Before IRIS foundation/training work, read
canonical/IRIS_SINGLE_VIEW_FOUNDATION_DECISION_V1.json and
docs/models/IRIS_SINGLE_VIEW_FOUNDATION_20261103.md.
The future product candidate is one real front-facing image→eight predicted
views→full3D. The user has DEFERRED this frontend implementation: finish the
current multiview corpus/materialization/model-truth scope first. Single-view
validation does not become an extra gate on current multiview work. Larger DINO
alone does not provide a multiview generator or prove dataset sufficiency.
This supersedes the old Small-only apparatus default and S-ceiling prerequisite,
not historical evidence or exact replay identities. Multi-view target renderings
must not leak into student conditioning. A larger encoder requires a versioned
frontend/projection/cache/checkpoint path; old eight-view/384 validators are not
a single-view product contract. No training authorization or empirical winner is
minted by the decision.
First read: CURRENT_STATE.md, then docs/platform/EXECUTION_LANES.md.
Use canonical/CANONICAL_HANDOFF_20261007.md for the sealed mechanics handoff.
If the user says only “continue / devam et”, resume from canonical main and the priority order in CURRENT_STATE.md. Do not resurrect an older runtime-audit, VF-11, recovery, research-branch or promotion-branch continuation merely because historical files still exist.
This repository must be resumable without conversational memory.
Resource capacity is an execution choice, not a scientific admission criterion. An online self-hosted runner does not prove that it can carry a corpus workload. After observed OOM or unsuitable resource capacity, move heavy research to the notebook/Colab compute-worker lane with exact input/release/Attempt bindings, bounded memory, serial subject checkpoints and resumable persisted outputs. Keep self-hosted coordination/lightweight validation separate. Do not repeatedly restart or resubmit the same heavy workload on the inadequate host. A notebook returns unqualified evidence; Platform import and independent courts still own admission. Resource exhaustion must not be counted as a subject truth failure.
The Platform and its orchestration are evolving engineering components, not an immutable constraint on authorized research. If a real research task exposes a missing capability, broken adapter, unsuitable lifecycle or incomplete execution path, diagnose and implement the owning generic Platform/orchestration repair, validate it, deploy the exact reviewed revision and resume the original work. Do not leave authorized work blocked merely because the current Platform cannot express or execute it.
A repair must preserve the scientific contracts: immutable evidence and lineage, explicit failure accounting, independent courts, exact dependency identities and the research/product separation. Disabling a guard, weakening a truth criterion, fabricating PASS, silently dropping subjects or introducing an untracked direct execution path is not a Platform repair. Platform limitations must be resolved through actual capability development, with failing evidence and subsequent execution receipts retained.
Report tests, submitted execution, completed execution and verified artifacts as separate states. A green contract workflow is not corpus closure. Continue an authorized task through its concrete deliverables; a status-only final response must not imply that agent work will resume automatically after the turn ends.
Before research execution, read realsasctl agent-context --id <subject UUID>
and enter with agent-enter: exact deployed main SHA, Attempt, released DAG,
immutable SubjectInput and explicit target. A live session must be explicitly
resumed; never resubmit a timed-out run blindly. research-run requires that
session ID. Exit with agent-exit, its exact scope hash, and the next action;
the Go owner seals current state/artifact bindings and returns a handoff hash.
The next entry must acknowledge that hash and stay on the same Attempt or an
explicit child. Read docs/platform/AGENT_HANDOFF.md for the command contract.
Read docs/platform/EXECUTION_LANES.md before stateful development. Use Go
realsasctl research-start and research-run for component replacement;
explicitly select the target and inspect the resolved reuse/execute scope.
Use product-compile for product construction and product-render only for an
exact qualified ProductRevision. Research results, including scoped FIT PASS,
must not silently mint product authority. Preserve independent artifacts and
report code/artifact identities with timings. Do not drive a new result by
searching old branch names or editing a historical execution ledger.
