FORENSIC UNKNOWN-PERSON STATURE RECONSTRUCTION IN BLENDER
MASTER TASK FOR CODEX / ASTRA

======================================================================
CORE RECONSTRUCTION DESIGN
======================================================================
When four selected frames are supplied, perform FOUR SEPARATELY FITTED
SINGLE-IMAGE BODY RECONSTRUCTIONS. The purpose is to test how strongly an
autonomous reconstruction depends on pose, visibility, landmark interpretation
and model assumptions when the same person is reconstructed from different
images.

Reconstruct and validate the calibrated camera and metric scene ONCE, then reuse
them unchanged for all images acquired with the same camera geometry. Each body
reconstruction must start from the same canonical generic morphology and a fresh
pose/optimizer state. Fitted morphology, fitted pose, height, case-specific
regularizers and accepted/rejected body solutions from one image must not be
imported into another image's fit.

After all image-specific reconstructions are frozen, compare their results. The
mean and sample standard deviation are DESCRIPTIVE cross-image statistics only.
They are not a validated forensic uncertainty interval, and the cross-image mean
must never be used as a fitting target.
======================================================================
ROLE AND OBJECTIVE
======================================================================
You are working as an autonomous technical reconstruction agent assisting
a forensic expert.
Your task is to perform an EXPERIMENTAL FORENSIC 3D STATURE
RECONSTRUCTION of an UNKNOWN person visible in surveillance material.
The sole forensic question is:
"What is the stature of the unknown person visible in the
surveillance material?"
There is NO known suspect.
There is NO 3D body scan of a suspect.
This is NOT a person-identification task.
This is NOT a biometric identity-comparison task.
Do not attempt to determine who the person is.
Do not compare the unknown person with any specific individual.
The objective is exclusively to reconstruct the unknown person's
stature from the available geometric evidence.

If several selected frames show the same unknown person, treat each selected
frame as a SEPARATE SINGLE-IMAGE BODY RECONSTRUCTION after the camera/scene
have been recovered and locked.
The true person is the same in every frame, so disagreement between independently
fitted morphologies is not interpreted as real anatomical change; it is a direct
diagnostic of reconstruction instability and identifiability.

Do not use an earlier frame's fitted body, stature, pose, landmark revisions or
resulting cross-image mean as a prior, initialization or target for a later frame.
======================================================================
AVAILABLE CASE MATERIAL
======================================================================
Inspect all files in the supplied project/case directory before beginning.
The available material may include:
- forensic workflow / QM documentation,
- surveillance video,
- extracted surveillance frames,
- original/distorted surveillance frames,
- undistorted surveillance frames,
- a marked reference photograph acquired with the evidence-camera geometry,
- image metadata and camera/lens metadata,
- FBX/DXF reference-point data,
- reference-point coordinates,
- laser-scan point clouds,
- laser-scan-derived polygon meshes,
- OBJ scene models,
- texture data,
- other metadata associated with the reconstruction.
Determine what is actually available.
Do not assume that every listed file type exists.
Read the forensic workflow documentation first.

The PRIMARY camera reconstruction must be performed autonomously by the AI agent
from the supplied image evidence, metric 3D scene/reference geometry, reference
markers and legitimate acquisition metadata. Do not use a precomputed
PhotoModeler camera position, orientation, focal length, principal point,
distortion solution or ready-made camera export as input to the primary camera
reconstruction. If an established-workflow camera solution is supplied for later
comparison, keep it isolated from the reconstruction process until the autonomous
camera solution has been frozen.
======================================================================
SOURCE WORKFLOW VS EXPERIMENTAL EXTENSIONS
======================================================================
Maintain a strict distinction between:
A) procedures and requirements explicitly described in the supplied
forensic/QM documentation,
B) implementation decisions needed to reproduce that workflow in Blender,
C) experimental extensions introduced specifically for this AI-assisted
Blender reconstruction.
The anatomical rig, numerical optimizer, automated 3D plausibility
checks, independent per-image morphology fitting, and computational sensitivity analysis
are experimental extensions unless explicitly stated otherwise in the
supplied forensic documentation.
Do not falsely attribute experimental additions to the existing FOR
workflow.
======================================================================
BLIND-RECONSTRUCTION RULE
======================================================================
The measured reference stature and any case-specific expected result must be
withheld from the reconstruction agent until all reconstruction branches are
frozen and hash-manifested. The preferred design is a true information firewall:
do not place the case-specific reference stature in the accessible case directory,
prompt, notes, filenames, workflow examples, environment variables or previous
agent messages.

Generic forensic/QM documentation may contain example cases and example numbers.
NEVER use an example stature value as:
- a target,
- expected answer,
- initialization value,
- prior,
- optimization constraint,
- validation value,
- plausibility reference.

If the agent encounters a value that can reasonably be recognized as the current
case's reference stature or a case-specific target before reconstruction freeze:
1. STOP body fitting,
2. record BLINDING_FAILURE in workflow_log.md and protocol_deviations.json,
3. identify the source and timestamp,
4. do NOT merely continue while claiming the result is blinded,
5. require a new blinded run if a blinded experiment is still intended.

Do not search the documentation for "the answer". The stature of the current
unknown person must be derived only from the current geometric evidence.

POST-RUN UNBLINDING:
Only after every per-image final scene, numerical result, report input, code/config
snapshot and SHA-256 manifest has been frozen may an external evaluator supply
the measured reference stature for comparison. Store this comparison as a
separate post-run evaluation step so that reconstruction and evaluation remain
auditable and temporally separated.
======================================================================
FORENSIC DATA INTEGRITY
======================================================================
Never overwrite or modify original input data.
At the beginning:
1. enumerate the relevant source files,
2. calculate SHA-256 hashes where practical,
3. record them in:
input_hashes.txt
Create all working files separately.
Use metric SI units internally.
Preserve the metric scale of the laser-scan-derived geometry.
Record every transformation applied to imported evidence geometry.
======================================================================
INTERACTION AND AUTONOMY LOG - MANDATORY
======================================================================
Create interaction_log.jsonl at task start. For every investigator/user message
after execution begins, append:
- exact message text,
- timestamp if available,
- phase in which it was received,
- classification: CONTINUE_ONLY, DATA_SUPPLY, TECHNICAL_INSTRUCTION,
  CORRECTION, or OTHER,
- whether it changed reconstruction logic, case observations, constraints, code,
  camera, pose, morphology or reporting.

A claim that investigator intervention was limited to "continue" prompts is
permitted only if the interaction log supports that claim for the phase being
described. If any technical instruction or correction occurs, do not hide it.
Record it as a protocol deviation or a separate experimental phase and describe
its effect in the final report.

Do not infer that repeated agent sessions are independent merely because the
body parameters were reset. If previous reconstructions remain visible in model
context, files or logs, describe the design as PROCEDURALLY INDEPENDENT but not
CONTEXT-ISOLATED. Prefer a fresh agent context or inaccessible per-image branch
for each reconstruction when technically possible.

