Completed Corpus370 audit and tested repairs — 2026-11-03: The sealed7a61 worker finished294 inputs:268 COMPUTED_UNQUALIFIED,26 execution failures;76 earlier preparation failures remain separate (47 mask,29 bind-vertex). All294 receipts/logs/archives and9,747 archived files verified by SHA/size/CRC; archive bytes4,079,694,727. No unresolved integrity error or evidence mutation. Geometry/normals/mechanical family PASS268 each; geometry/normals251 core+17 masked. Causal response23 PASS/245 FAIL (208 tangent,31 face-index,6 trajectory); TESSA14 PASS/254 FAIL (17 invalid faces,237 no passing candidate). Response PASS has ZERO valid tangent rows, and response/TESSA PASS intersection is ZERO. These are family measurements, not IRIS/TESSA eligible counts. Component sign is not global sign. Semantic lineage and train/val/unseen remain unminted. PR324 supplies tested versioned responseV2/independent court, exact TESSA face/incident-corner correspondence, actionable image diagnostics and live stdout/stderr/subject/stage/count/elapsed heartbeats. Archived-array witnesses reproduce the corrections; the trajectory guard still rejects. No training or requalification was run. Source regression passes do not promote old evidence. See canonical/CORPUS370_POSTRUN_AUDIT_20261103.json and docs/corpus/CORPUS370_POSTRUN_REPAIRS_20261103.md. Resume with a new exact Platform child Attempt/release and scoped repair worker: re-evaluate responseV2 and corrected TESSA on all268 sealed Stage47 inputs with explicit original dependencies. Repair independent source-only acquisition and diagnose26 execution/76 preparation failures; never fabricate zero actions, loosen exact-import/trajectory/sign guards or silently drop candidates. Then close the final family matrix, target projections/row masks, actual model loader and lineage split; present readiness coverage before final materialization. Keep8 real source views; the one-reference product frontend remains deferred.

Read-only live failure triage — 2026-11-03 18:37 Istanbul snapshot: 190/294 worker subjects accounted:164 COMPUTED_UNQUALIFIED,26 EXECUTION_FAILED. All26 failed archives and3 representative completed archives verified by size, SHA and ZIP CRC; all164 completed subject logs inspected.23 failures are zero-bone Kenney sources (18 Blocky +5 Cube Pets) unsupported by the pose-bone-only Stage47; one is image-buffer layout; two are admitted-import byte mismatches. Stage08 is PASS and Stage49 remains PENDING for all26: mandatory TeacherScene dependency still prevents independent source geometry on these sources. Do not drop whole subjects or fabricate rig/texture/sign authority to get around that dependency. Lane logs for164 completed subjects: geometry/normals/mechanical164 PASS each; response20 PASS/144 FAIL (119 tangent,19 face validity,6 trajectory); TESSA7 PASS/ 157 FAIL (15 invalid faces,142 no qualified candidate). These are intermediate lane outcomes, not final family/head admission; PASS logging may precede MASKED. Exact archived-array replay reproduced a TESSA identity-mesh face493 nearest relabeling error: vertex trajectories are identical, yet wrong face comparison produces response NRMS22566.394623195174 and a false fold comparison. Another representative has scalar agreement <6e-8 but tangent differences near2.0. A third has8 invalid faces and duplicate trajectory drift2.98e-8 vs1.14e-8 limit. Do not globally loosen thresholds or override SOURCE_FALLBACK acceptance. See canonical/CORPUS370_FAILURE_TRIAGE_20261103.json for exact identities, per-subject causes and pending generic repairs. No runtime/producer/threshold changes or new qualification were made; the running sealed7a61 notebook continues.

User sequencing / single-view deferred — 2026-11-03: Record one front-facing input→normalization→eight jointly predicted camera-conditioned views as the NEXT product research task. The user explicitly defers this implementation; finish current multiview corpus/IRIS/TESSA truth alignment and materialization first. Single-view validation is not a new blocker on that current scope. Larger DINO alone does not supply a multiview generator or guarantee corpus sufficiency. Existing truth/model/split/materialization gates and lack of training permission remain. See the updated single-view decision JSON and design document.

Single-view product / Large foundation decision — 2026-11-03: The user confirms one real RGB reference→full3D prediction. Eight calibrated research inputs do not establish this capability. Small is no longer the primary default and no Small-ceiling experiment is required before authorized Large research. Preferred candidate is DINOv3-L/16; DINOv2-L/14 is the required reference; SigLIP2-L is an alternative, Giant is optional. Empirical winner is UNSELECTED. See canonical/IRIS_SINGLE_VIEW_FOUNDATION_DECISION_V1.json and docs/models/IRIS_SINGLE_VIEW_FOUNDATION_20261103.md. Current S/384/eight-view frontend, cache, input projections and camera/scale assumptions require a new versioned path. No weights were swapped, no GPU comparison or optimizer ran. Keep source truths independent; hidden/generated views remain predicted input or target-only evidence, never extra observed truth. Final370/masks/lineage, materialization, sign/probe + tangent repair and model readiness remain open.

