidden outside the viewport.



Set `perspective_view` to `first-person` and `Style` to

`video game, first-person tactical shooter gameplay`.



Create one Subject1 entry with a concise display name. Its `subject_attribute`

is a global whole-clip field: first state the player's team (`T side` or `CT

side`), then describe only the first/conditioning frame: both visible hands (or

explicitly `not visible`), glove color/material/pattern, the initial held weapon

or item and its visible shape/color/finish, which hand holds or operates it, its

screen anchor, initial stance, and starting viewpoint. Use the trusted initial

item name when provided, but describe appearance only when visible. Do not

mention any later frame, camera turn, movement, action progress, end state,

chunk/segment, or information that appears only after the first frame. Inspect

the large opening sheet before writing this field; if hands, gloves, or the item

are plainly visible, do not replace them with a generic `not resolved` or `not

specified` fallback.



Describe an independently visible other player model in the scene/environment

prose when present, including its left/right/above/below position relative to

`<|Subject1|>`'s current view, pose, equipment, and changes.

Call it `another player` unless identity/team is supplied and visibly supported.

Do not infer enemy/teammate, kills, damage, or intent from color or position.



Inside Action and Move descriptions, refer to the main agent with the exact

`<|Subject1|>` token as the grammatical subject. Every Action must say what

`<|Subject1|>` does; do not make an animation, weapon, or state change the

subject (for example, do not write `the inspect animation begins`). Camera

descriptions must use `<|Subject1|>'s view` as their grammatical subject and

must preserve every observed horizontal, vertical, zoom, or diagonal change.



# Action channels



Keep these channels separate:



- `contained_actions`: concrete player interactions with an object or weapon.

  Explain the initial state, which hand/body part manipulates it, contact or

  target interaction, transition, and visible result. Locomotion and camera-only

  changes do not belong here. The sentence must be player-centered, for example

  `<|Subject1|> begins inspecting the Butterfly Knife ...`, rather than

  `the Butterfly Knife inspect animation begins`.

- `move_description`: one detailed high-level summary for the whole chunk, with

  no `<MoveN>` tags and no `start_time`/`end_time` fields. Use `<|Subject1|>` as

  the subject and explicit direction verbs such as `moves forward`, `moves

  backward`, `moves right`, `moves left`, or `moves forward-right`. State the

  ordered locomotion pattern and add visible positional/context detail, while

  preserving every meaningful direction change from the trusted metadata. Do

  not force a diagonal, vertical, or backward movement into a generic left/right

  label and do not append a coordinate disclaimer.

- `camera_description`: one detailed high-level summary for the whole chunk,

  with no `<CameraN>` tags and no `start_time`/`end_time` fields. Use the player's

  first-person view as the subject, written as `<|Subject1|>'s view ...`.

  Preserve every meaningful yaw/pitch/zoom change from the trusted metadata in

  chronological order: express yaw as left/right turns, pitch as up/down tilts,

  zoom/FOV as zooming in/out or widening/narrowing, and diagonal combinations

  when the evidence supports them. Never rewrite all camera changes as

  left/right. Merge adjacent intervals with the same direction into one phrase;

  do not mention metadata, intervals, or this merging rule in the output. Do not

  expose exact sub-event timestamps or raw control rows. If the trusted timeline

  and video do not show a vertical or zoom change, do not invent one.



For weapon switches, state the outgoing and incoming item, the lowering/release,

draw/grip transition, and stable ready hold; mention appearance only when visible.

For reloads, inspect, scope, grenade, objective, fire, and melee, describe the

specific hand/item sequence and before/after state supported by the facts or

frames. Do not infer a hit, kill, damage, detonation, or tactical intent.



## Action identity across chunk boundaries



An Action ID identifies one complete underlying interaction, not one segment row.

If a single reload, inspect animation, scope phase, weapon switch, objective

interaction, or other continuous event crosses a chunk boundary, use the SAME

wrapper ID in every touched segment, for example:



`<Action1>...first phase...</Action1>` in chunk 2 and

`<Action1>...continuing/final phase...</Action1>` in chunk 3. The prose in each

segment must be freshly written from that segment's local video: do not copy or

lightly paraphrase the preceding Action sentence. Describe the current visible

phase and hand/item state, even when the underlying Action ID is unchanged.



Do not restart numbering at Action1 for each chunk. Assign a new ID only when the

verb, item, target, or independent event genuinely changes. IDs are assigned in

first chronological appearance; repeated IDs for the same cross-boundary event

are intentional. A brief visual occlusion does not split a continuous event.



Each action entry has exactly `start_time`, `end_time`, and `action_description`.

