FIT8_ADMISSION.md

Fit8 diversity and source admission

Date: 2026-10-13. Selection is source discovery, not a research execution receipt.

FIT8_DIVERSITY_SELECTION_V1.json pins eight exact source files from the sealed October 4 corpus. Knight remains the separate fixed regression witness. All eight retrieved files matched the sealed file ledger. Import, pose, appearance, lineage and render admission remain pending; no optimizer has been started by this selection.

Subject	Coverage	Main stress
Snake	Serpentine	Continuous bends, self overlap, ground contact
Bat	Flying	Wing folds, thin surfaces, wing/body occlusion
T-Rex (Trex.fbx)	Dinosaur	Long tail, short forelimbs, jaw and nonhuman balance
Horse	Quadruped	Four contacts, crossed gait, neck/tail chains
Spider	Arthropod	Many limb chains, crossings and ground clearance
GreenBlob	Amorphous	Few controls, squash/stretch and contact
Shark	Aquatic	Axial swimming, fins and no ground assumption
George (textured)	Mechanical	Articulated rigid parts, local frames and materials

These labels are evaluation metadata. Production modules must not branch on them, source names, joint indices or pack IDs. A cosmetic sibling is not another coverage class. T-Rex is the actual dinosaur pack source, not the humanoid-rigged Ultimate Monsters/Big/Dino.gltf. Snake, Spider, T-Rex and Shark are FBX sources outside the previous GLTF-only truth audit; their raw hashes do not prove import or mechanical truth fidelity.

Admission starts at the first render

The following gates qualify each subject separately, in dependency order. A failure keeps that subject and its diagnostics in the roster; it does not trigger an automatic easier substitute. Replacement needs a versioned selection reason, retained rejected identity and refrozen court before fitting. Model scores must not determine replacement.

Existing boundary	Required evidence before continuing
Stage01/02: source bytes and provenance	Exact source, material/image dependencies and license bytes; source revision; verified ledger membership. Working derivatives stay outside immutable raw sources
Stage03/04: import and full subject	All meshes, armatures, rigid equipment and attachment nodes accounted for; parent/local/world matrices, authored axes, units and visibility recorded. Missing images, lost objects or ambiguous rig/animation bindings block admission
Stage05: pose and camera	Explicit source bind/rest pose and native first-frame pose, distinct when authored that way; one recorded subject-to-canonical transform for geometry, G, W, appearance and motion. Inverse binds and evaluated rest reconstruction agree. No humanoid A-pose assumption
Stage06/07: first eight observations	Source-native and imported source renders compared with the same camera, pose, material policy and frame settings; eight level orthographic views with one center/scale. No per-view pose adjustment, recentering, silhouette repair or attachment repositioning
Stage08: normalization	Whole subject including equipment fits the declared domain; axes, handedness, ground/root origin and inverse transform are explicit. Dynamic validation uses a separately sealed fixed camera envelope, never per-frame recentering
Training admission	Per-subject approved pose/scene/render evidence, usable geometry/rig/skin/appearance/native-motion truth, source dependency closure and leakage component assignment. Geometry-only or partially imported truth is not silently promoted

Source bind/rest, frame0 and animation start are separate addresses. Frame0 may be an authored pose rather than rest identity. Compare against the appropriate source reference; never force identity to hide a rest-frame discrepancy.

Attachment validation must independently check parent and target correspondence, bind/rest transforms, local grip/anchor, offset, rotation, scale and any inherited root/object motion. Record the source relation even when the equipment is a separate rigid mesh or node rather than skinned vertices. Proximity to a hand, zero centroid residual or a plausible picture alone does not establish a correct grip. No selection label is an attachment consumer instruction.

Store per-view source/import/rest/frame0 evidence, visible ownership and contact residuals, clipping/coverage loss and material/alpha differences. Check terminal poses and loop boundaries for each native motion where applicable. Do not invent looping semantics or require ground contact from flying/aquatic actors. Thresholds, reference hashes, sample times and acceptance interpretation must be sealed before measuring a court; this planning document is not such a quantitative seal.