Actual thin input availability — 2026-11-03: Read-only native transport run 41627179582 completed, artifact13226387715. A fresh reconstruction of the pinned runtime was compared against original input bytes:5818 unique objects come from the already-published base ZIP,5338 from authenticated runtime reconstruction, and811 from the exact overlay (18188466raw bytes). Thin ZIP is7586146bytes, SHA edf024254862c22a22db9dcbd1e7333d7fef6c27234e24909ae8967db2f6e0d5. Connected Drive upload created file1ZNIfLyw4qMoISgSwjr4O7F5Co_4eJv1y in the existing7a61 compute folder; downloaded readback matched that exact SHA and size. Original8.27MB capsule seal4140247a, all18286 file rows,294 subject inputs, scientific7a61 code and worker result identity remain unchanged. Do not await the failed monolithic publisher or upload its final readiness receipt. New generator tools/notebooks/build_corpus370_thin_compute_v1.py verifies mounted existing base ZIP and small overlay directly, reports source-copy/input progress and stops on missing/drifted inputs. No unbounded publisher wait. It is a compute worker, NOT corpus qualification/materialization/training.

Confirmed transport failure / thin recovery — 2026-11-03: Publication 41616613311 failed at10:18:55UTC with Drive403 RATE_LIMIT_EXCEEDED for rclone's shared project202264815644. Its full GitHub backup had reached729808896bytes when the10minute step limit killed it; no full backup artifact was finalized. Only the8.27MB seal reached Drive. The old notebook's wait loop did NOT detect terminal publication failure, so its repeated messages were not progress. Stop that wait cell. Do not resubmit the same monolithic transport.

The existing Drive base bundle4462b7f9 covers5818 unique capsule objects. Remaining non-runtime inputs are348 objects/10851385bytes; fresh pinned-runtime reconstruction plus exact original cache overlays can preserve all6134 runtime file rows without uploading the full runtime. Thin transport keeps original seal 4140247a and scientific7a61 code unchanged, checks every destination SHA/size, preserves source/Attempt/split boundaries, and grants no authority. Its native review-branch job is read-only repacking of sealed bytes, not new measurement. Obtain and verify the small transport artifact, upload it through connected Drive, verify readback, then deliver the new notebook with no publisher wait. Main/PR contract success alone is NOT input availability or completed science.

Compute publication recovery — 2026-11-03: Export41526186943 successfully sealed294 exact inputs in about4minutes. Archive2037365166bytes has SHA 4aa855302017162c94d34638d995c03e8019da21fb21bb214d166785d085ace3, seal SHA 4140247ae8fd6ea93c367cb953ed8ee115e1df9709a640250ac9fcfac18d5ee1. Its Drive upload was cancelled at the90minute job boundary; Drive target was still empty on readback. This is transport failure, not a new subject/court FAIL. Original metadata/notebook survived as artifact13185224022. Native7a61 deployment and exact API/Engine/reuse smoke completed successfully in41526187153. PR318 then merged as1e305f16f3dc895b85d98968ff2ce408473be524; its native deployment also passed in41614704651. Scientific input/code remains sealed7a61. Publication recovery reuses that exact existing ZIP, verifies its SHA/size/seal, stages in a new directory, preserves a complete authenticated GitHub artifact, and uses bounded network waits plus visible rclone progress. Drive readiness receipt is uploaded LAST, after a remote checksum check. No scientific execution, new Campaign or qualification is created. See docs/platform/CORPUS370_COMPUTE_PUBLICATION_V1.md. Recheck actual publication receipt and Drive size/hash before calling the notebook's inputs ready.

Resource/offload repair — 2026-11-02: Native runner updated to 2.338.2, connected and accepted readback job125536951888, which completed successfully. Its 20:39 UTC readback exposed Engine oom-kill, not merely runner unavailability. Latest witnessed V4 state: 33 COMPLETED / 1 FAILED / 260 OPEN; IRIS33 / TESSA2 are partial court coverage, NOT final eligibility. The old queue-only diagnosis below is historical. The user explicitly reaffirmed notebook offload when the self-hosted machine cannot carry the workload.

PR317 bounds the action/face working tensors and nearest queries, preserves full-mesh scale/global adjacency/all samples, caches exact source-only TESSA measurements, and batches the independent causal court. Cross-platform golden floats use1e-13 equality tolerance (observed LAPACK1ULP), far below unchanged scientific court thresholds; masks/addresses remain exact. Resource MemoryError now escapes family accounting instead of becoming a scientific subject FAIL.

The exact Go lifecycle descriptor targets ONLY V4 and preserves completed artifacts. Automatic native continuation is DISABLED. Deployment uses one native slot; heavy corpus work moves to a separate, serial high-RAM compute notebook. The read-only compute export request pins exact V4 PREPARATION, all294 member Attempt/SubjectInput identities, unchanged inputs/evaluator bytes, and the76 explicit preparation failures. Export workflow publishes the real capsule and generated Run All notebook to Drive only after canonical-main merge. Worker outputs are raw, persisted per subject, with exact resume hashes and original parent identities. No Registry qualification or training authorization is minted by the notebook. Native import/independent validation, unresolved global sign/graphics, final authority matrix, actual lineage-safe split, target projection and the final materialization notebook remain required. Do not call the compute notebook a completed corpus materialization or claim that imports already exist.