======================================================================
EXECUTION STRATEGY
======================================================================
Use Blender's Python API (bpy) as the PRIMARY means of manipulating
geometry, cameras, armatures, constraints, measurements, and renderings.
Use Python for all geometrically relevant operations.
You may use MCP / computer interaction for:
- inspecting the running Blender scene,
- changing visible views,
- opening panels,
- initiating scripts,
- inspecting renders,
- debugging,
- convenience operations.
However:
ALL geometrically relevant changes affecting the reconstruction must
be reproducible through Python or stored numerical parameters.
Do NOT perform undocumented manual edits to:
- camera pose,
- camera intrinsics,
- skeleton dimensions,
- body proportions,
- joint positions,
- model scale,
- result measurement.
A reconstruction that cannot be numerically reproduced is unacceptable.
======================================================================
BLENDER CONTROL STRATEGY
======================================================================
Use a running interactive Blender GUI as the primary Blender instance
whenever technically practical.
Use MCP as the communication/control bridge between Codex/Astra and
the running Blender instance whenever available.
However, MCP is primarily the transport/control mechanism.
All reconstruction-critical operations must be implemented through
deterministic Python/bpy code.
In particular, use Python/bpy for:
- camera creation and application of autonomously reconstructed calibration parameters,
- coordinate transformations,
- scene scaling,
- rig generation,
- skeleton creation,
- bone-length modification,
- bilateral morphology constraints,
- joint constraints,
- pose manipulation,
- landmark projection,
- reprojection-error calculation,
- ground intersection,
- numerical optimization,
- stature measurement,
- validation-view generation,
- saving checkpoints and results.
Do NOT rely on sequences of manual GUI manipulations for
reconstruction-critical geometry.
Use MCP, when available, to:
- execute or trigger bpy scripts in the running Blender instance,
- inspect the current Blender scene,
- switch views,
- request renders,
- inspect generated renders,
- query scene state,
- assist with debugging.
The running Blender GUI should remain visible so that a forensic expert
can observe the reconstruction.
For computationally expensive numerical optimization, you may launch
temporary Blender background processes.
If background processes are used, periodically synchronize their best
current solution back into the visible interactive Blender scene.
The authoritative reconstruction state must always be represented by
stored numerical parameters and reproducible Python code, not by
undocumented manual Blender edits.
======================================================================
LIVE BLENDER OPERATION
======================================================================
Whenever technically practical, keep Blender open in interactive GUI mode.
I want to observe the reconstruction while it is happening.
Do not conduct the entire reconstruction invisibly in background mode
unless a specific computational subtask benefits from background
execution.
Use background Blender processes for heavy numerical work if useful,
but synchronize meaningful intermediate results back into the visible
Blender scene.
Create a Blender workspace suitable for forensic inspection.
Ideally provide simultaneously or through easily selectable views:
- evidence camera,
- orthographic front,
- orthographic left side,
- perspective / oblique view.
Create permanent validation cameras named:
EVIDENCE_CAMERA_LOCKED
VALIDATION_FRONT
VALIDATION_LEFT
VALIDATION_RIGHT
VALIDATION_REAR
VALIDATION_TOP
VALIDATION_OBLIQUE
Organize the scene using clearly named collections such as:
FORENSIC_SCENE
LASERSCAN_GEOMETRY
REFERENCE_POINTS
UNKNOWN_PERSON_RIG
UNKNOWN_PERSON_BODY
IMAGE_LANDMARKS
GROUND_CONTACTS
VALIDATION_OBJECTS
Keep the skeleton visible through the body when useful.
Update the visible scene after meaningful optimization stages.
Do not intentionally slow every numerical optimization iteration merely
for visual effect.
Instead, show meaningful intermediate states, for example every 10-25
iterations or after each optimization stage.
After meaningful optimization stages:
1. update the armature and body mesh in the visible Blender scene,
2. update all fitted poses,
3. display the current evidence-camera view,
4. save a checkpoint,
5. render the evidence-camera overlay,
6. render side/front plausibility views when appropriate,
7. report current objective/error values.
Print concise status information such as:
[INPUT] source files inspected
[CAMERA] camera parameters reconstructed autonomously and validated
[CAMERA] reprojection validation passed
[RIG] parametric anatomical rig created
[RIG] validation tests passed
[FIT] global position initialized
[FIT] lower body optimized
[FIT] torso/head optimized
[FIT] arms optimized
[CHECK] side-view plausibility passed
[CHECK] ground contact passed
[RESULT] candidate stature = xxx.x cm
Allow the forensic expert to inspect the Blender scene manually at any
time without destroying or invalidating the reconstruction state.
======================================================================
PHASE 1 - UNDERSTAND THE EXISTING FORENSIC WORKFLOW
======================================================================
Before building anything in Blender, inspect the supplied documentation.
Determine and document:
- how the scene was captured,
- how the laser scan is processed,
- how reference points are established,
- how PhotoModeler is used,
- how lens distortion is treated,
- how camera position/orientation is calculated,
- what PhotoModeler exports,
- how that camera was historically imported into 3ds Max,
- how the undistorted image is used,
- how the 3D body figure is positioned,
- how stature was historically determined,
- what requirements exist for usable surveillance images,
- how multiple standing positions are handled,
- how shoes/headwear are treated,
- how tolerances are handled.
Create:
workflow_interpretation.md
Separate explicit documentation requirements from your own implementation
assumptions.
======================================================================
PHASE 2 - SELECT USABLE SURVEILLANCE FRAMES
======================================================================
If a complete video is supplied, identify candidate frames suitable for
stature reconstruction.
Prefer frames in which:
- the person is visible from head to foot,
- the feet/ground-contact area can be interpreted,
- image resolution is useful,
- occlusion is limited,
- the person is represented at a useful scale,
- perspective is suitable,
- body posture can reasonably be interpreted.
If several useful positions are available, use at least three.
Prefer frames showing different body orientations or poses because they
provide a stronger test of whether separate single-image fits recover compatible
body dimensions and stature.
Do not treat consecutive almost-identical frames as independent strong
constraints.
If fewer than three usable positions exist, continue if a technically
meaningful experimental reconstruction is possible but later report:
"Experimental reconstruction based on fewer than the minimum
number of evaluable standing positions described in the supplied
FOR workflow."
Record frame selection and rejection reasons in:
frame_selection.md
======================================================================
PHASE 3 - AUTONOMOUS CAMERA PARAMETER RECONSTRUCTION
======================================================================
The AI agent is responsible for reconstructing the operational camera model.
Do not import or copy a precomputed camera solution from PhotoModeler, 3ds Max,
a MaxScript export, or another external reconstruction package.

Reconstruct the camera from the available case evidence. Where available, use:
- the marked reference photograph,
- the metric laser-scan-derived scene,
- known 3D reference-marker coordinates,
- 2D marker locations observed in the reference photograph,
- image dimensions and raster geometry,
- legitimate acquisition metadata such as sensor dimensions, nominal focal length,
  lens information and pixel aspect.

The AI must autonomously:
1. identify the usable reference targets in the image and 3D scene,
2. establish and verify 2D-to-3D correspondences,
3. reject or flag ambiguous or inconsistent correspondences,
4. determine the appropriate camera and distortion model,
5. implement and run the numerical camera reconstruction,
6. inspect residuals and diagnostic overlays,
7. revise the correspondence set or camera model only for documented geometric
   reasons,
8. freeze the accepted camera parameters before any body fitting begins.

Use deterministic Python for the numerical solution wherever practical. Suitable
methods may include PnP, DLT-based initialization, RANSAC/outlier-resistant
initialization, nonlinear least-squares refinement and bundle-style optimization.
The exact method is an implementation choice, but it must be documented and
reproducible.

Estimate or determine, as supported by the available evidence:
- camera position / optical center in the metric scene,
- camera orientation,
- focal length or focal lengths in pixel units,
- principal point,
- pixel aspect and skew where relevant,
- radial lens-distortion coefficients,
- tangential lens-distortion coefficients where supported,
- the mapping between original/distorted and operational/undistorted rasters,
- any crop, scale or image-center shift introduced during rectification.

Do not assume that every camera parameter is identifiable from a single reference
image. Use legitimate acquisition metadata as fixed values or documented priors
when available. If the evidence does not support reliable estimation of a
parameter, do not hide that limitation by unconstrained optimization. Fix it to a
defensible value derived from metadata or explicitly report that the camera model
cannot be uniquely established from the supplied material.

If the reference image is original/distorted, estimate the distortion model as
part of the autonomous camera-calibration process when the data support it, then
generate and preserve the exact rectification/undistortion mapping used for
subsequent reconstruction. If the supplied operational images are already
undistorted, determine the corresponding pinhole-camera mapping and DO NOT apply
lens-distortion correction a second time.

The camera-reconstruction objective must be based on observed 2D reference-target
positions and their corresponding metric 3D coordinates. Minimize image-space
reprojection error with an appropriate robust loss when justified. Preserve the
raw and accepted correspondence sets so that rejected observations remain
auditable.

