JEPA-Anything epistemic report

Type: agentic-system-analyses/types/agentic-system-epistemic-report.md

Source-and-claim boundary

The question is what reliance JEPA-Anything licenses when it converts a world-model request into a design and skeleton, and whether its checks support claims about a learned world model. Scope, exclusions and frozen evidence identity are see boundary; the sole registered source is see Source register. All direct inspection below uses SRC-1 at the frozen commit, supplying implementation or doctrine/design evidence, never observed operation.

The model-assisted compiler is a declared workflow. Its deterministic supports implement narrower checks. The relevant claims are EPI-CLM-world-model-compiler and RT-CLM-validation-scope. These are not claims that a passing configuration proves a successful world model. The source explicitly distinguishes the two.

Entry point or operation	Source path in SRC-1	Coverage and supporting records	Limit and prevented conclusion
Implicit/explicit host compilation and repair	jepa-anything-skill/SKILL.md; jepa-anything-skill/agents/openai.yaml	Direct complete reads; see RT-RTE-task-compilation, EPI-CLM-world-model-compiler	Reference-specific interpretation supplied by runtime; no host trace, so no exercised extraction, repair or completion finding.
Standalone validator CLI/API	jepa-anything-skill/scripts/validate_design.py	Direct dispatch/report/load and claim/output branch reads; see EPI-RTE-design-diagnostics	Remaining schema, leakage, state and audit branches covered provisionally by runtime, not exhaustively audited here; no complete validator-correctness finding.
Generator CLI and direct renderer	jepa-anything-skill/scripts/generate_scaffold.py	Direct complete executable body; see RT-RTE-scaffold-admission, EPI-RTE-template-specialization	Direct renderer bypasses the CLI validity gate. No package-generation trace supplied.
Planned-claim and evidence-envelope helpers	jepa-anything-skill/assets/scaffold/src/jepa_task/evaluation.py.tmpl	Direct complete read; see EPI-RTE-evidence-envelope	Metadata comparison, not empirical verification or statistical acceptance.
Activity, capacity, geometry helpers	jepa-anything-skill/assets/scaffold/src/jepa_task/activity.py.tmpl; jepa-anything-skill/assets/scaffold/src/jepa_task/capacity.py.tmpl; jepa-anything-skill/assets/scaffold/src/jepa_task/model.py.tmpl	Direct complete reads; see EPI-RTE-measurement-checks	Callers supply purported measurements; provenance and actual backend computations are not enforced.
Generated tests and consumers	jepa-anything-skill/assets/scaffold/tests/test_contract.py.tmpl; jepa-anything-skill/assets/scaffold/src/jepa_task/usage.py.tmpl	Direct test prefix and complete dispatcher; see EPI-RTE-evidence-envelope, RT-RTE-state-consumption	Remaining tests and adapter checks supplied by runtime. Callback content and its truth conditions are not determinable from this dispatcher.
Core projection, QR hook, EMA	jepa-anything-core/src/jepa_anything_core/opf.py; jepa-anything-core/src/jepa_anything_core/baselines.py	Supplied runtime findings: RT-RTE-core-projection, RT-CMP-orthogonal-factor-projection	No training/optimizer trace or parameter values; mutability does not establish improved capacity.
Core numerical reports	jepa-anything-core/src/jepa_anything_core/audit.py	Direct basis, reconstruction, round-trip and aggregate audit reads; see EPI-RTE-numerical-audits	Factor-audit internals uninspected here; aggregate wiring is not exhaustive correctness evidence.
Core losses and historical variance	jepa-anything-core/src/jepa_anything_core/losses.py	Direct tracker, streaming wrapper, prediction and composite objective reads; see EPI-RTE-variance-history	Primitive activity/variance loss internals not exhaustively inspected; no actual optimizer consumer supplied.
Synthetic audit	recipes/synthetic-linear-dynamics/run_recipe.py	Direct activity generator and report/CLI; supplied structural account: RT-RTE-structural-checks	Recipe audit branches remain sampled; oracle-coordinate activity cannot establish learned activity.
Checkpoint verifier	checkpoints/verify_manifest.py	Direct verification and CLI; see RT-RTE-structural-checks	File-check helper internals supplied by runtime; integrity is not performance or scientific acceptance.
Repository check dispatch	Makefile; .github/workflows/ci.yml	Direct Make read, supplied boundary CI wiring; see RT-RTE-structural-checks	Nonzero status constrains check completion, not independently inspected merge/publication policy.

