Current work order — 2026-10-28: Follow Truth -> LOFO notebook -> Shape integration. Reuse the same eight sources and preserved measurements. Close the outstanding real truth admission/coverage/observability and executable V3.1 training gates; the first new model training uses LOFO, not the older all-eight fit-first bootstrap. Knight is separate regression. The source work requests only the existing sparse appearance admission after successful exact-main deployment; it authorizes no training or model/product promotion.
A100 delivery continuation — 2026-10-26: Continue through the optimized notebook under the explicit user request. Preserve the same FIT8 and pre-Truth baseline. Read the executable-source and delivery work order. Recover/verify the eight exact teacher payloads, close real truth producers and qualification, integrate actual target/head/trainer/Compiler bindings, seal the Platform external worker and benchmark full A100 cost before calling it ready. Source unit courts do not authorize a shortened recipe or synthetic corpus.
Truth Foundry handoff — 2026-10-26: The V1 geometric/mechanical response design and same-eight FIT8 research arms are now defined in Truth Foundry V1 and the execution plan. Primary truth is measured source-space response and compositing; attachment, grip and support remain Compiler consumer interpretations. Independent numeric kernels/local package tools are implemented; real teacher adapters, real truth qualification, real V3.1 model integration and the executable training seal remain pending. Reuse the existing eight source lineages; do not reopen the 139 census. Next: Platform-bound teacher equivalence, intervention/renderer adapters and source truth qualification. The existing read-only Skin inspector is requested again after exact-main deployment; read its receipt for live model/parent state. Production still requires exact Skin parents/model admission and actual Mechanics/Graphics/Motion/Proof consumer closure; IRIS/TESSA is not the only proven blocker. No scientific PASS, model promotion or product activation here.
Active research handoff — 2026-10-26: IRIS/TESSA V3.1 is frozen as a pre-Truth architecture baseline, after actual joint V2, endpoint V2 and DINO fusion report closure. Read the freeze record and current research state. User-selected next research work is an additive extension of the existing FIT8 truth packages, retaining the same eight subjects; Knight is separate regression. Real-data/training state remains Platform-owned. This source record changes no ModelRelease, product activation or qualification. The product-state snapshot below remains scoped historical/current product context.
Date: 2026-10-24
Status: ACTIVE RESPONSIBILITY-CLOSURE PROGRAM
Canonical code line: main only
Primary objective: close Skin → Mechanics → Graphics → Motion → Proof without reopening solved boundaries or creating duplicate authority.
Research objective: make every new hypothesis testable through Platform Experiments/Attempts with exact reuse, lineage and independent qualification.
The current product architecture is governed by responsibilities, not by a numbered-stage mental model:
Input → Shape → Mesh → Rig → Skin → Mechanics → Graphics → Motion → Proof → ProductRevision
Legacy stage IDs remain valid implementation lineage where existing executors/receipts still use them, but new work should be described first by responsibility, typed contract, authority and dependency.
The Compiler participates only where canonical/discrete authority is required. It must not become a duplicate perception model, a hidden repair bucket, or a second owner of work already performed elsewhere.
The next product blocker is fresh current Skin closure. Mechanics follows Skin. Graphics follows the mechanical baseline. The shared-scene Graphics Truth Foundry program is preserved, but it is not allowed to bypass Skin/Mechanics closure.
- One authority per truth. Do not encode the same live state in Git source and Platform state.
- Learned output is evidence/proposal unless explicitly promoted by a court. IRIS/TESSA/AXIS/MIRA do not self-mint canonical product authority.
- Compiler authority is bounded. Compiler validates/binds/resolves/mints where required; it does not duplicate perception or invent missing evidence.
- No subject-specific production tricks. Shared rules/parameters must remain generic; diagnostic subject differences belong in evidence and Experiments.
- Unknown is not false. N/A/UNKNOWN/invalid evidence remains explicit abstention.
- Reuse is exact, not approximate. An artifact is reused only when governing code/model/policy/input lineage satisfies the exact consumer contract.
- Merge, model promotion and product activation are separate transitions. None implies another.
- FIT8 is not unseen proof. If a parameter/hypothesis was selected by observing FIT8, independent LOFO/unseen evidence remains required for generalization claims.
- Source art wins for appearance. Learned completion/support may assist but does not silently replace authored visual truth.
- No new stage chain. New subsystems appear as responsibilities/capabilities and typed contracts.
CURRENT_STATE.md and this file must reflect the active program. Historical plans remain useful evidence but must not steer continue / devam et back into retired work.