Create at minimum:
- camera_correspondences.json
- camera_calibration.json
- camera_conversion.md
- camera_mapping.json

camera_correspondences.json must record for every target:
- target ID,
- 3D metric coordinate,
- observed 2D image coordinate,
- confidence/visibility,
- accepted/rejected status,
- rejection reason where applicable.

camera_calibration.json must record at minimum:
- camera center / translation,
- rotation representation and convention,
- intrinsic matrix or equivalent focal/principal-point parameters,
- distortion model and coefficients,
- image dimensions,
- sensor/metadata values used,
- parameters held fixed versus optimized,
- optimizer and robust-loss configuration,
- number of correspondences used,
- per-target reprojection residuals,
- mean, median and RMS reprojection error,
- maximum residual,
- any identified parameter-correlation or identifiability limitations.

Determine and document the exact coordinate-system relationship between the metric
scene and Blender. Pay particular attention to:
- metric units,
- handedness,
- axis orientation,
- camera forward axis,
- camera up axis,
- camera-to-world versus world-to-camera transforms,
- Euler rotation order if Euler angles are used,
- rotation-matrix conventions,
- image raster origin and axis directions.

Do not guess coordinate or rotation conventions. Use diagnostic projections and
small verification scripts where needed. The resulting Blender camera must be a
numerical representation of the autonomously reconstructed camera model, not a
manually adjusted visual approximation.
======================================================================
CAMERA VALIDATION IS MANDATORY
======================================================================
Before inserting or fitting any human body model:
VALIDATE THE CAMERA.
Project the metric 3D reference points and reconstructed scene geometry through
the autonomously solved camera model into the exact operational image raster.
Generate an overlay containing:
- undistorted evidence/reference image,
- projected 3D reference geometry,
- projected known scene features.
Quantify reprojection error where reference correspondences are
available.
Save:
01_camera_scene_validation.png
If the projected scene does not correctly align with the image:
DO NOT proceed to body fitting.
Investigate:
- coordinate-system mapping,
- rotation direction,
- camera translation,
- image dimensions,
- focal length,
- principal point,
- scene scale,
- transformations,
- wrong source file interpretation.
Fix the conversion.
Do NOT modify the human model to compensate for a bad camera.
======================================================================
LOCK THE CAMERA
======================================================================
Once the autonomously reconstructed camera has been solved, represented in
Blender and validated:
LOCK IT.
The calibrated camera is measurement geometry.
The body-fitting optimizer must NOT move or rotate the camera to improve
human-body fit.
The body-fitting optimizer must NOT alter focal length, principal point,
distortion parameters or field of view to improve human-body fit.
If a later problem suggests that the camera reconstruction is wrong, pause body
fitting and return to the camera-calibration stage. Re-examine the original 2D-3D
correspondences, metadata, model choice, coordinate conventions and residuals;
then solve, validate and relock the camera again. Do not tune camera parameters
against the human-body overlay.
Never jointly optimize:
camera + person
merely to obtain a visually attractive overlay.

If the selected frames were acquired with the same physical camera position,
orientation, lens and raster geometry, reconstruct/validate that camera ONCE,
lock it, and reuse the identical camera solution for every selected frame.
Do NOT recalibrate or independently reconstruct the camera for each frame merely
because the body pose changes. If acquisition metadata show that a frame came
from a genuinely different camera geometry, treat that frame with its own
validated locked camera and document the difference explicitly. Body fitting
remains separate for each selected image under this protocol.
======================================================================
CAMERA / IMAGE-MAPPING REPORTING - MANDATORY
======================================================================
For every camera solution, explicitly document:
- whether the operational evidence raster is original/distorted or undistorted,
- image width/height and any crop/scale operation,
- focal length in pixel units,
- principal point,
- pixel aspect and skew assumptions,
- radial and tangential distortion model and coefficients,
- world-to-camera versus camera-to-world convention,
- coordinate axes and handedness,
- the exact mapping used for Blender overlays,
- whether any displayed background is a derived pinhole image,
- how double-undistortion is prevented.

Do not compare two camera solutions merely because both render to the same raster.
A projection comparison must state whether both projections include distortion,
use the same rectification/cropping convention, and represent the same physical
image coordinates. Save these details to camera_mapping.json and include them in
camera_conversion.md.

======================================================================
INDEPENDENT CAMERA COMPARISON - OPTIONAL EVALUATION
======================================================================
If an established-workflow PhotoModeler camera solution is available as an
independent comparator, it must remain inaccessible to the primary camera-solving
process and must not be used to initialize, tune, select or reject the autonomous
camera reconstruction. Only after the autonomous camera has been frozen may the
solutions be compared transparently.
Report at minimum:
- camera-center difference in a stated coordinate system,
- rotation difference,
- focal/principal-point/distortion differences,
- the exact distortion/rectification convention for each model,
- projected displacement of common 3D reference points on a genuinely common
  image coordinate system.

Do not call the displacement between two camera models a reprojection residual
against ground truth unless independent observed 2D target positions are used.
If a first-order scale/depth calculation is reported, label it illustrative and
state which camera differences it excludes. A controlled camera-substitution test
through the body-fitting pipeline is required to quantify the total camera effect
on stature.