No target code or tests were executed. Missing host traces prevent conclusions about actual model compliance; missing measured results, weights and datasets prevent conclusions about learned task success; missing interventions prevent causal attribution. Uninspected branches above prevent a system-complete negative about all possible acceptance mechanisms. The supported negative is narrower: the inspected claim-envelope helper does not decide statistical support, and the inspected dispatcher does not require those helpers before calling callbacks. The opaque host model's internal knowledge acquisition and any task-specific learned prediction semantics remain not determinable; see RT-CMP-host-model.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
RT-OBJ-design-protocol	Mathematical/design guidance and constraints	see record	SRC-1: jepa-anything-skill/SKILL.md	Delivery to host does not establish consumption or the truth of every recommendation.
RT-OBJ-design-config	Task identity, causal availability declarations, proposed architecture; diagnostic predicates are a distinct operative part	see record	SRC-1: jepa-anything-skill/scripts/validate_design.py	World-facing declarations remain unverified; defaults are hypotheses.
EPI-OBJ-claims	Planned comparison contracts, not supported comparative outcomes	see record	SRC-1: jepa-anything-skill/scripts/validate_design.py	Metric/direction/scope specify a future check, not a measured result.
EPI-OBJ-evidence-envelope	Caller assertions of measurement identity and passed audits	see record	SRC-1: jepa-anything-skill/assets/scaffold/src/jepa_task/evaluation.py.tmpl	No measured score, statistical criterion or audit attestation is required by this helper.
EPI-OBJ-audit-measurements	Caller-supplied geometry, coordinate activity and capacity assertions, separately checked	see record	SRC-1: jepa-anything-skill/assets/scaffold/src/jepa_task/model.py.tmpl	Authentic measurement origin remains unknown.
RT-OBJ-scaffold-manifest	Specialized executable obligations, saved configuration and diagnostics	see record	SRC-1: jepa-anything-skill/scripts/generate_scaffold.py	Code emission/retention is not empirical endorsement.
EPI-OBJ-audit-reports	Computed numerical properties of supplied tensors	see record	SRC-1: jepa-anything-core/src/jepa_anything_core/audit.py	Numerical scope only; no world-model semantics follow.
EPI-OBJ-variance-tracker	Sample count, coordinate means and squared-deviation sums	see record	SRC-1: jepa-anything-core/src/jepa_anything_core/losses.py	Sufficient statistics describe supplied samples, not generalization.
RT-CMP-orthogonal-factor-projection	None individuated as a world-facing proposition from available basis state	see record	SRC-1: jepa-anything-core/src/jepa_anything_core/opf.py	Basis mutability and geometry are inspectable; actual learned content is not.

Authority-route ledger

Each row separates functions even where it cites the same canonical route. Identity, endpoints, ownership and persistence fields owned by that route are see the cited record. Architectural status below is this member's assessment. No row claims observed activation.