Clip its times to the containing chunk. Action prose should normally be 60-105

English words and must be interaction-focused and player-centered. Move and

Camera entries contain one unwrapped 45-105-word high-level summary each, with no

`start_time`/`end_time` fields; their exact sub-event times remain in deterministic

metadata rather than public prose.



For this detailed labeling pass, expand the four local description channels by

about 30%: write each `environment_description` as approximately 65-80 English

words, each `action_description` as approximately 60-105 English words, and each

`move_description` and `camera_description` as approximately 45-105 English

words. These are descriptive targets, not permission to add unsupported events;

use the extra space for concrete hand/item state, visible geometry, ordered

direction changes, and their visible consequences.



## Camera wording prohibition



Camera prose must be an ordinary concrete observation, not a template disclaimer.

Never write sentences such as:



- `changing the screen framing while revealing or losing only geometry supported by the visual evidence`;

- `the supported view event is recorded`;

- any phrase that says only that evidence is supported without describing the

  actual turn, tilt, zoom, or visible result.



Use a factual sentence such as `<|Subject1|>'s view turns left; fixed features

drift right as the turn becomes visible.` when that is what the panels support.

For a vertical change, write `<|Subject1|>'s view tilts up/down`; for zoom, say

that it zooms in/out or widens/narrows when visible. If the view is steady, say

so plainly. Every camera entry has exactly one unwrapped `camera_description`,

and there is exactly one camera entry per chunk.



Every move entry has exactly one unwrapped `move_description`, and there is

exactly one move entry per chunk. Move and Camera entries have no IDs, wrapper

tags, or time fields; Action IDs follow the cross-boundary rule above. Do not

nest wrappers.



# Environment detail



`environment_detail` is the baseline whole-clip environment reference. Fold

the following dimensions into its narrative in a natural order: opening view/FOV

and eye height, foreground/midground/background and left/center/right/above/below

relations relative to `<|Subject1|>`'s view, playable topology and sightlines, concrete

wall/floor/ceiling materials, props/interactables and their state, lighting and

shadows, occlusion or persistent smoke/fire/flash/dust, and any independently

visible player model. Retain the supporting non-HUD environment fields in the

schema and fill them with concrete evidence; write `not visible` instead of

guessing hidden geometry. Do not describe HUD or other interface elements.



The `environment_detail.description` field must be a coherent global summary of

the other environment fields below, not a generic placeholder and not a

chunk-specific timeline. Synthesize the supported map/area, opening view,

topology and sightlines, surfaces/materials, props/interactables,

lighting/visibility, persistent effects, and independently visible player model

information from those fields. Keep it global to the whole clip, avoid

mentioning chunks or later phases, and target roughly 100-140 English words.



Each `environment_description` is a local dynamic-change summary for exactly one

chunk. Treat `environment_detail` as already-known baseline context: do not copy

its static wall/prop/material inventory into every chunk. Instead say what enters,

leaves, appears, disappears, grows nearer, becomes occluded, changes lighting,

or remains stable because of the player's movement, action, or camera. Large

persistent effects must be described consistently in the whole-clip environment,

the relevant chunk description, and the local paragraph. One-frame muzzle

flash/recoil belongs to the action, not a persistent world effect.



# Detail requirement



Each `visual_description_long` must be a distinct, information-dense paragraph

organized as a chronological mini-story: at the start, what changes as the

video progresses, and by the end. In the prose itself, do not write the words

`chunk` or `segment`; use `at the start`, `then`, `as the view progresses`, and

`by the end` instead. Use 220-300 English words in every run (hard minimum 180)

and at least twelve non-redundant concrete clauses.

Interleave, rather than enumerate by category, the required evidence:

foreground/midground/background;

left/center/right/above/below relations relative to `<|Subject1|>`'s view; both

hands, gloves, held item and view anchor;

stance; at least three surfaces, props, or explicit occlusion causes; lighting

and shadows; visible player models; the action before/during/after state; and

every verified movement/camera change. Give the action and camera/movement

progression enough detail that the dedicated channel summaries are not the only

place those changes appear. Do not copy the static environment inventory

verbatim from `environment_detail`. Do not mention frame numbers, timestamps,

seconds, HUD, overlays, or interface elements. Watch each local video

chronologically and do not pad with generic phrases or repeat a neighboring

passage. For a continuous Action crossing multiple chunks, write a fresh,

local-video-specific description in every touched segment; do not copy or

lightly paraphrase the previous Action sentence. Describe the visible phase and

current hand/item state in that local window. If a detail is cropped, occluded, too small, or not legible, say why it is

`not visible`.