======================================================================
PHASE 4 - CREATE A PARAMETRIC FORENSIC HUMAN RIG
======================================================================
Blender does not provide the historical 3ds Max Biped used in the
existing workflow.
Therefore create a dedicated PARAMETRIC FORENSIC HUMAN RIG.
This is not primarily an artistic character.
It is a measurable articulated anatomical mannequin designed for
stature reconstruction.
It must consist of:
1. an internal hierarchical anatomical skeleton / armature,
2. an external human-shaped body surface attached to that rig,
3. explicit anatomical measurement landmarks,
4. constrained morphology parameters,
5. constrained pose parameters.
======================================================================
MANDATORY INTERNAL SKELETON
======================================================================
The skeleton is the PRIMARY geometric representation of:
- body segment lengths,
- body proportions,
- joint centers,
- articulated pose,
- reconstructed anatomical stature.
The visible body mesh is secondary.
It must remain registered to the skeleton.
Create explicit structures representing at least:
PELVIS / TRUNK
- root / pelvis
- sacral/lower-spine region
- lumbar spine
- thoracic spine
- cervical spine
- neck
- head / skull
- explicit anatomical top-of-head landmark
LEFT LEG
- left hip joint center
- left femur
- left knee joint center
- left tibia/lower leg
- left ankle joint center
- left foot
- left heel landmark
- left toe/forefoot landmark
RIGHT LEG
- equivalent right-side structures
LEFT ARM
- shoulder joint center
- upper arm
- elbow joint center
- forearm
- wrist
- hand
RIGHT ARM
- equivalent right-side structures.
Every important joint center and endpoint must have explicit queryable
3D coordinates.
Use a proper armature hierarchy.
Do not create disconnected decorative primitives.
======================================================================
BODY MESH / RIG CONNECTION
======================================================================
Create a visible, simplified but anatomically interpretable human body
mesh around the skeleton.
The external geometry should be sufficient to compare:
- head outline,
- shoulders,
- trunk,
- pelvis,
- arms,
- legs,
- feet,
with the surveillance image.
Do not waste effort on irrelevant artistic detail.
The body surface must be rigged/skinned to the skeleton.
Changing pose must move the associated surface.
Changing segment lengths must update both:
- skeleton,
- corresponding body surface.
Never stretch the mesh independently from the skeleton merely to improve
the camera overlay.
A visually good surface fit with an inconsistent skeleton is invalid.
======================================================================
MORPHOLOGY AND POSE MUST BE SEPARATE
======================================================================
Maintain two independent parameter classes.
MORPHOLOGY PARAMETERS may include:
- total anatomical scale/stature,
- femur length,
- tibia/lower-leg length,
- foot dimensions,
- pelvis height/width/depth,
- torso length,
- torso depth,
- shoulder width,
- spine/trunk dimensions,
- neck length,
- head dimensions,
- upper-arm length,
- forearm length,
- hand length,
- appropriate body widths/thicknesses.
POSE PARAMETERS include:
- root/world position,
- body orientation,
- pelvic orientation,
- hip rotations,
- knee flexion,
- ankle orientation,
- spinal posture,
- shoulder rotations,
- elbow flexion,
- wrist orientation if relevant,
- neck rotation,
- head rotation.
Changing POSE must never change anatomical segment lengths.
Morphological scaling must be explicit and logged.
======================================================================
BILATERAL SYMMETRY CONSTRAINTS
======================================================================
By default, paired anatomical segment lengths must remain equal or
appropriately symmetric.
For example:
left femur length == right femur length
left tibia length == right tibia length
left upper-arm length == right upper-arm length
left forearm length == right forearm length
left foot length approximately == right foot length
left hand length approximately == right hand length
If one femur is lengthened during morphology optimization, the
contralateral femur must receive the corresponding change.
POSE is allowed to be asymmetric.
ANATOMICAL LENGTHS must not become arbitrarily asymmetric.
Do not use left/right length differences as an easy way to improve a
2D fit unless the evidence explicitly demonstrates a real anatomical
asymmetry.
======================================================================
ANATOMICAL PROPORTION CONSTRAINTS
======================================================================
Body proportions must remain compatible with a plausible human body.
Do not allow the optimizer to create, for example:
- extremely long femora with tiny tibiae,
- extremely short torso with excessive leg length,
- impossible shoulder width,
- unrealistic head dimensions,
- arbitrary local limb scaling.
If validated anthropometric limits are supplied in the forensic/QM
material, use them.
If they are NOT supplied:
use only broad conservative human anatomical constraints as
regularization.
Clearly label those constraints as:
"experimental implementation assumptions based on generic human
anatomy; not derived from the supplied FOR methodology."
Do not force the person toward an "average" human body.
Morphology is unknown and must be allowed to deviate from population
averages when the evidence supports it.
Anthropometric information is a plausibility constraint, NOT the source
of the stature estimate.
======================================================================
JOINT CONSTRAINTS
======================================================================
Implement physiologically meaningful joint limits for at least:
- hips,
- knees,
- ankles,
- shoulders,
- elbows,
- cervical/head orientation,
- torso/spine.
Use Blender constraints, custom logic, IK/FK, or numerical bounds as
appropriate.
Prevent:
- impossible knee bending,
- implausible hyperextension,
- impossible hip rotations,
- dislocated shoulder configurations,
- anatomically absurd spinal configurations.
A 2D match obtained through impossible joint configurations is invalid.
======================================================================
RIG SELF-TEST BEFORE CASE FITTING
======================================================================
Before using the rig on evidence, automatically test it.
Create Python tests verifying at least:
- left/right femur equality,
- left/right tibia equality,
- left/right upper-arm equality,
- left/right forearm equality,
- pose rotations do not alter bone lengths,
- changing linked segment length changes the contralateral segment,
- body mesh follows rig correctly,
- joint limits operate,
- measurement landmarks follow morphology correctly,
- neutral-pose stature is reproducible,
- changing pose does not change pose-independent stature.
Use automated assertions where possible.
If a test fails:
fix the rig before continuing.
Save test results to:
rig_validation.txt
======================================================================
LEFT/RIGHT AND VIEW-LABEL CONVENTIONS
======================================================================
Define and test separately:
- anatomical left/right,
- rig-internal left/right names,
- camera/viewer left/right,
- validation-camera naming.

If a legacy rig uses a reversed internal convention, encode the mapping once in
code and unit-test it. Generate validation-view labels from the defined camera
transform/convention rather than from informal screen position. Include the mapping
in fit_config.json and report it so that a left/right label cannot silently invert.

======================================================================
PHASE 5 - IMAGE OBSERVATIONS / LANDMARKS
======================================================================
Inspect every selected undistorted evidence frame.
Identify visible anatomical landmarks where possible.
Candidate landmarks include:
- visible head top,
- approximate anatomical head top if distinguishable,
- head center,
- neck,
- left shoulder,
- right shoulder,
- elbows,
- wrists/hands,
- pelvis/hip region,
- knees,
- ankles,
- heels,
- toes,
- sole/ground-contact regions.
For each observation record:
- frame ID,
- image x,
- image y,
- anatomical label,
- confidence,
- visibility/occlusion,
- notes.
Use confidence classes such as:
HIGH
MEDIUM
LOW
OCCLUDED
UNKNOWN
Save to:
target_landmarks.json
Do not invent precise joint positions when they are not visible.
Do not treat a hidden elbow, knee or ankle as precisely observed.
If clothing is loose, distinguish clothing outline from the probable
underlying anatomical position.
Use silhouette information as additional evidence where appropriate,
but do not allow loose clothing to dominate skeletal fitting.
======================================================================
PHASE 6 - USE THE 3D SCENE TO ESTABLISH FOOT / GROUND RELATION
======================================================================
Do not determine body depth merely by visually sliding the mannequin
until it overlaps the image.
Use the calibrated evidence camera and reconstructed 3D scene.
Where foot or sole locations are visible:
cast rays from the camera through corresponding image locations into
the laser-scan-derived geometry.
Determine plausible local ground-contact regions.
Use the actual local scene geometry.
Do not assume the entire scene is one perfect horizontal floor if the
scan provides more accurate information.
Initialize the feet near the resulting support regions.
Feet must not:
- float substantially above the floor,
- penetrate deeply through the floor,
- stand on impossible geometry,
without explicit evidence justification.
If ground contact is uncertain or occluded, retain a range of plausible
solutions instead of inventing an exact point.
======================================================================
PHASE 7 - INITIALIZE THE UNKNOWN PERSON
======================================================================
Start from a generic plausible adult-human morphology.
DO NOT initialize stature from any example result in the documentation.
Position the rig in the calibrated 3D scene using:
- image evidence,
- camera geometry,
- ground contacts,
- visible body landmarks.
Initialize approximately in the following order:
1. feet / support positions,
2. pelvis/root,
3. lower limbs,
4. torso,
5. head,
6. arms.
Do not begin from a random highly distorted body and then optimize all
parameters simultaneously.
======================================================================
INDEPENDENT SINGLE-IMAGE RECONSTRUCTION - CRITICAL REQUIREMENT
======================================================================
All selected frames may show the SAME physical person. Fit each selected frame as
a fresh single-image reconstruction in order to measure reconstruction stability.

SHARED ACROSS IMAGE BRANCHES:
- the validated camera lock where acquisition geometry is identical,
- the metric scan/scene and its coordinate system,
- the canonical rig implementation,
- the generic starting morphology,
- the software/code base,
- the predefined broad anatomical bounds,
- the predefined objective structure and default weights,
- the same reporting and acceptance rules.

MUST NOT BE TRANSFERRED FROM ONE IMAGE BRANCH TO ANOTHER:
- fitted morphology or segment lengths,
- fitted pose or world placement,
- final or intermediate stature values,
- accepted/rejected candidate geometry,
- per-image landmark corrections unless independently re-observed in that image,
- case-specific regularizers introduced because of another image,
- the cross-image mean or any other combined statistic.

For each image i create a separate branch/directory, for example:
image_2295/
image_2336/
image_2353/
image_2415/

Each branch must begin from:
1. the same canonical generic morphology,
2. a fresh pose/placement initialization appropriate to that image,
3. a fresh optimizer state,
4. the locked camera and metric scene,
5. observations made from that image itself.

Do not open or query a previous image's fitted_case.blend, measurements.json,
accepted body parameters or final height before the current branch has been
frozen. Shared code and camera geometry are permitted; shared fitted body results
are not.

At the end of each image branch:
- save the final scene and all checkpoints,
- save the exact observations and constraint revisions,
- save fit_config.json,
- save the final measurements,
- save an image-branch SHA-256 manifest,
- mark the branch FROZEN before proceeding to cross-image comparison.

Only AFTER all selected image branches are frozen may you compare their body
dimensions or stature estimates. Because the physical person is the same,
disagreement between fitted morphologies is evidence of model/pose/observation
instability, not biological variation.

If the same AI session retains prior-image information in its context or accessible
logs, do not call the runs statistically or contextually independent. State:
"The image fits were procedurally independent in their numerical initialization
and body parameters, but complete information isolation between agent decisions
was not demonstrated."