Repository observation on 2026-10-24:
main branch protection = OFF
repository rulesets = NONE
Required administrative remediation:
- PR-only main changes;
- required CI/status checks;
- no force-push/delete;
- explicit minimal bypass/emergency policy;
- keep exact-head merge leases in automated merges.
Until GitHub protection/ruleset is enabled, treat this as an open governance/security gap. A source-side test cannot prevent a privileged direct push after the fact.
Live model/channel/product state is Platform-owned. Source capability descriptors contain static requirements and policies only.
MIRA V5.5 exact runtime has been recovered and registered as a Platform DRAFT ModelRelease. Recovery evidence proves runtime identity/parity/compute behavior only.
Do not promote from DRAFT solely because historical Knight/FIT1 evidence exists.
Current execution must establish that the exact registered runtime still participates correctly in the present typed product boundary.
Before executing or importing anything:
- verify exact deployed main and Platform health;
- inspect Platform model-release/channel operational state;
- inspect the explicit authority-host Knight roots;
- identify whether exact current typed prerequisites exist:
CanonicalMeshIR.v1;CanonicalRigIR.v1;- required
RiggingSurfaceIR/carrier evidence; TESSAMaterialSupportFieldIRor exact declared equivalent;
- validate hashes/bindings/authority metadata;
- report READY / MISSING / STALE / INCOMPATIBLE per prerequisite;
- mint no qualification, inference result, promotion or product state.
Do not reinterpret legacy NPZ/JSON files as current canonical typed parents merely because the underlying numbers look compatible.
If exact typed Mesh/Rig parents are absent:
- run/reuse the current Mesh responsibility to obtain
CanonicalMeshIR; - run/reuse the current Rig responsibility to obtain
CanonicalRigIR; - preserve exact evidence and authority bindings;
- do not create a bespoke Knight conversion path.
With exact parents available:
CanonicalMesh
+ CanonicalRig
+ qualified carrier/surface evidence
+ exact MIRA ModelRelease
↓
MIRA runtime inference
↓
W_M proposal/evidence
Execution must obey the recovered compute contract. CPU/cross-device equivalence must not be assumed.
Compiler consumes the exact MIRA proposal and exact typed parents, enforces simplex/binding/coverage/repair budgets and either:
PASS → CanonicalSkinIR.v1
FAIL → typed failure evidence; no authority minted
No hidden fitting, semantic transfer, GSA ownership or silent post-hoc correction is allowed.
A fresh successful Skin execution may support a new current-execution evaluation scope. Promotion to RESEARCH must remain Platform-owned and policy-reviewed.
Do not require FIT8/LOFO/unseen merely to create a research execution channel if the research policy does not claim generic/product qualification. Conversely, do not let a RESEARCH promotion masquerade as genericity or production qualification.
product.skin may enter the admitted capability graph only when its model requirement can resolve an allowed MIRA channel according to the exact execution purpose:
- research:
RESEARCHorCANDIDATEas allowed by policy; - product:
PRODUCTION.
Source must not hard-code a particular current release ID.
Mechanics starts only after exact Mesh/Rig/Skin closure.
Responsibilities:
- post-bind frame/deformation observability;
- dynamic deformation-domain coherence;
- stretch/area/conditioning validity;
- contact/mechanical constraints where product semantics require them;
- motion-probe consumption and qualification;
- typed mechanical failure signatures.
Mechanics must not own:
- draw order;
- visibility/occlusion policy;
- appearance/material truth;
- graphics semantic grouping;
- motion clip selection.
Exit criterion:
the exact canonical Mesh/Rig/Skin puppet is mechanically/deformably admissible under the declared probe policy.
Before new graphics training, close the product contracts that already exist architecturally.
- one product-facing presentation-evidence assembly path;
- explicit pre-M endpoint references and post-M canonical address binding;
- validity/authority/provenance semantics aligned between IRIS evidence and Compiler Graphics;
- pair order/occlusion separated from global ray visibility;
- classification of each evidence family as learned, deterministic query-time or post-M Compiler-derived;
- source-preserving CAA/material addressing retained.
Research hypothesis:
geometry, mechanics, graphics and topology evidence queried from a shared scene representation may provide a more generic basis for presentation relations than a fixed semantic segmentation/part taxonomy.
This is a falsifiable research hypothesis, not a product assumption.
After baseline contract closure, activate the derived-truth program as its own responsibility/capability family:
Authored full-truth assets
↓
Source census / dedup lineage
↓
TeacherSceneIR adapters
↓
exact 8-view + ID + ray-stack renders
↓
bounded interventions
↓
causal source-space + visible response
↓
truth qualification courts
↓
qualified role-free graphics truth
Do not manually paint semantic masks. Do not train heuristic pseudo-labels merely to fill losses.