Notebook delivery checkpoint — 2026-11-02 23:20 Europe/Istanbul: Canonical/deployed main remains c16d9a17d73a308322ead20e065fa54d3f2be019. Active exact294 Campaign is 1613ca08-c05f-4df4-96d2-a675d142c70c, EngineRelease b0f6500e-d4af-4c7b-bdc1-1d105fad60aa; do not blindly resubmit or cancel it. PREPARATION at corpus370_truth_preparation/c16d9a17d73a308322ead20e065fa54d3f2be019_isolated/PREPARATION.json has SHA 451189d0d2221119727ea0015f6b43de83befd759f2c92a4339cbd5898c1fba8. Last verified live readback at 20:19 Istanbul (17:19 UTC): 26 COMPLETED / 268 OPEN / no failed Attempts. Current court coverage was IRIS26 / TESSA2, including26 surface+normal,3 causal response,2 mechanical teachers; this is partial evidence, NOT final corpus eligibility or a sealed split. All global sign and graphics cells remain UNKNOWN; lineage/split and the final projection index still require closure. The final production notebook has not been emitted.

A new readback requested at 23:10 Istanbul is still QUEUED without any active workflow in this repository. The native realsas-wsl-1660ti runner is not accepting that job; physical host/runner availability is not observable through the available read-only repository APIs. Do not claim that measurements are currently advancing or that the runner is definitively offline. Agent has no remote WSL terminal. Restore authorized runner availability, let the already queued readback finish, then inspect the actual current Campaign state before choosing any lifecycle change. Preserve every original scope and court FAIL.

Review branch corpus370-numeric-throughput-v1 prepares two numeric cost fixes: action-batched independent response measurement (unchanged thresholds/basis/ reduction order, all8 outputs bit-identical to frozen scalar regression), and one hash-checked process-local source evaluation per TESSA candidate ladder. Every candidate gate still runs; source masks/normals/response/strain are reused only for exact input identity. Local600-face/64-action kernel benchmark: .89117s scalar vs .06429s batched (~13.86x), NOT an end-to-end corpus estimate. 17 focused numeric/court/teacher/scope tests passed. Implementation comparison changes ONLY Stage49; stages01–08/47/48 retain identity. Code has NOT been deployed or used to mint new family authority. Do not merge it over an unknown active scope or schedule a duplicate corpus simply to report a speedup.

Materialization execution checkpoint — 2026-11-02 15:43 UTC: Main is 2f92a42cdad18d43797f98c66c7dd2eeda8f5dc9 (PR314/315 merged). Both exact-main deployments and real API→Engine/reuse smokes passed. Original Campaign 14431b73-3369-46d4-9e93-dd36ed1686cb is terminal: 288 Temporal cancellations, 4 previously failed workflows, 2 completed; original artifacts and failures retained. Child Campaign 614fa91b-9b2a-4b81-8db5-1ae7127107b1 release ea6b4c8e-bffd-4fea-9e51-119df4c815e3 submitted 294 members in successful workflow 41472938393: 32 exact external-dependency closures repaired, 262 SubjectInputs reused unchanged. Its PREPARATION SHA is 84c0a0ecac5897479606ed6d39c2565461df5faf75ef5b60789f10b5be273eba. Initial observer: 2 COMPLETED / 17 FAILED / 275 OPEN at 15:39:30. Failures include INTERVENTION_BASELINE_ARTIFACT_MISSING:04_FULL_SUBJECT_ADMISSION: earlier repair changed execution/reuse checks but failed to restrict the immutable frozen-stage set passed to strict resolver verification. PR316 now captures only actually produced baseline artifact IDs before sealing the spec; strict cold-cache/artifact-ID checks remain unchanged. Its canonical lifecycle descriptor targets ONLY this exact child cohort; the following explicit V4 continuation changes no scientific adapter (direct_changed_stage_ids=[]). Do not resubmit either old Campaign or edit their immutable specs.

Notebook path is now implemented, not yet authorized: PR316 contains per-subject/head checkpoint JSON+Markdown, numeric IRIS/TESSA projections, full exact-byte closure exporter, resumable sealed-bundle materializer with consumer verification, and a four-cell Colab Run All generator (Drive mount, pinned code/archive/seal, elapsed time, visible progress, exact resume). No production notebook was emitted: generator rejects OPEN checkpoints. Focused numeric/materialization/notebook tests: 21 passed. Full source CI and PostgreSQL-backed Platform CI are still required for final PR316 head. The actual report in Drive is 41472668388_CORPUS370_MATERIALIZATION_CHECKPOINT_V1.md (file 1FnG5-r30yGBKSx-NUdr69lhlhMaO43Zt). It is explicitly BLOCKED, not final N: 2 Registry-complete old subject sweeps support IRIS surface+normals; both have TESSA teacher FAIL (TESSA_V5_NO_MECHANICALLY_ADMISSIBLE_FALLBACK) and causal response FAIL (CAUSAL_COURT_TANGENT). Other 368 required family sweeps are unresolved. Global sign/graphics authority and real lineage audit/split remain open. Do not turn UNKNOWN into MASKED, reuse local component sign as global, substitute teacher geometry for student Z/G, or start an optimizer. New V4 results must be inspected before final admission, target projection, split, full bundle seal and actual notebook delivery. The notebook generator expects already projected per-subject manifests; that final corpus-wide projection index still needs binding to completed Registry family evidence.