For Fit8 or later multi-subject genericization campaigns, do not create a
notebook-owned authority or eight ad-hoc agent continuations. Read
canonical/FIT8_PLATFORM_CAMPAIGN_V1.json and
docs/platform/FIT8_PLATFORM_CAMPAIGN.md; use realsasctl campaign-start,
campaign-run and campaign --id. Every Fit8 subject enters the released DAG
at Stage01 and targets the same final product-closure stage. Knight is a separate
regression member. Colab/A100 may execute preregistered learned work, but it is a
compute worker only; imported receipts/checkpoints remain non-authoritative until
the subject-bound Compiler stages validate them. A campaign failure is a signal
to fix the owning generic module, never permission for a subject-specific branch.
Finite source-apparatus measurement may use the same Platform campaign API with
an explicit released component target, preserving all source-entry roots, the
entire cohort and its separate regression. Its goal_complete closes only that
measurement target; full_closure_complete remains false. This bounded research
scope must not enter an unrelated downstream training wait or claim FIT8/product
closure. The system-development FIT8 campaign still targets product closure.
Targets include model inference and Compiler modules, not only render. Treat DAG nodes/edges as versioned release data: add, remove or rewire in a new EngineRelease, inspect graph-aware impact, and preserve unrelated artifacts. Keep implementation and plan changes on canonical main through short reviewed PRs; research continuation belongs to the Attempt, never a long-lived branch.
M/G/W preservation is a boundary of the current Knight witness, not a global architecture rule. Models, heads, datasets and evaluations may change under a declared intervention. Only their actual DAG consumers invalidate; independent artifacts retain identity. A shared trunk change legitimately affects its heads. Never hide a dependency hash to make a requested scope appear smaller.
Use standard GitHub-hosted ubuntu-latest for host-independent CPU CI while
the repository is public; private-repository jobs skip. Notebook/Colab is the
default research/GPU lane. Local WSL/GTX remains an explicit native-host or
selected inference option. Old frozen recovery is manual provenance only.
Absolute G3 remains an open gap without cancelling teacher-relative FIT1 PASS.
Read current authority before historical provenance:
canonical/EXPERIMENT_REGISTRY_V3.jsoncanonical/SCIENTIFIC_JOURNAL_V2_20260909.jsonlcanonical/ACTIVE_RUN_V2.jsoncanonical/MAINLINE_EXECUTION_PLAN_V2.jsoncanonical/REALSAS_CANONICAL_ARCHITECTURE_V2_20260920.jsoncanonical/COMPLETE_APPEARANCE_AUTHORITY_V1_20260920.json
Historical provenance only:
canonical/EXPERIMENT_REGISTRY_V2.jsoncanonical/SCIENTIFIC_JOURNAL_V1.jsonl
main is the single canonical continuation line.
Long-lived research/promotion/integration branches must not be used as scientific, product or execution authority. Temporary review branches are allowed only as short-lived source-code collaboration; they must merge promptly and must never become a second continuation state.
Git is source-code history, not product/research state.
The normative state model is:
Git/main = code truth
Artifact Registry = immutable typed artifacts + dependencies
Attempt = research truth
ProductRevision = production truth
Workflow Engine = durable execution truth
Proof/Qualification = promotion eligibility
Research experiments belong in Attempt, not branch chronology. Product promotion belongs in ProductRevision, not Git merge semantics.
Research execution should reuse exact qualified upstream artifacts according to dependency identity. Product compile must execute the required model/Compiler construction path for a new subject unless an exact semantic cache hit is proven. Product render/playback consumes a qualified ProductRevision and must not rerun model inference merely to render it.
The Go control plane is the intended sole owner of durable product/research state, Artifact Registry metadata, Attempt/ProductRevision lifecycle, promotion, dependency scheduling and workflow state.
Current known enforcement gap: legacy GitHub Actions and direct $REALSAS_AUTHORITY_ROOT scripts can still bypass Platform state. Until that gap is closed, such direct paths are allowed only for CI/reproducibility of already sealed evidence. They must not be treated as the normal way to mint new research/product authority.
The active platform program is to make this rule mechanically unavoidable.