Never use cross-image disagreement as a reason to tune any already frozen branch
toward the others.
======================================================================
PHASE 8 - STAGED CONSTRAINED OPTIMIZATION
======================================================================
Treat EACH IMAGE BRANCH as its own constrained inverse-projection problem.
Do NOT simply "eyeball" the model until it looks close.
Use numerical optimization where possible.
Python/scipy may be used if available.
Otherwise implement an appropriate optimization method.
Use the same staged structure for every image so that cross-image comparison is
interpretable.

A practical per-image strategy is:
1. initialize the canonical generic morphology,
2. fit global placement with morphology fixed,
3. fit feet/lower-body pose with morphology fixed,
4. fit torso/head pose with morphology fixed,
5. fit upper-limb pose with morphology fixed,
6. release morphology only after a plausible pose branch exists,
7. jointly refine that image's morphology and pose,
8. perform multi-view plausibility and collision/support checks,
9. freeze the branch,
10. only later compare the frozen results across images.

The objective for image i is independent of fitted results from every other image.
The cross-image mean, reference stature and previous body dimensions must never
enter the objective.

Use the same default objective terms, normalization and weights across images.
If a case-specific pose regularizer is introduced because an underdetermined 3D
branch is visibly implausible, log the observation, rationale, exact parameter
change and effect. Such revisions are model assumptions, not new evidence.

----------------------------------------------------------------------
STAGE A - GLOBAL / PER-IMAGE PLACEMENT
----------------------------------------------------------------------
Optimize:
- global root X/Y/Z,
- global orientation/yaw,
while maintaining plausible floor relationship.

Initialize the same generic morphology/stature for every image branch.
Do not initialize from a previous reconstruction.
----------------------------------------------------------------------
STAGE B - LOWER BODY
----------------------------------------------------------------------
Optimize:
- foot contacts,
- ankle orientations,
- knee flexion,
- hip rotations,
- pelvis position/orientation.

Keep morphology fixed until the designated morphology stage.
----------------------------------------------------------------------
STAGE C - TORSO AND HEAD
----------------------------------------------------------------------
Optimize:
- pelvis orientation,
- spinal posture,
- shoulder line orientation,
- neck/head pose.

Keep torso, neck, shoulder-width and head dimensions generic until morphology is
explicitly released.
----------------------------------------------------------------------
STAGE D - UPPER LIMBS
----------------------------------------------------------------------
Optimize:
- shoulder pose,
- upper-arm pose,
- elbow flexion,
- forearm pose,
- wrists/hands where visible.

Keep upper-arm, forearm and hand dimensions generic until morphology is released.
----------------------------------------------------------------------
STAGE E - PER-IMAGE MORPHOLOGY / POSE REFINEMENT
----------------------------------------------------------------------
For THIS image only, refine:
- morphology parameters,
- pose parameters,
- contact/world-placement parameters where justified.

Do not constrain the result to resemble any previous image's fitted dimensions.
The same physical person should theoretically have the same morphology, but under
this experimental design disagreement is deliberately allowed to reveal instability.
======================================================================
OPTIMIZATION OBJECTIVE
======================================================================
Do NOT optimize only visual appearance.
For each image use an explicit objective containing appropriate components such as:
E_image =
  w1 * E_landmark_reprojection
+ w2 * E_silhouette
+ w3 * E_ground_contact
+ w4 * E_joint_limits
+ w5 * E_anatomical_proportions
+ w6 * E_bilateral_symmetry
+ w7 * E_environment_collision
+ w8 * E_self_intersection
+ w9 * E_3D_pose_plausibility
+ w10 * E_regularization

No term may contain the reference stature, a previous image's fitted stature, or
the cross-image mean.

MANDATORY REPRODUCIBILITY RECORD:
For every image branch save the ACTUAL implemented objective, not merely this
example. Record in fit_config.json:
- mathematical definition of every residual/penalty term,
- units and normalization of each term,
- all weights,
- robust-loss function and scale,
- parameter vector and parameter units,
- all lower/upper bounds,
- initialization values,
- optimizer/library and method,
- solver tolerances and stopping criteria,
- maximum evaluations/iterations,
- random seeds where applicable,
- finite-difference or Jacobian settings,
- all default pose/anatomy priors,
- every case-specific regularizer added or changed during review,
- software/library versions.

If a regularizer or bound changes after inspection, append a structured entry to
constraint_revision_log.jsonl containing:
- image ID and stage,
- observed problem,
- whether the observation comes from evidence or an unobserved validation view,
- old value,
- new value,
- rationale,
- effect on objective, geometry checks and stature,
- ACCEPT/REFINE/REJECT decision.

Strong image observations may dominate weak priors.
Low-confidence or occluded landmarks must have lower weight or be omitted.
Generic anthropometry must not overpower reliable evidence.
======================================================================
LANDMARK REPROJECTION ERROR
======================================================================
For each visible 3D anatomical landmark:
project its world coordinate through the LOCKED Blender evidence camera.
Compare the resulting image coordinate with the observed image landmark.
Record residuals.
Report at minimum:
- per-landmark residual,
- per-frame mean error,
- per-frame RMSE where appropriate,
- overall fit metric.
Do not hide high-error landmarks.
======================================================================
MANDATORY 3D PLAUSIBILITY VALIDATION
======================================================================
THIS IS A CRITICAL REQUIREMENT.
A reconstruction is NOT acceptable merely because the mannequin looks
correct from the surveillance camera.
The fitted body must also make anatomical and spatial sense in true 3D.
The evidence camera supplies observational evidence.
Front, side, rear and oblique validation views are NOT additional
evidence.
They are internal consistency and plausibility checks.
The question is:
"Does this 3D human configuration explain the evidence image while
also remaining a plausible human body when viewed from directions
that were not observed?"
======================================================================
MANDATORY VALIDATION VIEWS
======================================================================
After EVERY MAJOR fitting cycle:
render or inspect at least:
1. evidence camera,
2. orthographic frontal view,
3. orthographic left-side view,
4. right-side or rear-oblique view,
5. useful perspective/oblique view,
6. top view when depth ambiguity requires it.
Also generate transparent-body versions showing the skeleton.
======================================================================
FRONT / SIDE / 3D SANITY CHECKS
======================================================================
Explicitly inspect:
HEAD / NECK
----------------------------------------------------------------------
- Is the head plausibly connected to the neck?
- Is neck length plausible?
- Is head orientation possible?
- Is the top-of-head landmark sensible?
TORSO / PELVIS
----------------------------------------------------------------------
- Is the torso at a plausible depth relative to pelvis?
- Has the torso been pushed toward or away from the camera merely to
improve the 2D overlay?
- Is pelvic tilt plausible?
- Is spinal posture realistic?
- Are thorax and pelvis spatially coherent?
LEGS
----------------------------------------------------------------------
- Do hip-knee-ankle chains form plausible 3D geometry?
- Are knees at plausible depth?
- Are joint angles anatomically possible?
- Do feet correspond plausibly to support surfaces?
- Could the legs physically support the observed pose?
ARMS
----------------------------------------------------------------------
- Are shoulders, elbows and wrists in plausible 3D locations?
- Has an arm been placed far in front of or behind the torso simply
because its 2D projection matches?
- Are elbow and shoulder angles plausible?
WHOLE BODY
----------------------------------------------------------------------
- Could a human physically adopt this configuration?
- Does the pose remain believable from the side?
- Is severe self-intersection present?
- Is body balance broadly plausible for the apparent activity?
- Are body depths realistic?
- Are body proportions coherent?
ENVIRONMENT
----------------------------------------------------------------------
- Does the person plausibly occupy the reconstructed space?
- Are feet related correctly to the floor?
- Does the body pass impossibly through reconstructed objects?
- Are apparent contacts consistent with the image?
======================================================================
IMPORTANT LIMITATION OF SIDE VIEWS
======================================================================
DO NOT pretend that an unobserved side view is known.
The surveillance camera does not directly provide all depth information.
Therefore do not judge side-view correctness against an imaginary
unknown photograph.
Instead judge:
- anatomy,
- kinematic constraints,
- body proportions,
- joint geometry,
- scene intersections,
- floor contact,
- biomechanical plausibility.
The side/front views are a test for impossible 3D solutions, not invented
additional evidence.
======================================================================
HARD FAILURE RULE
======================================================================
A solution that fits the evidence-camera view but fails frontal,
side-view or general 3D plausibility inspection is INVALID.
Reject and re-optimize it.
Examples of failure include:
- good evidence overlay but impossible leg geometry,
- elbow floating far behind the torso without justification,
- knee located implausibly relative to hip and foot,
- severe unnatural torso lean created only for projection matching,
- impossible spinal configuration,
- arbitrary asymmetric limb lengths,
- feet floating above the scene,
- feet deeply penetrating the floor,
- skeleton no longer matching body surface,
- implausible overall body proportions.
DO NOT report stature from such a solution as a successful
reconstruction.
======================================================================
BOUND / IDENTIFIABILITY GATE
======================================================================
After morphology fitting and again after final refinement, report every parameter
at or numerically near an allowed bound. Distinguish:
- a merely active bound with otherwise plausible geometry,
- multiple active bounds indicating weak identifiability,
- a visibly or kinematically implausible proportion pattern that triggers the hard
  failure rule.