Worker/lifecycle repair source — 2026-11-02: PR314 now includes supervised process isolation for Python Engine activities, bounded concurrency, exact native-descendant cleanup, and the Go-owned exact-Campaign cancellation path. canonical/CORPUS370_WORKER_ISOLATION_REPAIR_V1.json binds only the original 294 Attempt IDs and exact PREPARATION SHA; deployment may apply that explicit lifecycle repair before the unchanged OPEN-Attempt guard. The original Campaign has NOT yet been cancelled at this checkpoint. A child Campaign must declare its terminal parent and per-member release/input intervention; retain every actual qualified prefix artifact and compute previously unexecuted stages. Read docs/platform/ENGINE_ACTIVITY_ISOLATION_V1.md. Unit/source tests pass; PostgreSQL-backed cancellation/continuation checks must pass CI before merge. libspatialindex explicitly forbids concurrent thread access; no current native threaded Stage49 result is silently promoted into a new apparatus. The user asked for a status report during work; this did not cancel the corpus goal.

First real execution readback — 2026-11-02 14:28 UTC: Original Campaign 14431b73-3369-46d4-9e93-dd36ed1686cb remains active; do not resubmit it. All 294 Attempts are OPEN. The 14:20 observer exported 16 committed Stage47 manifests and saw 14 Stage49 executions running. Three Stage47 ledgers failed with SOURCE_EXTERNAL_RESOURCE_MISSING_OR_ESCAPED: GLB external textures were omitted by the transport packer because it collected dependencies only for .gltf. This is a packaging failure, not a morphology/truth rejection. Repair both packers with the common binary-safe reader, retain raw bytes and URI guards, then bind a repaired bundle to an explicit continuation; do not overwrite V2. The old observer incorrectly counted all OPEN Attempts as failed; its reported failed_attempt_count=294 is not 294 terminal failures. The pending/failure counter repair and short reference-only initial snapshots are in PR314. Readback run 41461627488 reports 12 CPUs, ~7.7 GiB RAM, ~2.4 GiB available and ~0.9 GiB swap used. Investigate stage-family progress/worker logs before changing execution capacity. Bound the geometry nearest-surface workload without changing queries or targets; exact chunk-invariance tests cover the proposed repair. Latest historical Drive geometry-repair receipt is 17 complete / 122 failed (97 nonclosed-manifold, 22 topology mismatch, 3 triangle collapse), not Truth139 readiness. Global-solid binding, other families, lineage review and all final model/materialization readiness remain open. No notebook or optimizer yet.

Actual Corpus370 execution — 2026-11-02: PR313 merged as 36d26f2c39112f46ab82045582b4276bdb8e9108; exact-main deployment run 41455024458 succeeded. Actual workflow 41455683161 sealed all 294 prepared inputs and submitted Campaign 14431b73-3369-46d4-9e93-dd36ed1686cb, Experiment 52c61b66-15f5-4249-9f95-13c780216679, EngineRelease 00aceafd-8b20-40b5-836f-fc5f306b6432. The 76 excluded preparation failures remain bound in PREPARATION.json; 294 submitted is not an eligibility count. First Registry readback at 13:58 UTC observed 294 OPEN Attempts and no committed Stage49 families. Resume/read those exact Attempts; do not resubmit. Read-only workflow corpus370-closure-input-readback-v1.yml projects Registry-committed family outputs into a clearly partial matrix/roster and publishes it to Drive. The same first partial court aggregation has SOURCE_IMPORT/CAMERA 341 PASS / 29 FAIL, FBX RIG 111 PASS and SKIN 108 PASS / 3 MASKED. glTF source rig courts still require exact importer equivalence; numeric global-solid 31/62 PASS is source-world evidence with current camera/canonical binding still open.

Corpus370 execution continuation — 2026-11-02: The user requires actual subject-family closure before materialization/training. The active operator request is canonical/CORPUS370_TRUTH_OPERATOR_REQUEST_V2.json, binding the exact failure-preserving V2 bundle (370 candidates; 294 prepared / 76 explicit failure rows). The finite Platform target is 49_SUBJECT_TRUTH_FAMILIES: exact Stage47 response followed by independent surface geometry, mechanical topology, causal response and TESSA V5 courts. Family failures are retained independently; stage accounting success is not a family PASS or training readiness. Stage48 graphics measurement remains a separate scope and is not repeated by this target. Resume the original Campaign/Attempts and inspect Registry outputs; never blindly submit a replacement after a timeout. Aggregate those outputs with existing source/camera and rig/skin evidence, resolve global-solid and graphics authority, seal lineage/split/materialization, and inspect the final model/target bindings before delivering the materialization notebook. No optimizer is authorized here. AGENTS.md requires repairing missing Platform capabilities while preserving scientific evidence, rather than leaving authorized work blocked or bypassing it.