The latest developed FIT1 mechanics line is carrier-native: Stage19 owns frozen mechanical carrier M; AXIS V5.4.1 owns the hard causal required tree plus deterministic 256-bin XYZ; MIRA V5.5 predicts W_M directly on that exact M; Compiler validates/materializes the required tree and exact (M,G,W_M) bindings without semantic skin transfer.
Coincident controls are handled subject-free after binding: a mechanically unobservable subtree may inherit its parent frame; any coincident subtree with non-zero final W_M influence fails closed and requires explicit AXIS orientation/tail evidence.
Knight FIT1 mechanics are closed relative to the sealed source/teacher motion envelope. The absolute frozen G3 threshold remains inconclusive because the sealed teacher itself is outside that absolute threshold. This is not FIT8/FITK/unseen/product/generalization evidence.
Canonical public model naming is IRIS -> TESSA -> AXIS -> MIRA; legacy Geppetto/Arachne/ATLAS names may remain only where compatibility requires them.
The user approved the single-M source-wins + existing CAA cutover. Read
CURRENT_STATE.md and tests/fixtures/single_m_caa_knight_v1/baseline.json.
Use tools/demo/render_knight_m_caa_v1.py --output <directory> for exact sealed
RUN regression, never the old retarget trial. Geometry, topology and recorded
motion all use M. There is no independent visual mesh/rebind continuation.
The previous main is retained only at archive/visual-v-before-m-caa-20261011.
Priority after source cutover: grip/weapon relation, then semantic overlap/draw order. Preserve the approved CPU pixels and recorded M poses while investigating. Native exact parity at diagonal depth ties and linear-light composition remain open; do not mint scientific/product PASS from engineering regression. IRIS shared-head research remains a proposal awaiting user approval.
For source work, generate the current map and inspect edited files with
tools/realsas_architecture.py map and impact; see
docs/platform/DYNAMIC_ARCHITECTURE.md. This is discovery, not reuse authority.
An unknown impact must be investigated; it never means zero consumers. At exit,
record exact source/Attempt/artifacts and the pending decision in the handoff.
Geometry, Mechanics and Appearance are co-equal product authorities; Presentation is first-class editable addressing authority. A visually incorrect puppet is not accepted merely because mechanics are valid.
Runtime consumes qualified product artifacts. It may not create a second topology, silently re-solve skin, perform donor search/generative correction, or mutate scientific authority during playback/render.
canonical/TRUTH_CORPUS_V1_STATUS_20261004.md remains corpus authority. Current status is SEALED_RAW__STRUCTURALLY_AUDITED__ADMISSION_PENDING. Do not overclaim it as training-ready canonical truth.
The user authorized diverse Fit8 as the next program. Read
canonical/FIT8_DIVERSITY_SELECTION_V1.json, canonical/FIT8_PLATFORM_CAMPAIGN_V1.json,
docs/corpus/FIT8_ADMISSION.md, docs/platform/FIT8_PLATFORM_CAMPAIGN.md and
docs/platform/PRODUCTIZATION.md. Eight source identities are verified; every
subject must independently close source/import/rest pose, camera, observation and
normalization before learned-stage evidence is consumed. Knight stays the separate
regression witness. Shared model parameters and common contracts are required; no
per-character production fix or silent substitution. New Fit8 scientific state is
owned by the platform campaign/Attempts, never by notebook completion alone.
A FIT/witness PASS is scoped evidence for the exact subject/apparatus. Engineering CI is not unseen/generalization evidence. Knight FIT1 teacher-relative mechanics closure is not unseen/generalization or product authority.
CURRENT_STATE.mdcanonical/CANONICAL_HANDOFF_20261007.mdSYSTEM_INDEX.mdcanonical/V2_IMPLEMENTATION_READINESS.jsoncanonical/KNIGHT_AXIS541_MIRA55_CARRIER_NATIVE_PROMOTION_20261007.jsoncanonical/MAINLINE_EXECUTION_PLAN_V2.jsondocs/platform/adr/0002-compiler-platform-boundary.mddocs/platform/adr/0003-go-control-plane.mdcanonical/TRUTH_CORPUS_V1_STATUS_20261004.md- older recovery/scientific material only as needed for provenance