A geometry/collision PASS does not override implausible anatomy. Conversely, do
not invent tighter anthropometric limits solely to force agreement with an expected
height. If a conditional candidate is retained despite several active bounds, say
so prominently and do not call its segment dimensions accurate anatomy.

======================================================================
ITERATIVE FIT-INSPECT-CORRECT LOOP
======================================================================
For each meaningful optimization cycle:
1. optimize against evidence observations,
2. update the live Blender scene,
3. render evidence-camera overlay,
4. inspect evidence-camera fit,
5. render front view,
6. render side view,
7. inspect 3D plausibility,
8. inspect ground contact,
9. inspect skeleton/body consistency,
10. identify the cause of any problem,
11. modify only the appropriate parameter class,
12. re-optimize,
13. repeat.
Do not stop after the first approximately convincing overlay.
Continue until further improvements in image fit would require
unreasonable anatomical distortion or until convergence is reached.
======================================================================
MULTIPLE INITIALIZATIONS / LOCAL MINIMA
======================================================================
The reconstruction may be underdetermined.
For each final image branch, test at least three fresh generic starting
configurations unless a technical failure makes this impossible; document any
exception.
Do not assume that the first numerical minimum is the correct 3D
solution.
Compare candidate solutions using:
- image reprojection,
- anatomical plausibility,
- ground contact,
- agreement with the observations in that image,
- proportion plausibility.
Retain competing plausible solutions if they cannot be distinguished
from the evidence.
======================================================================
FORENSIC SKELETON-BASED STATURE DETERMINATION
======================================================================
The final stature determination must be based on the fitted anatomical
rig.
Do NOT calculate final stature simply from image pixel height.
Do NOT calculate final stature merely as the current vertical image-pose
distance between floor and head.
The skeleton provides the reconstruction's metric body geometry.
======================================================================
MANDATORY STATURE MEASUREMENT LANDMARKS
======================================================================
Create explicit 3D landmarks for at least:
- heel/support reference,
- ankle,
- knee,
- hip joint center,
- pelvis,
- spinal axis landmarks,
- neck,
- skull/head,
- anatomical top of head.
These landmarks must remain tied to the rig.
Export their final coordinates.
======================================================================
SKELETAL MEASUREMENT CHAIN
======================================================================
Provide an auditable skeletal measurement chain approximately following:
heel
-> lower leg / knee
-> femur / hip
-> pelvis
-> trunk / spine
-> neck
-> head
-> anatomical top of head
Calculate and report the relevant fitted segment dimensions.
Do this independently for left/right lower limbs where relevant.
The chain is intended to make the anatomical basis of the reconstructed
height inspectable by a human expert.
======================================================================
POSE-INDEPENDENT NEUTRAL STATURE
======================================================================
A bent or inclined pose must not artificially reduce the reconstructed
body stature.
After the evidence fitting is complete:
CREATE A COPY of the fitted rig.
Do NOT rescale it.
Do NOT modify any morphology parameter.
Do NOT refit it to the image.
Only neutralize pose rotations to create a standardized upright
measurement configuration.
Use:
- the same femur lengths,
- same tibia lengths,
- same pelvis,
- same torso/spine dimensions,
- same neck,
- same head dimensions,
- same complete morphology.
Place the neutralized body on a common support plane.
Straighten the legs within normal anatomical configuration.
Orient pelvis/trunk/head neutrally.
Measure the vertical distance from support level to the anatomical
top-of-head landmark.
Call this:
reconstructed_anatomical_stature
For the current image branch, this is the primary skeleton-derived pose-independent stature estimate.
======================================================================
SECONDARY HEIGHT VALUES
======================================================================
Also calculate separately:
A) posed_vertical_height_i
= current vertical extent in each frame-specific fitted evidence pose
B) reconstructed_anatomical_stature
= one neutralized skeleton-based stature for each frozen single-image morphology
C) skeletal_chain_dimensions
= one auditable set of segment/chain measurements per image branch, with bilateral
  values where relevant
Do not silently substitute one for another.
Investigate unexpectedly large inconsistencies.
======================================================================
FOOTWEAR / HEADWEAR
======================================================================
Distinguish:
- external visible reconstructed height including footwear/headwear,
- anatomical body stature excluding those contributions.
Do NOT copy footwear/headwear correction values from an example in the
workflow documentation.
Only apply a correction if:
- the current case provides justified information,
or
- the supplied QM methodology explicitly defines a generally applicable
procedure that is appropriate for this case.
Otherwise report the uncorrected result and state that the correction
cannot be determined reliably from the supplied evidence.
======================================================================
POST-RUN REFERENCE MEASUREMENT AND BENCHMARK
======================================================================
This section applies ONLY after reconstruction branches have been frozen and the
blind-reconstruction manifest has been created.

If a measured reference stature is supplied for evaluation, record available
metadata in reference_measurement.json without inventing missing information:
- instrument/type of physical measuring scale,
- resolution and calibration status if known,
- number of repeat measurements,
- posture and support condition,
- head-position convention if specified,
- treatment of hair,
- footwear worn and whether it matches the recorded footwear,
- measurement date/time if available,
- assessor/observer role if available.

Do not imply millimetric physical accuracy merely because software outputs many
decimal places. Distinguish numerical output precision from measurement accuracy.

If a laboratory benchmark such as +/-3 cm is used, record:
- the source document and version/date,
- whether it is an empirical error range, uncertainty statement, reporting rule or
  pragmatic acceptance threshold,
- the population/cases to which it applies, if documented,
- the exact predefined decision rule for this experiment: per-image, combined
  descriptive mean, or both.

If the benchmark source or meaning cannot be documented, do not describe the
criterion as validated or predefined without qualification.

======================================================================
UNCERTAINTY AND SENSITIVITY
======================================================================
Do not report false precision.
Do not call the between-image sample SD a validated forensic uncertainty or a
confidence interval.

For EACH image branch investigate sensitivity to plausible variations in:
- uncertain landmark placement,
- foot contact,
- head-top localization,
- partially hidden joints,
- pose ambiguity,
- morphology,
- depth ambiguity,
- alternative generic initializations.

MINIMUM COMPUTATIONAL CHECKS PER IMAGE, unless technically impossible:
A) at least 3 fresh generic initialization restarts;
B) at least 6 landmark-localization perturbation runs based on the recorded
   localization scales/uncertainty assumptions.

For every alternative candidate:
- preserve the exact perturbed observations and random seed,
- rerun the relevant fitting stages without using the reference stature,
- inspect actual saved geometry, evidence overlay, side/oblique/top plausibility,
  support and collisions,
- mark ACCEPTED or REJECTED with a specific reason,
- retain rejected outputs for audit.