Pinned numeric-runtime repair — 2026-10-28: Controlled sealed replay reproduced all eight admitted hashes with glibc 2.39 and all eight native-host hashes with glibc 2.43 on the same CPU and Blender. This isolates a runtime dependency; version/executable pinning alone was insufficient. The next repair declares a separate evaluator derivative plus pinned loader/libc/libm, retaining the original executable and every prior result. Native CI must reproduce all eight existing admitted imports before merging this repair; new teacher results still require deployed-main Platform run/Registry readback. No tolerances change. Read the numeric runtime binding.

Actual observation-rest repair — 2026-10-28: #277 merged as ed92b3d0f2668ff5994909b2d5a56d923aac64a4; deployment succeeded. Actual run 40731693360, campaign 6b8975af-d389-4584-ba7c-1387cf048c88: four cohort members and separate Knight regression completed; T-Rex, Horse, Spider and George failed the exact import-byte gate. All nine Attempts are terminal. The official 4.0.2 binary hash matched its pin, but the native host did not reproduce all bytes that the same binary reproduced locally. Read the measured runtime boundary. Do not blindly resubmit, weaken the gate, claim a complete cohort or begin training. Readback revision 5 observes this exact repair. Next inspect complete runtime provenance and establish a common sealed import context for observation and teacher; executable identity alone is insufficient.

Observation-rest continuation — 2026-10-28: Sparse Appearance admission completed through real Platform Attempt/Registry readback (run 40676103301). Continue with observation-rest calibration V2: the same eight subjects plus separate regression, exact admitted import replay, and a finite nine-node Platform research release. Preserve V1 measurements and native clips; V2 does not assume their temporal compatibility or grant training. Read actual new PREPARATION/OBSERVATION before claiming cohort completion.

User-selected continuation — 2026-10-28: Complete the existing Truth Foundry -> IRIS/TESSA V3.1 LOFO notebook -> matching checkpoint/Shape integration. FIT8 is the system-development cohort; the first new model training is 7-train/1-held-out LOFO, not an all-eight fit-first campaign. Read the current delivery continuation before the older snapshots below. First close the hash-pinned sparse appearance Platform admission after exact-main deployment and inspect its real receipt. #244's original run stopped before admission; new source is not a PASS. No training, ModelRelease promotion or product activation is minted here.

Actual calibration failure diagnosed — 2026-10-27: PR225 merged as 0b5c650c070e27b03ffd2c90cd7fc560459e6f6d. Readback run 40477560825, artifact 12720313840 (archive SHA256 f322366772997fd4b54a1c2e589aedadf1c38fb2d1c6b3c1d5e2f9efc244efbf), recovered every bound failure record: all eight failed on KeyError: observation_set_hash before Blender. NormalizationDomainIR actually exposes observation_set_binding_hash. This is an adapter/schema defect, not a morphology, model or teacher-response result. The repair uses that real wire field and declares direct Source01/Source03/Normalization08 parents. The native synthetic court now traverses the actual Compiler wire/adapter/Blender boundary. The declared idle-regression drain preserves all original history and lets the existing Go owner reconcile Temporal terminal proof; it cannot terminate active computation, other stages or unrelated Attempts. The requested repair campaign has the same roster, the old EngineRelease as explicit baseline and requires the old campaign drained before any submission. Read its actual receipt; source tests do not prove real calibration or notebook readiness.

FIT8's goal remains the whole common source-to-Compiler-to-runtime/proof chain on all eight morphologies, with shared learned parameters and common rules. A Knight pass or eight training fits cannot substitute for that closure; unseen generalization remains a separate court.

Read-only diagnosis remains available during a blocked upgrade — 2026-10-27: PR224 merged as 6448f2c48a5a4fe0df80e0c5560284e655ce972c. Its deployment run 40475763050 stopped at OPEN_ATTEMPTS_REQUIRE_OPERATOR_DRAIN_BEFORE_UPGRADE; the older services remain active. The calibration regression Attempt is still waiting for external evidence. Do not disable this guard or cancel unrelated work. The hash-pinned, GET-only diagnostic readback may use the existing API after either successful or failed deployment; it must not be gated by an upgrade that it is needed to diagnose. The original failed calibration is not retried.

Calibration readback — 2026-10-27: PR223 merged as 9bc200b5ec530d269792b01839d0c410ea0460e5; all 13 source checks passed and exact-main developer deployment succeeded. Actual campaign d713b26b-5bbf-4d55-b6b8-9e10b9a81029, Experiment e0aa1792-64e7-48b0-8f95-fd2f44e4cbe5, EngineRelease abf94e64-8b0b-4fb8-bf59-a5e87430d796: the first observer snapshot reports all eight cohort Attempts passed Stage08 then failed Stage47. The separate regression reached the explicit Stage09 external-evidence wait. Readback run 40414940754 / artifact 12696105110 preserves those exact Attempts, not successful causal truth. Detailed bound Engine ledgers/Blender logs are requested by the read-only readback descriptor; diagnose before any declared repair execution. The original failed history must remain intact. Notebook delivery remains open.

Source response calibration — 2026-10-27: All eight historical native teachers now have exact-array and exact-NPZ replay proof (run 40405919318). The new isolated kinematic/UV/material-input producer and same-cohort Platform calibration path are implemented for source validation; real calibration artifacts must still be read back. A required external seal holds the learned boundary; no old/new training is launched by calibration. Read the calibration scope. Renderer truth, qualification, final targets/trainer binding and notebook remain open. Do not relabel native replay or source CI as causal truth admission.