After source admission, compare the first compiled M/G/W + appearance render to that admitted reference at matching cameras and poses, before broader motion testing. Source/render admission cannot close compiled contact, attachment or semantic draw-order failures. Preserve every failed image and attribute it to source, import, pose/scene, observation, model, Compiler or Runtime ownership. An average score must not conceal a failed subject or view.

Fit and regression rules

Use the same architecture and one shared parameter set per model across the eight subjects, with the same Compiler contracts. No per-subject checkpoint, anatomy shortcut, renderer patch or hidden teacher input can support the shared Fit8 claim. Source rig/mesh/weights are teacher/evaluation truth, not image-only model inputs. Native motion can be explicit court input, with its authority recorded. Declared shared architecture changes are allowed, with impact and matched baseline comparisons; they invalidate the real consumers.

Report every subject and module, worst-case failures and timing alongside any aggregate. A fix must rerun affected admitted subjects and the matched Knight regression. Independent qualified artifacts retain exact identities under the Go released-DAG resolver; missing artifacts do not become implicit reuse.

The old September Fit8 prereg is a separate historical selection and must keep its bytes. The new roster does not imply an optimizer schedule, checkpoint seal, new live Attempt or product promotion. Execute scientific work only through the existing Go entry/run/exit contract on deployed main.

Lineage and unseen

Split LOFO by leakage-connected skeleton, topology and material lineage, across pack boundaries. All eight lineage component IDs are still pending. Different coverage labels do not establish lineage independence.

KayKit is excluded from this roster. However, the older Fit8 prereg already names Skeleton Mage and Skeleton Warrior; a creator-level unseen claim requires an exact past-use/derivative audit. Reserve only demonstrably unconsumed identities and their connected lineages, seal the holdout before new fitting, and distinguish creator holdout from a fully unseen creator. Fit8 success is fit evidence; LOFO and unseen are later separate courts.

First implementation work before admission

Source inspection at the recorded main baseline exposes concrete apparatus gaps:

Existing implementation	Observed capability and required work
tools/observation/blender_controlled_8view_v1.py:reset_and_import	FBX-only import; frame0 plus armature REST. Add an explicit GLTF/FBX source contract and isolated source bind/rest evaluation covering animated object/root transforms and NLA/action state before comparing renders
Materializer armature_rows / mesh_rows	Bones, parents, groups and bounds are recorded; complete node local/world matrices and inverse-bind reconstruction evidence are not emitted. Seal these alongside rigid attachment relations
Materializer force_unlit	Transfers base color to emission; source alpha has no explicit preserved alpha connection. Qualify alpha/material policy against the source reference before calling its silhouettes authoritative
adapters/preflight_v2.py:materialize_observation_render_stage	Materializes already supplied exact raster bytes. It is not a source renderer. Bind new admission evidence through the actual Stage03–08 consumers; an external report alone must not mark those missing checks passed

These are source-inspection limits, not observed failures of any selected character. Extend the common apparatus, prove pose/scene/material fidelity on real sources and seal an admission court before optimizer work. Do not change a source pose, conceal missing material, or convert format silently just to obtain a picture. The source identity verifier below cannot close these gaps.

Common source/rest engineering producer

tools/observation/blender_controlled_8view_v2.py imports FBX with animation disabled and GLTF/GLB through an explicit animation-free derivative outside raw source directories. Nodes, default morph state, skins, materials and binary buffers remain unchanged in that derivative. The producer records imported node/bone frames, effective parent/world transforms and a common invertible normalization, then renders a fixed eight-view orbit with explicit clip planes. The existing v1 Knight producer is unchanged.

blender --background --factory-startup --python-exit-code 1 \
  --python tools/observation/blender_controlled_8view_v2.py -- \
  --source /absolute/path/exact-roster-source.gltf \
  --expected-source-sha256 <exact-roster-sha256> \
  --subject-id <roster-subject-key> \
  --out-dir /absolute/path/fresh-candidate-directory

Defaults are 1024 px, shared frame scale 1.12, minimum border margin 8 px and Cycles CPU with seed 0 / 16 samples / no denoising / Standard view transform. NumPy must be available to Blender's own Python, not only the external training interpreter. Hosted Ubuntu CI installs python3-numpy for this boundary. Each receipt binds the source, producer/helper bytes, Blender binary, camera, scene snapshot and pixels. Output must be fresh; a failed/partial directory is retained for diagnosis and cannot carry a stale success receipt into a retry.