Experiment 0: Truth Foundry qualification
Experiment 1: IRIS role-free evidence learning
Experiment 2: TESSA presentation A/B
Experiment 3: Compiler Graphics product court
Experiment 4: genericity / holdouts / articulated-machine falsification
Compiler product court should compare student-evidence products with teacher-evidence products, not only model-level AUROC/loss.
Consume valid Mesh/Rig/Skin/Mechanics and motion source. Own:
- clip application/retargeting;
- motion intent;
- declared secondary motion;
- runtime pose generation.
Motion may not repair Skin or decide Graphics order.
Aggregate fail-closed product evidence:
- mechanical/deformation validity;
- contact/attachment residuals;
- graphics coverage/order/visibility/occlusion;
- source-preserving appearance;
- motion/runtime behavior.
Only after this closure may a ProductRevision be promoted.
Once the end-to-end responsibility chain is operational:
Regression witness only. Used to protect known working behavior and exact lineage.
Shared-architecture closure over diverse known subjects with one shared system. No subject-specific code.
If FIT8 is used to choose a hyperparameter/policy, record that selection lineage explicitly.
Family/subject holdout evidence designed to reduce selection leakage.
Independent generalization evidence. This is the appropriate place for stronger genericity claims.
Articulated machines (for example tank/excavator-like structures) are high-value because humanoid/character semantic shortcuts should fail there.
A small parameter experiment must be cheap operationally and rigorous scientifically.
Example: AXIS rig capacity N.
Experiment: AXIS_RIG_CAPACITY_SEARCH
Attempt N=36 → exact reused upstream artifacts → affected execution/proof → evidence
Attempt N=64 → exact reused upstream artifacts → affected execution/proof → evidence
Attempt N=k → ...
Each Attempt is immutable. The Platform records exact code/model/data/policy identity and failure signature.
If no fixed N can satisfy observed subjects, the conclusion may be that the fixed-capacity assumption is falsified. Do not endlessly increase a constant when the actual architecture assumption is failing.
A new adaptive policy becomes a new hypothesis/Experiment and must apply the same generic rule to every subject; subject-name conditionals are forbidden.
The canonical responsibility/capability graph is the only invalidation map. Do not maintain a second hand-written experiment dependency graph.
A changed responsibility invalidates only its affected outputs and transitive downstream consumers. Exact compatible upstream artifacts are reused.
contract/unit/synthetic
↓
Knight regression
↓
1–2 diagnostic subjects
↓
FIT8
↓
LOFO
↓
unseen
Do not pay for the most expensive run when a cheaper court can already falsify the hypothesis.
Keep these separate:
source PR merge
model registration
model evaluation
model channel promotion
capability admission
ProductRevision promotion
scientific genericity claim
Every transition needs its own authority and evidence. A later transition may depend on an earlier one, but must never be inferred from it automatically.
Do not:
- revive a hidden persistent identity/owner system as a prerequisite;
- create a second visual mesh competing with canonical Mesh;
- let IRIS output canonical Mesh;
- let TESSA become canonical topology authority;
- let MIRA mint canonical Skin;
- turn Compiler into a perception network;
- introduce named body-part taxonomies as product truth;
- use subject-specific fixes to make FIT8 pass;
- weaken exact lineage guards to resume old checkpoints/artifacts;
- treat historical evidence as fresh current execution;
- activate Graphics Truth Foundry work before current Skin/Mechanics closure unless a measured blocker explicitly requires it.
1. merge the refreshed current-state/current-plan governance record
2. record and remediate main branch protection/ruleset gap at GitHub admin level
3. implement and run read-only Knight Skin readiness inspection
4. close exact current Mesh/Rig prerequisite lineage
5. execute exact MIRA V5.5 current Skin attempt
6. run Compiler Skin qualification and persist typed proof/failure receipt
7. create current-execution model evaluation and reviewed RESEARCH promotion if justified
8. admit/activate Skin for research only when the model requirement resolves legitimately
9. close Mechanics
10. close Graphics baseline contracts
11. run Graphics Truth Foundry + shared-scene experiments
12. close Motion and Proof/ProductRevision
13. run FIT8 → LOFO → unseen evidence sequence
When ambiguous, prefer the option that preserves one authority, exact lineage, minimal invalidation and fewer hidden branches.
When the user says only continue / devam et, execute the first unresolved item in this order. Do not resume the retired 2026-10-16 plan by default.