A100 delivery continuation — 2026-10-26: The user explicitly selected continuing through the optimized FIT8 notebook. Candidate observation-to-Dense3D and V3.1 TESSA source, chunk-independent rolling attention/incremental KV, semantic-update packing, exact durable resume and heartbeat primitives are now implemented. They remain source foundations: real truth qualification, target parameterization, full trainer/consumer integration and Platform execution seal are open. Read the delivery work order. A read-only exact-byte teacher recovery/Platform inspection is requested after deployment. Independent Compiler work may continue while future training runs. No notebook readiness, scientific PASS or product activation is claimed.

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.

RealSaS-OPT — Current State

Date: 2026-10-24
Canonical code line: main only
Active program: close the canonical product responsibilities in order: Skin → Mechanics → Graphics → Motion → Proof → ProductRevision, while the Platform remains the operational authority for experiments, releases, artifacts and model selection.
Architecture view: responsibility/capability map. Legacy numbered stages remain implementation details and historical lineage, not the primary mental model.

1. Executive state

RealSaS is no longer operating as a loose collection of notebook/stage experiments. The current architecture is organized around stable product responsibilities and explicit authority boundaries:

Input
  ↓
Shape
  ↓
Mesh
  ↓
Rig
  ↓
Skin
  ↓
Mechanics
  ↓
Graphics
  ↓
Motion
  ↓
Proof
  ↓
ProductRevision
  ↓
Studio / product consumption

Learned systems produce evidence or proposals. The Compiler validates, binds, resolves contradictions and mints canonical product IR where the domain requires discrete authority. The Compiler must not duplicate learned perception, silently repair unrelated responsibilities, or become a second implementation of the same work.

The architectural objective is a simple map where each responsibility has one clear purpose, one authority boundary and explicit typed inputs/outputs. Avoid duplicate authority, hidden data loss, unnecessary branching and repeated work.

2. Operational authority model

The current control plane distinguishes source, execution, scientific intent and product state:

Concern	Authority
reviewed executable source	Git `main`
exact executable snapshot	`EngineRelease`
research question / preregistration	`Experiment`
one execution	`Attempt`
immutable bytes / typed evidence	Artifact Registry / CAS
model checkpoint identity + evaluation envelope	`ModelRelease`
selected model for RESEARCH/CANDIDATE/PRODUCTION	Platform model channels
canonical product state	`ProductRevision`
canonical product/scientific decisions	Compiler + qualification/proof courts

A merge to main is not model promotion. Model registration is not qualification. Qualification is not product activation. Product activation is not a genericity claim.

Operational model/channel state must be read from Platform (/v1/ops/status, model-release/channel APIs), not copied into source descriptors as live truth.

2.1 Preserved V2 continuity anchors

The responsibility map supersedes stage-centric navigation; it does not erase the still-valid V2 implementation continuity records. canonical/V2_IMPLEMENTATION_READINESS.json and the existing 43-stage dependency DAG remain machine-readable implementation/provenance anchors wherever legacy executors and receipts still reference them. They are not a competing architecture authority and must not override current responsibility ownership.

The registered governance gate id PLATFORM_EXECUTION_LANES remains the active control-plane execution-lane continuity gate on main. It records the boundary between modular research execution and product execution; it does not create a second scientific architecture or restore stage-centric ownership.

The historical governance context token CARRIER_NATIVE_FIT1_SCOPED_PASS__PLATFORM_LANES_IMPLEMENTED remains a rehydration/continuity anchor for the proven V2 control plane. It records prior closure context only; it does not promote MIRA, activate Skin, prove genericity, or override the current responsibility work order.

The established product-quality separations also remain explicit: Geometry, Mechanics, and Appearance are distinct concerns. QualifiedMeshIR remains the preserved typed static-qualified mesh authority entering the historical canonical-carrier boundary; the current product Mesh responsibility mints CanonicalMeshIR.v1 under the present contract. SurfaceAddressing remains the stable canonical-M address boundary used to bind downstream evidence without changing Mesh authority. Complete Appearance Authority (CAA) remains the source-preserving Appearance authority downstream of canonical M. These continuity anchors are preserved while the higher-level operational map moves to responsibilities/capabilities.

3. Current responsibility status

Input — ACTIVE

Source/input capability and execution-lane contracts are implemented. Exact source identity and provenance remain required.

Shape — ACTIVE

Shape is an evidence/proposal responsibility. IRIS/TESSA learned state does not own canonical Mesh authority.

Mesh — ACTIVE

CanonicalMeshIR.v1 is Compiler-minted after static qualification. Learned topology is proposal evidence only. Mesh geometry/topology basis cannot be silently replaced downstream.

Rig — ACTIVE

CanonicalRigIR.v1 is Compiler-minted against the exact CanonicalMesh parent. AXIS may propose structure, but AXIS does not mint canonical Rig authority. Post-bind observability/deformation behavior must not be pushed back into Rig merely because Rig precedes Skin/Mechanics.