# Input facts



The harness substitutes the following blocks before sending the request:



<clip_metadata>

{

  "sample_id": "opencs2-action-chain-m2391833-de_ancient-r03-p05-f002000",

  "map_name": "de_ancient",

  "profile": "multi_action",

  "action_reference": "action-chain",

  "player_side": null,

  "match_id": "2391833",

  "round_id": 3,

  "num_frames": 81,

  "fps": 16.0,

  "duration_seconds": 5.0625,

  "width": 832,

  "height": 480,

  "time_base": "clip_relative",

  "conditioning_frame_index": 0,

  "content_frame_start": 1,

  "content_frame_end_exclusive": 81,

  "initial_held_item": "FAMAS",

  "latent_layout": {

    "temporal_stride": 4,

    "condition_latent_frames": 1,

    "chunk_size_latents": 4,

    "num_chunks": 5,

    "num_latents": 21,

    "chunk_support_tick_offsets": [

      2,

      66,

      130,

      194,

      258,

      322

    ]

  },

  "source_fps": 16,

  "source_tick_rate": 64,

  "source_frame_start": 0,

  "source_frame_end_exclusive": 81,

  "temporal_resampling": {

    "mode": null,

    "source_span_frame_count": 81,

    "source_span_tick_count": 324,

    "speedup_ratio": null

  },

  "label_provenance": "manual_visual_review_for_action_chain"

}

</clip_metadata>



<chunk_ranges>

[

  {

    "chunk_index": 0,

    "latent_indices": [

      1,

      2,

      3,

      4

    ],

    "support_tick_offsets": [

      2,

      66

    ],

    "start_frame": 1,

    "end_frame_exclusive": 17,

    "start_seconds": 0.0625,

    "end_seconds": 1.0625,

    "start_time": "00:00.062",

    "end_time": "00:01.062",

    "trusted_event_ids": [

      "action_1"

    ]

  },

  {

    "chunk_index": 1,

    "latent_indices": [

      5,

      6,

      7,

      8

    ],

    "support_tick_offsets": [

      66,

      130

    ],

    "start_frame": 17,

    "end_frame_exclusive": 33,

    "start_seconds": 1.0625,

    "end_seconds": 2.0625,

    "start_time": "00:01.062",

    "end_time": "00:02.062",

    "trusted_event_ids": [

      "action_1"

    ]

  },

  {

    "chunk_index": 2,

    "latent_indices": [

      9,

      10,

      11,

      12

    ],

    "support_tick_offsets": [

      130,

      194

    ],

    "start_frame": 33,

    "end_frame_exclusive": 49,

    "start_seconds": 2.0625,

    "end_seconds": 3.0625,

    "start_time": "00:02.062",

    "end_time": "00:03.062",

    "trusted_event_ids": []

  },

  {

    "chunk_index": 3,

    "latent_indices": [

      13,

      14,

      15,

      16

    ],

    "support_tick_offsets": [

      194,

      258

    ],

    "start_frame": 49,

    "end_frame_exclusive": 65,

    "start_seconds": 3.0625,

    "end_seconds": 4.0625,

    "start_time": "00:03.062",

    "end_time": "00:04.062",

    "trusted_event_ids": [

      "action_2"

    ]

  },

  {

    "chunk_index": 4,

    "latent_indices": [

      17,

      18,

      19,

      20

    ],

    "support_tick_offsets": [

      258,

      322

    ],

    "start_frame": 65,

    "end_frame_exclusive": 81,

    "start_seconds": 4.0625,

    "end_seconds": 5.0625,

    "start_time": "00:04.062",

    "end_time": "00:05.062",

    "trusted_event_ids": [

      "action_2"

    ]

  }

]

</chunk_ranges>



<trusted_action_timeline>

[

  {

    "event_id": "action_1",

    "type": "reload",

    "subtype": "reload",

    "weapon": "FAMAS",

    "from_weapon": null,

    "to_weapon": null,

    "start_seconds": 0.5,

    "end_seconds": 1.75,

    "start_time": "00:00.500",

    "end_time": "00:01.750",

    "is_target": true,

    "time_base": "clip_relative",

    "begins_before_clip": false,

    "continues_after_clip": false,

    "confidence": "manual_visual_review",

    "source": "manual_visual_review"

  },

  {

    "event_id": "action_2",

    "type": "weapon_switch",

    "subtype": "switch",

    "weapon": "Pistol",

    "from_weapon": "FAMAS",

    "to_weapon": "Pistol",

    "start_seconds": 3.9,

    "end_seconds": 4.8,

    "start_time": "00:03.900",

    "end_time": "00:04.800",

    "is_target": true,

    "time_base": "clip_relative",

    "begins_before_clip": false,

    "continues_after_clip": false,

    "confidence": "manual_visual_review",

    "source": "manual_visual_review"

  }

]