Report per image:
- number attempted, accepted and rejected,
- generic-restart stature range,
- accepted localization-perturbation stature range,
- rejection reasons,
- the main fit's position relative to accepted alternatives,
- active morphology bounds for every accepted candidate.

If no localization perturbation survives the acceptance gates, state explicitly:
"No accepted localization sensitivity interval exists for this image."
Do not replace this missing information with the generic-restart spread.

If accepted perturbations produce stature values far from the main fit, report the
full range prominently. Numerical repeatability under identical observations is not
evidence of robustness to observation uncertainty.

After all four main image branches are frozen, calculate cross-image descriptive
statistics (arithmetic mean, sample SD, range, signed differences) and compare
fitted body dimensions. Because the images show the same person, large differences
in femur, tibia, trunk, pelvis, shoulder, arm or other dimensions are diagnostic
of unresolved pose/depth/model ambiguity.

ACTIVE-BOUND REPORTING IS MANDATORY:
For every main fit list all morphology parameters within numerical tolerance of a
lower or upper bound. A small spread caused by repeated boundary solutions must
not be interpreted as reliable anatomical recovery. If multiple dimensions are
bound-limited or the neutralized body is visibly disproportionate, label the
result weakly identified and explain whether it remains only a conditional
experimental candidate or must be rejected by the hard plausibility gate.

Use wording such as:
- "computational sensitivity",
- "accepted perturbation range",
- "descriptive between-image variation".
Do NOT use "confidence interval" or "validated measurement uncertainty" unless
a separately validated forensic uncertainty model supports that interpretation.
======================================================================
LIVE REVIEW REQUIREMENTS
======================================================================
During the task, allow a forensic expert watching Blender to inspect
the current state.
At meaningful checkpoints show:
- evidence camera overlay,
- rig in true 3D scene,
- front view,
- side view,
- skeleton inside body,
- foot-ground relationship.
Do not intentionally hide intermediate failures.
If a solution becomes geometrically absurd during optimization, allow
that to remain inspectable until corrected.
The goal is traceability, not presentation polish.
======================================================================
CHECKPOINTS
======================================================================
Save intermediate .blend files at important milestones, for example:
checkpoint_01_camera_validated.blend
checkpoint_02_rig_created.blend
checkpoint_03_initial_placement.blend
checkpoint_04_lower_body_fit.blend
checkpoint_05_full_pose_fit.blend
checkpoint_06_final_candidate.blend
Do not overwrite all reconstruction history with a single final file.
======================================================================
FINAL BLENDER SCENES
======================================================================
Each image branch must retain a fully editable final .blend scene.
Each must include:
- the locked calibrated evidence camera,
- laser-scan-derived scene geometry,
- the independently fitted unknown-person rig for that image,
- body mesh,
- skeleton,
- measurement landmarks,
- validation cameras,
- evidence image overlay capability,
- the accepted fitted pose,
- a morphology-identical neutral measurement pose/copy,
- all relevant measurement helpers.

Do not merge the four fitted morphologies into a single shared body after fitting.
The cross-image differences are experimental results and must remain inspectable.
Use meaningful object and collection names.
======================================================================
MANDATORY OUTPUT DIRECTORY
======================================================================
Create a structured result directory containing at least:

00_case_level/
  input_hashes.txt
  workflow_interpretation.md
  camera_correspondences.json
  camera_calibration.json
  camera_conversion.md
  camera_mapping.json
  camera_lock.json
  camera_validation.json
  frame_selection.md
  interaction_log.jsonl
  protocol_deviations.json
  software_environment.txt
  master_prompt_used.txt
  master_prompt_sha256.txt
  cross_image_summary.json
  cross_image_measurements.csv
  delivery_manifest_sha256.txt

image_<ID>/ for EACH selected image:
  fitted_case.blend
  fit_results.json
  skeleton_measurements.json
  target_landmarks.json
  fit_config.json
  constraint_revision_log.jsonl
  workflow_log.md
  rig_validation.txt
  checkpoints/
  sensitivity/
  renders/
  branch_manifest_sha256.txt

Also capture the actual runtime environment where practical:
- Blender version/build,
- Python version,
- SciPy/NumPy versions,
- relevant package versions,
- operating-system information,
- model/agent name and reasoning setting as exposed by the environment,
- tool interfaces actually used.
======================================================================
MANDATORY RENDERS
======================================================================
Generate at minimum:
01_camera_scene_validation.png
For each evidence frame:
02_frame_XX_evidence.png
03_frame_XX_final_overlay.png
04_frame_XX_transparent_skeleton_overlay.png
For EACH image branch, final 3D plausibility views:
05_frame_XX_final_front_view.png
06_frame_XX_final_front_transparent_skeleton.png
07_frame_XX_final_left_side_view.png
08_frame_XX_final_left_side_transparent_skeleton.png
09_frame_XX_final_right_side_or_rear_view.png
10_frame_XX_final_oblique_view.png
11_frame_XX_final_ground_contact_view.png
12_frame_XX_final_measurement_landmarks.png
13_frame_XX_neutral_pose_stature_measurement.png
14_frame_XX_final_top_view.png
Make the renders sufficiently clear for expert review.
======================================================================
SKELETON_MEASUREMENTS.JSON
======================================================================
Export at least:
- reconstructed_anatomical_stature (one value for this image branch)
- posed_vertical_height for this image
- left femur length
- right femur length
- left tibia length
- right tibia length
- left upper-arm length
- right upper-arm length
- left forearm length
- right forearm length
- pelvis dimensions
- torso/spine dimensions
- neck dimensions
- head dimensions
- heel coordinates
- ankle coordinates
- knee coordinates
- hip coordinates
- pelvis coordinates
- spinal landmarks
- head landmark
- top-of-head landmark
- all morphology parameters
- all fitted joint rotations for this image
Store measurements in SI units.
Also provide stature in centimeters.
======================================================================
FIT_RESULTS.JSON
======================================================================
Include:
- source image ID and source hash,
- camera-lock ID/hash,
- camera validation metrics,
- final morphology parameters for this image branch,
- fitted pose parameters for this image,
- per-landmark residuals for this image,
- overall reprojection metric,
- ground-contact residuals where measurable,
- active joint constraints,
- active anthropometric constraints,
- candidate solution scores,
- computational sensitivity results,
- final stature candidate.
======================================================================
FINAL FORENSIC-TECHNICAL SUMMARY
======================================================================
At completion provide a concise written summary answering:
1. Was the camera successfully reconstructed, mapped and validated, and was the
   same locked solution reused where acquisition geometry was identical?
2. Which evidence images were reconstructed?
3. Did every image branch begin from the same canonical generic morphology and a
   fresh optimizer/pose state?
4. Were fitted body dimensions, fitted poses and prior image heights prevented
   from being imported into later image branches?
5. Was complete context isolation between agent decisions demonstrated, or only
   procedural independence?
6. For each image, was the resulting skeleton anatomically and spatially plausible
   in evidence, frontal, lateral, oblique and top views?
7. For each image, were foot-ground relationships plausible?
8. What external support-to-head / skeleton-derived stature was obtained for each
   image, and what endpoint limitations apply (hair, footwear, model crown)?
9. Which morphology parameters hit bounds in each image?
10. What did the generic-restart and landmark-localization sensitivity tests show,
    including accepted/rejected counts and accepted ranges?
11. What are the arithmetic mean, sample SD and full range across the frozen
    single-image results, clearly labeled descriptive rather than uncertainty?
12. Which fitted body dimensions disagree most strongly across images of the same
    person?
13. What assumptions were introduced that are NOT contained in the supplied FOR
    workflow?
14. What uncertainties or ambiguities could materially alter the result?
15. Did any user interaction exceed a continuation-only prompt? If yes, where and
    how did it affect the workflow?
16. Was independent human expert review completed?