Skin — BOUNDARY IMPLEMENTED; PRODUCT ACTIVATION OPEN

CanonicalSkinIR.v1 exists as the canonical Skin authority. The intended path is:

CanonicalMesh
+ CanonicalRig
+ qualified carrier/surface evidence
+ exact admitted MIRA ModelRelease
        ↓
MIRA carrier-native W_M proposal
        ↓
Compiler Skin qualification
        ↓
CanonicalSkinIR

MIRA cannot mint canonical Skin authority.

The Skin source descriptor now carries static requirements/policy only. It deliberately does not mirror whether a particular MIRA release is currently registered, promoted or channel-bound. That live state belongs to Platform.

MIRA V5.5 — EXACT RUNTIME RECOVERED; DRAFT MODELRELEASE REGISTERED

The exact recovered runtime package is operationally registered in Platform as a DRAFT ModelRelease.

Pinned runtime identity:

architecture: RealSaS.MIRA.DirectCarrierSimplex.Axis541.v5_5
runtime sha256: b5ad2d56a9006058cba30c0b123d4e111f3eea3b78ed42f3542cbd095983139a
runtime size:   558840979 bytes

Recorded PASS evaluation scopes are limited to recovered-runtime identity/parity/compute evidence:

RUNTIME_PACKAGE_IDENTITY_V1          PASS
HISTORICAL_READOUT_PARITY_V55        PASS
HISTORICAL_COMPUTE_CONTRACT_V1       PASS
FIT8                                 NOT_RUN
LOFO                                 NOT_RUN
UNSEEN_GENERALIZATION                NOT_RUN

No MIRA promotion has been authorized. No model channel was changed by recovery. product.skin is not active merely because the runtime is now registered.

Current next scientific requirement is a fresh current-execution Knight Skin proof under the current typed Mesh/Rig/Skin boundary. Historical Knight W_M bytes remain evidence, not a substitute for this current execution.

Mechanics — NEXT AFTER SKIN

Mechanics owns post-bind deformation/mechanical validity, dynamic mechanical checks, contact/mechanical constraints where appropriate, and motion-probe qualification. It must not absorb Graphics responsibilities.

Target question:

Is this exact Mesh/Rig/Skin puppet mechanically and deformably valid?

Graphics — PLANNED AFTER MECHANICS BASELINE

Graphics owns source-preserving presentation decisions such as drawables/fragments, visibility, overlap/order, occlusion, appearance support and graphics-side attachment/order decisions.

The current research direction is shared-scene, role-free graphics evidence rather than a fixed semantic part taxonomy. The proposed Graphics Truth Foundry / Truth139 program is a future Graphics research program, not a prerequisite for current Skin closure.

Important protected principles:

IRIS produces scene/evidence, not Mesh authority.
TESSA remains structural proposal only.
Compiler remains canonical discrete authority.
no hidden learned owner/identity system is required as a prerequisite; stable typed addressability comes first.
pair relations and global ray visibility are distinct.
UNKNOWN/N/A must remain abstention, never silently become zero.
source art wins for appearance.

Before new Graphics training, close the current product evidence contracts and test the shared-scene hypothesis with explicit A/B and Compiler product courts.

Motion — OPEN

Motion should consume already valid Mesh/Rig/Skin/Mechanics state. It owns clip application/retargeting, motion intent and secondary motion where defined. It must not repair Skin or own Graphics draw order.

Proof — OPEN

Proof is the final fail-closed product court across mechanical, deformation, contact, coverage, visibility/occlusion, presentation and motion behavior. Only qualified closure may mint/promote a ProductRevision.

4. Current Knight position

Knight remains the primary regression witness, not proof of genericity.

Existing historical/sealed Knight evidence is valuable for exact lineage, regression and comparison. It must not be relabeled as a fresh current ProductRevision or as current canonical typed parents when those exact typed artifacts have not been freshly produced/verified.

The immediate Skin task is therefore fail-closed:

inspect current authority-host/Platform state read-only;
prove whether exact current typed Mesh/Rig parents already exist;
reuse only exact compatible sealed artifacts;
otherwise execute the missing canonical responsibilities;
run exact MIRA runtime inference;
run Compiler Skin qualification;
record the resulting evidence/receipts as a current Attempt;
only then consider RESEARCH-channel qualification/promotion policy.

Do not manufacture CanonicalMesh/CanonicalRig from legacy files merely because equivalent-looking geometry/skeleton data exists.

5. Research operation rules

Experiments are immutable scientific comparisons, not mutable notebooks.

Hypothesis / question
        ↓
Experiment
        ↓
Attempt(s)
        ↓
Artifacts / ModelRelease / EngineRelease identities
        ↓
EvidenceBundle / courts
        ↓
qualification decision
        ↓
separate source merge / model promotion / product activation decisions

For parameter searches (for example AXIS rig capacity N), each value is a separate Attempt under one Experiment. Do not overwrite prior attempts. The dependency graph determines exact reuse and downstream invalidation.

FIT8 used to select a parameter becomes development/selection evidence for that parameter; it cannot then be presented as independent unseen proof. LOFO/unseen remain separate genericity evidence.