Unlit conversion retains active upstream color/alpha and downstream shader graphs. An explicit imported legacy OPAQUE mode keeps its dormant BSDF alpha inactive across the CPU backend. The first common run exposed this distinction on the FBX sources: transporting the inactive alpha value as transparency made three subjects lose views and Spider nearly disappear. A common mode transport fix corrected that engineering failure; no source names or anatomy labels enter the implementation. Native FBX opacity semantics still require independent source/material parity before admission. Unknown/new material conventions must be qualified with their exact importer and renderer; an engineering picture is not authority to infer a source opacity repair.

The October 13 engineering receipt records all eight exact sources producing 64 nonempty 256 px candidate views, minimum margins of 12–13 px and no raw mutation in Blender 4.0.2. Synthetic Blender tests separately exercise nested root/rigid attachment transforms, active linked alpha with a downstream shader and dormant opaque alpha. Run:

python -m unittest tests.repository.test_source_scene_contract_v2 -v
blender --background --factory-startup --python-exit-code 1 \
  --python tests/fixtures/source_scene_v2/blender_smoke.py

This is source engineering, not a new Go Attempt or source/truth admission. Snapshot consistency checks bindings inside the imported scene; effective local matrices are reconstructed from imported world matrices. They do not prove raw node correspondence, authored axes/units, inverse-bind reconstruction or correct grip. Independent rest/frame0/native motion/material references and typed Stage03–08 consumers remain pending. The engineering receipt must not substitute for those gates or mark the first compiled puppet render accepted.

Four-model notebook preparation

The common source fixes, scoped IRIS bind court, shared runner, notebook builder and Go evidence consumers are now described in IRIS_SHARED_COHORT_RUNNER_V1.md. The eight controlled bind observations pass the independent IRIS preparation court. This does not close the full native-motion/material/lineage truth courts above. A real 16-step CPU prefix and exact resume tests validate engineering; physical A100 and live Go qualification retain their separate gates.

FIT8_TRAINING_READINESS_V1.json is the delivery checklist, not a second Attempt/workflow ledger. Its initial status is blocked. Public fit order remains IRIS → TESSA → AXIS → MIRA, with one shared parameter/optimizer state per model over all eight admitted subjects. Existing single-subject runners and solved-tree checkpoints are not a common Fit8 entrypoint. The record names the concrete source gaps for each model.

The first deliverable is the actual IRIS A100 fit notebook, after admission, released Go cohort bindings and frozen fit/evaluation policy. It must embed frozen hash-bound source, run all in a clean runtime, mount Drive, use /content for active work/cache, print short true step/loss/per-subject/time/GPU logs and persist compact results. Checkpoints bind code/inputs/policy and model, optimizer, AMP, RNG and progress; atomic save and interruption/resume must be preflighted. No live GitHub token/checkout is required. A selection preview or CPU source-preparation notebook cannot be presented as this A100 fit notebook. Notebook path, hash, observed download URL and preflight evidence stay null until the exact delivered file and its required evidence exist.

Source identity reproduction

python tools/corpus/verify_fit8_sources_v1.py \
  --audit-bundle /absolute/path/RealSaS_Truth_Corpus_AUDIT_UPLOAD.zip \
  --source-root /absolute/path/materialized-corpus \
  --output /absolute/path/source-identity-receipt.json

The source root preserves the roster's raw/... relative paths. The audit ZIP contains the selected GLTF files; selected FBX bytes can be supplied from the immutable corpus or the recorded official locators. The verifier checks corpus seal, manifest, ledger and all eight exact bytes. It performs no import/render, qualification, Registry write, model fit or product promotion.