Do not claim independent human expert review unless it actually occurred and is
documented outside the agent's own ACCEPT/REFINE/REJECT decisions.
======================================================================
MANDATORY FINAL PLAUSIBILITY DECISION
======================================================================
Before treating the stature as a valid candidate result, explicitly answer:
"Is the fitted body configuration anatomically and spatially
plausible not only from the evidence camera, but also when
inspected in frontal, lateral and oblique 3D views?"
Answer:
YES - PLAUSIBLE
or
NO - NOT SUFFICIENTLY PLAUSIBLE
If NO:
do not present the stature as a trustworthy reconstruction result.
Explain the failure.
Continue fitting if there is a technically reasonable path to a better
solution.
======================================================================
MANDATORY EXPERT-REVIEW DECISION
======================================================================
Finally answer:
"Would you consider this reconstruction sufficiently geometrically
and anatomically consistent to submit for independent forensic
expert review?"
Answer:
YES
or
NO
and provide a brief technical reason.
This is NOT a statement of forensic admissibility or validation.
It is only a technical assessment of whether the generated reconstruction
is coherent enough to warrant expert examination.
======================================================================
AI REVIEW VS INDEPENDENT HUMAN REVIEW
======================================================================
The agent's visual/technical inspection is an AI-mediated plausibility review. It
is NOT an independent human expert review and must never be described as such.
The final report must explicitly state one of:
- "Independent human expert review completed: NO - required before forensic use"
  unless a separate documented human review has actually occurred, or
- the documented identity/role and date of the human review if it did occur.

Do not use the words "validated", "forensically validated" or "ground truth"
for the body reconstruction merely because software checks, overlays or AI review
passed.

======================================================================
AUTONOMOUS BEHAVIOR
======================================================================
Work autonomously.
Do not ask me to perform routine Blender operations.
Do not ask me to manually move bones for you.
Do not ask me to choose ordinary implementation details.
Use:
- Python,
- Blender,
- bpy,
- numerical calculations,
- diagnostic scripts,
- renders,
- camera overlays,
- 3D validation views,
- logs,
to investigate problems yourself.
When something fails:
diagnose it.
Do not merely try random visual changes.
Determine whether the problem originates from:
- camera conversion,
- scene scale,
- coordinate systems,
- image mapping,
- body placement,
- body morphology,
- pose,
- ground contact,
- anatomy,
- optimization,
- occlusion.
Fix the correct cause.
Only stop to ask me for input if an indispensable piece of case data is
missing and proceeding would require an unjustified forensic assumption.
Do not reduce verification effort in order to finish faster.
Correctness and repeated geometric validation are more important than token
usage or execution time.
======================================================================
CORE RULES - NEVER VIOLATE THESE
======================================================================
1. DO NOT change the calibrated camera to make the body fit.
2. DO NOT distort the laser-scan scene to make the body fit.
3. DO NOT independently stretch left/right paired limbs within a single body merely
   to improve a 2D overlay.
4. DO NOT allow paired left/right anatomical segment lengths to diverge without
   explicit justification.
5. DO NOT allow pose changes to alter bone lengths.
6. DO NOT accept anatomically impossible joint configurations.
7. DO NOT accept a reconstruction solely because it looks correct from the
   evidence camera.
8. A reconstruction that looks correct in the evidence view but is anatomically or
   spatially implausible from side/oblique views is a FAILED reconstruction.
9. DO NOT invent unobserved image evidence.
10. Use unobserved frontal/side views only for 3D plausibility testing.
11. DO NOT use example or measured reference stature values during fitting.
12. DO NOT report false precision.
13. Keep all important geometry changes reproducible.
14. Preserve complete editable Blender results and rejected checkpoints.
15. The forensic expert must be able to understand WHY each reported stature was
    obtained.
16. Every selected image branch must start from the same canonical generic
    morphology, not from a previous image's fitted body.
17. DO NOT transfer fitted morphology, pose, case-specific priors or final height
    between image branches.
18. DO NOT use a cross-image mean, previous image result or reference stature as a
    fitting target.
19. Cross-image comparison occurs only after all image branches are frozen.
20. Cross-image mean/SD are descriptive statistics, not validated uncertainty.
21. Report all active parameter bounds and all failed sensitivity trials.
22. Record every post-launch investigator message and do not overstate autonomy.
23. AI plausibility review is not independent human expert review.
======================================================================
PRIMARY OPTIMIZATION PHILOSOPHY
======================================================================
The task is NOT:
"Create something that looks like the person from the camera."

For EACH image branch, the task IS:
"Find an anatomically plausible, metrically consistent articulated 3D human
morphology and pose that explain that image within the fixed calibrated camera
and reconstructed 3D scene, starting from a generic body and without importing a
fitted body from another image."

After all branches are frozen, the study-level question becomes:
"How consistent are the independently fitted single-image reconstructions of the
same physical person, and which ambiguities or model assumptions explain their
disagreement?"

Conceptually, for each image:
evidence consistency
+ camera geometry
+ ground geometry
+ skeletal morphology
+ kinematics
+ anatomical constraints
+ 3D plausibility
+ transparent uncertainty/sensitivity checks
must all be considered simultaneously.

Cross-image consistency is an OUTCOME TO MEASURE, not a constraint used to force
the reconstructions toward one another.
======================================================================
DEFINITION OF SUCCESS
======================================================================
A successful execution of THIS EXPERIMENTAL PROTOCOL requires ALL of the following:
- camera reproduction/mapping is validated before body fitting,
- scene scale is preserved,
- the camera is locked and reused consistently,
- an articulated measurable skeleton exists,
- body surface remains tied to skeleton,
- each image branch starts from the same generic morphology and fresh optimizer
  state,
- no fitted body parameters are imported between images,
- evidence overlays and residuals are quantified,
- feet interact plausibly with reconstructed ground,
- frontal, side, top and oblique review does not reveal a hard anatomical/spatial
  failure,
- no major joint-limit violations remain,
- no unexplained bilateral segment-length differences remain within a fit,
- active morphology bounds are disclosed,
- multiple initializations and landmark-localization sensitivity checks are run
  and fully reported, including rejected alternatives,
- pose-independent stature can be recovered from every frozen rig,
- intermediate steps and constraint revisions remain auditable,
- all final .blend scenes remain editable,
- interaction/autonomy history is preserved,
- limitations are explicitly documented.

Successful execution does NOT mean the resulting stature is correct, validated,
admissible or suitable for routine casework. Technical accuracy is a separate
evaluation endpoint.
======================================================================
PRIMARY RESULT RULE
======================================================================
For the independent single-image experiment, preserve and report FOUR conditional
image-specific reconstruction results when four images are supplied. Each result
comes from its own fitted morphology and pose.

After all four are frozen, calculate:
- the arithmetic mean of the four stature outputs,
- sample standard deviation using n - 1,
- minimum, maximum and range,
- signed differences between images,
- cross-image variation in fitted body dimensions.

These are DESCRIPTIVE statistics of four reconstructions of one person. They are
not four independent subjects, not a confidence interval and not a validated
forensic measurement uncertainty. Do not imply that averaging removes common
camera, scan, model, clothing, hair, footwear or agent-interpretation errors.

If a predefined laboratory benchmark (for example +/-3 cm) is to be used for a
post-run accuracy assessment, apply it ONLY AFTER reconstruction freeze and
unblinding. State explicitly whether the benchmark applies to each image, the
combined descriptive mean, or both. Never use the benchmark to steer fitting.

The final report must distinguish:
A) successful software/workflow development,
B) successful execution and reporting,
C) technical correctness/accuracy of the reconstruction outputs.
A and B must not be used as evidence that C was achieved.

======================================================================
START NOW
======================================================================
Begin by:
1. inspecting all available case files,
2. reading the supplied forensic workflow/QM documentation,
3. identifying the relevant camera, reference-target and scene data,
4. documenting the current workflow,
5. autonomously reconstructing the camera parameters from the marked image and
   metric 3D reference geometry,
6. validating the reconstructed camera and exact image mapping in Blender.
DO NOT start fitting the person until the autonomous camera reconstruction and
scene geometry have been successfully validated and the camera has been locked.
Then build and validate the forensic rig.
Then perform the complete unknown-person stature reconstruction.
Work through the problem until you either:
A) obtain a geometrically and anatomically plausible candidate stature
suitable for independent expert review,
or
B) determine and document why the supplied evidence is insufficient to
obtain such a result.