Change classes must stay distinct:

MODEL change → new ModelRelease candidate;
COMPILER/code change → new EngineRelease;
POLICY change → new policy/EngineRelease identity as applicable;
DATA/split change → new dataset/evaluation identity;
PRODUCT wiring change → separate product activation/revision.

The Compiler decides domain canonicalization/proofs; it does not decide whether a scientific hypothesis is globally “true”.

6. Platform state

Operational platform primitives now include:

immutable EngineRelease/source binding;
capability/responsibility projection per exact release;
Experiments and Attempts;
Artifact Registry/CAS with streaming large-artifact ingestion;
exact ModelRelease identity and evaluation state;
Platform-owned promotion policy;
model channels;
ProductRevision wiring;
self-hosted developer deployment;
real API → Engine smoke and exact-reuse proof;
authority-host operator workflows.

A partial/research EngineRelease now admits only product capabilities whose declared dependencies exist in that exact release graph. Source capability catalog and release-admitted capability set are intentionally distinct.

7. Governance and security gap

As of this 2026-10-24 snapshot, GitHub reports:

main protected: false
repository rulesets: none

This is an unresolved repository-governance risk. Current PR/CI discipline is procedural; GitHub does not yet enforce it against direct writes to main.

Required admin-level remediation:

protect main or install an equivalent repository ruleset;
require PR-based changes;
require the relevant CI/status checks before merge;
prevent force-push/delete of main;
restrict bypass to an explicitly reviewed emergency path;
preserve exact-head merge leases in automation.

Source-side CI is not a substitute for branch protection because CI can detect a bad direct push only after the write occurred.

8. Deferred / future research

Not on the immediate Skin/Mechanics critical path:

Graphics Truth Foundry / Truth139 full builder implementation;
broad shared-scene graphics model retraining;
one-image → canonical-view synthesis;
large corpus (139 / 10k+) training optimization;
broad appearance/semantic ontology redesign;
alternative carrier representations not motivated by a measured blocker.

These are preserved research programs and should enter through Experiment/Attempt lifecycle rather than replacing the current product backbone by default.

9. Immediate work order

1. keep current source/Platform authority boundaries clean
2. close repository governance documentation and track main-protection remediation
3. run read-only Knight Skin readiness inspection
4. establish exact current CanonicalMesh + CanonicalRig + required Skin inputs
5. execute current MIRA → Compiler Skin proof on Knight
6. record fresh evaluation evidence; promote MIRA only under Platform-owned reviewed policy
7. activate Skin only when its required model channel/admission rule is legitimately satisfied
8. close Mechanics responsibility
9. close baseline Graphics contracts
10. activate Graphics Truth Foundry experiments and shared-scene A/B
11. close Motion
12. close final Proof/ProductRevision
13. run FIT8 genericity closure, then LOFO/unseen as independent evidence

10. Required read order for agents

CURRENT_STATE.md
CURRENT_PLAN.md
AGENTS.md
SYSTEM_INDEX.md
platform/OPERATIONAL_SPINE.md
capability/responsibility contracts relevant to the current task
exact canonical/historical receipts only as referenced evidence

Historical numbered-stage documents are lineage/reference material. They must not override current responsibility ownership or recreate a parallel authority graph.

When the user says only continue / devam et, continue the immediate work order above from the first unresolved item. Do not silently return to the retired 2026-10-16 IRIS→TESSA/FIT8 work order, and do not activate future Graphics research before the current Skin/Mechanics prerequisites are closed.

Navigation Menu

Repository navigation

CURRENT_STATE.md

CURRENT_STATE.md

RealSaS-OPT — Current State

1. Executive state

2. Operational authority model

2.1 Preserved V2 continuity anchors

3. Current responsibility status

Input — ACTIVE

Shape — ACTIVE

Mesh — ACTIVE

Rig — ACTIVE

Skin — BOUNDARY IMPLEMENTED; PRODUCT ACTIVATION OPEN

MIRA V5.5 — EXACT RUNTIME RECOVERED; DRAFT MODELRELEASE REGISTERED

Mechanics — NEXT AFTER SKIN

Graphics — PLANNED AFTER MECHANICS BASELINE

Motion — OPEN

Proof — OPEN

4. Current Knight position

5. Research operation rules

6. Platform state

7. Governance and security gap

8. Deferred / future research

9. Immediate work order

10. Required read order for agents

FilesExpand file tree

CURRENT_STATE.md

Latest commit

History

CURRENT_STATE.md

File metadata and controls

RealSaS-OPT — Current State

1. Executive state

2. Operational authority model

2.1 Preserved V2 continuity anchors

3. Current responsibility status

Input — ACTIVE

Shape — ACTIVE

Mesh — ACTIVE

Rig — ACTIVE

Skin — BOUNDARY IMPLEMENTED; PRODUCT ACTIVATION OPEN

MIRA V5.5 — EXACT RUNTIME RECOVERED; DRAFT MODELRELEASE REGISTERED

Mechanics — NEXT AFTER SKIN

Graphics — PLANNED AFTER MECHANICS BASELINE

Motion — OPEN

Proof — OPEN

4. Current Knight position

5. Research operation rules

6. Platform state

7. Governance and security gap

8. Deferred / future research

9. Immediate work order

10. Required read order for agents