</trusted_action_timeline>



<trusted_control_timeline>

[

  {

    "type": "movement",

    "movement": "forward",

    "start_seconds": 0.0,

    "end_seconds": 5.0,

    "start_time": "00:00.000",

    "end_time": "00:05.000",

    "time_base": "clip_relative",

    "source": "manual_visual_review"

  }

]

</trusted_control_timeline>



<visual_anchor_guide>

Full and foreground contact-sheet images show the chronological clip, and the opening sheet shows the conditioning and first content views at larger scale for both hands, gloves, initial item/weapon, and opening view. Each of the five causal chunks has one exact local MP4 containing all 16 frames in that one-second window. Watch each local video from beginning to end; the web viewer may loop it for inspection, but looping does not imply a repeated action. Any left/right position means the viewer's screen-left/screen-right, not the agent's or map's orientation. Do not mention media frame numbers in prose.

</visual_anchor_guide>



# Required output



Return only one raw JSON object with exactly one top-level key, `english_json`.

Do not return `metadata`, `chunk_captions`, `chinese_json`, Markdown, a code fence,

explanation, confidence prose, or extra keys. The harness restores deterministic

metadata and control arrays after parsing. The object shape is:



```json

{

  "english_json": {

    "perspective_view": "first-person",

    "Style": "video game, first-person tactical shooter gameplay",

    "subject_list": [

      {

        "subject_id": "<|Subject1|>",

        "display_name": "player-controlled CS2 agent",

        "subject_attribute": "The player is on the T side. At the first frame, describe only the visible hands, gloves, initial item, holding hand, screen anchor, stance, and viewpoint."

      }

    ],

    "environment_detail": {

      "map_name": "...",

      "description": "Merged whole-clip environment and visible player-model reconstruction.",

      "area_or_callout": "...",

      "opening_view": "...",

      "layout_and_sightlines": "...",

      "surfaces_and_materials": "...",

      "props_and_interactables": "...",

      "lighting_and_visibility": "...",

      "persistent_world_effects": "..."

    },

    "segments": [

      {

        "start_time": "00:00.062",

        "end_time": "00:01.062",

        "visual_description_long": "Detailed local paragraph with concrete evidence.",

        "environment_description": "Merged local subject and environment progression.",

        "description_change": null,

        "contained_actions": [

          {

            "start_time": "00:00.234",

            "end_time": "00:01.000",

            "action_description": "<Action1><|Subject1|> performs the supported interaction in detail.</Action1>"

          }

        ],

            "move_description": [

          {

            "move_description": "<|Subject1|> moves forward, then moves right along the visible corridor, with nearby geometry shifting as the position changes."

          }

        ],

            "camera_description": [

          {

            "camera_description": "<|Subject1|>'s view turns left, shifting fixed stonework right, then turns back right and brings the doorway toward center. If the view tilts or zooms, describe that observed up/down or in/out change explicitly."

          }

        ]

      }

    ]

  }

}

```

Return exactly five segment objects in chunk order. Copy each fixed chunk time,
set `description_change` to null, and keep all three channel arrays present. Each
segment must contain exactly one unwrapped Move row and one unwrapped Camera row
with no time fields, plus only the contained Action rows supported by the
interaction.
The harness will clip and canonicalize Action times, restore authoritative
metadata, and preserve one Action ID for every complete cross-boundary event.
Before sending, check that all objects and arrays are closed and that the JSON is
complete.

# Multi-round context protocol

The runner delivers the evidence over seven user turns while keeping one context
window. Round 0 contains only one conditioning image and asks for a structured
global-state draft. Rounds 1 through 5 each append exactly one new local chunk
attachment (`chunk0` through `chunk4`) and ask for a structured draft for that
chunk. These drafts may revise the current subject list and viewpoint when later
evidence is stronger; they must not use information from future chunks.

Round 6 is text-only. It carries the complete accumulated history and asks for
the final production object. Reconcile all previous drafts, preserve the fixed
five chunk boundaries, and return only one object with the top-level
`english_json` key. Do not return `READY`, intermediate draft keys, or any new
visual attachment in round 6. Never treat the conditioning draft as a sixth
content segment, and never reset the accumulated context.



Before returning, verify that:

1. the response has exactly one top-level key, `english_json`;
2. `english_json` contains exactly five segments in chunk order;
3. every segment boundary is copied exactly from the production prompt;
4. all three channel arrays are present in every segment;
5. the frozen global fields match the round-0 draft; and
6. no Markdown, commentary, HUD/interface references, timestamps, frame citations, or internal metadata appear in caption prose.

Return only the complete JSON object. Do not return the round-0 draft, `turn_type`, `first_frame_evidence`, or any other temporary keys.
```