Route ID	Route function	Architectural status	Object/candidate ID	Content/update relation	Transition or check target	Evaluator/condition and domain	Activation and timing	Possible or observed result	Implemented force	Epistemic authority and scope	Operational authority: behavior permitted, blocked, or changed	Behavioral-authority path: consumer, channel, force, horizon	Evidence source ID and local anchor	Claim IDs or none	Mismatch marker or none	Gap/limit
RT-RTE-task-compilation	content transformation	doctrine only	RT-OBJ-design-config	truth-apt transformation: acquisition/import	User/project causal facts enter proposed contract	Host interpretation; open task domain	Requested before validation	Proposed normalized facts	Instruction only	Source warrant unknown; original project facts not supplied	Shapes proposal and questions	see RT-BAP-skill-guidance	SRC-1: jepa-anything-skill/SKILL.md	EPI-CLM-world-model-compiler	none	No actual import trace or semantic fidelity check.
RT-RTE-task-compilation	content transformation	doctrine only	RT-OBJ-design-config	truth-apt transformation: indeterminate	Task facts plus defaults become design hypotheses	Host model guided by references	Requested during proposal/repair	Design choices and compiler assumptions	Instruction only	Candidate status, not accepted efficacy	Host selects design and repair	see RT-BAP-skill-guidance	SRC-1: jepa-anything-skill/SKILL.md	EPI-CLM-world-model-compiler	none	No concrete task/design pair to distinguish entailed extraction, reshaping and non-entailed conjecture; defaults alone are not outcome propositions.
EPI-RTE-design-diagnostics	check/evidence production	implemented	RT-OBJ-design-config, EPI-OBJ-claims	truth-apt transformation: entailed derivation	Declared mapping satisfies coded predicates	Deterministic rule checks; schema and reciprocal evidence-plan domain	Explicit API/CLI call before rendering	Checks, counts, valid bit	Diagnostic; standalone exit status	Bounded facts about declarations under this implementation	Feedback permits host repair; does not repair itself	CLI caller/host; JSON diagnostics, advisory; current repair round	SRC-1: jepa-anything-skill/scripts/validate_design.py	RT-CLM-validation-scope	none	Rule completeness and factual truth not established.
RT-RTE-scaffold-admission	disposition/acceptance	implemented	RT-OBJ-design-config	no content change	Admission for fixed-template generation	CLI consumes report.valid; destination/format checks	Before CLI emission	Invalid rejected; valid proceeds	Enforcing CLI gate	Acceptance of declared completeness for skeleton generation only	Allows filesystem package creation	see RT-RTE-scaffold-admission	SRC-1: jepa-anything-skill/scripts/generate_scaffold.py	RT-CLM-validation-scope	CLI-only	Direct renderer does not require valid report.
EPI-RTE-template-specialization	content transformation	implemented	RT-OBJ-scaffold-manifest	truth-apt transformation: non-ampliative reshaping	Config values become literals and saved config/report	Fixed mapping/repr/substitution	Render call	Specialized interface declarations	Fixed code generation	Preserves selected declarations, not their empirical warrant	Sets later branch constants and checks	see RT-BAP-generated-contract	SRC-1: jepa-anything-skill/scripts/generate_scaffold.py	EPI-CLM-world-model-compiler	none	Does not certify semantic adequacy of original design.
RT-RTE-scaffold-admission	retention	implemented	RT-OBJ-scaffold-manifest	no content change	Package/report/config promoted from temporary tree	Missing/empty destination check	After rendering	Persisted package or exception cleanup	Filesystem replacement; no version rollback	Retention carries declarations and diagnostic result, not accepted scientific claims	Later import/use made available	see RT-BAP-generated-contract	SRC-1: jepa-anything-skill/scripts/generate_scaffold.py	none	none	Not lifecycle integration of an accepted model-quality claim.
EPI-RTE-evidence-envelope	check/evidence production	implemented	EPI-OBJ-claims	no content change	Planned claims reciprocally covered by experiments	Equality/membership; exact plan metadata	Explicit helper call	None or exception	Local enforcing exception	Plan coherence only	Blocks caller continuation if uncaught	Explicit caller/test; synchronous call; enforcing locally; current call	SRC-1: jepa-anything-skill/assets/scaffold/src/jepa_task/evaluation.py.tmpl	RT-CLM-validation-scope	none	No changed claim status.
EPI-RTE-evidence-envelope	check/evidence production	implemented	EPI-OBJ-evidence-envelope	no content change	Result identity matches a planned comparison	Equality/set comparisons; metadata domain	Explicit caller/test invocation	None or exception	Local exception only	Metadata alignment, not statistical support or authentic audits	Allows caller to continue when no mismatch	Explicit caller/test; synchronous call; enforcing locally; current call	SRC-1: jepa-anything-skill/assets/scaffold/src/jepa_task/evaluation.py.tmpl	RT-CLM-validation-scope	none	No mandatory coupling to state consumer or measurement helpers.
EPI-RTE-measurement-checks	check/evidence production	implemented	EPI-OBJ-audit-measurements	no content change	Shapes, full rank, condition ceiling, synthesis NMSE	Numeric/type/range predicates; supplied geometry metadata	Explicit audit call	None or exception	Local exception	Supplied envelope within declared numerical bounds	Caller continuation locally blocked on mismatch	Explicit backend/test caller; Python call; enforcing locally; current call	SRC-1: jepa-anything-skill/assets/scaffold/src/jepa_task/model.py.tmpl	none	none	Does not compute rank or tie evidence to a tensor.
EPI-RTE-measurement-checks	check/evidence production	implemented	EPI-OBJ-audit-measurements	no content change	Every supplied factor/encoder coordinate standard deviation meets floor	Finite nonnegative arrays and thresholds; coordinate activity domain	Explicit audit call	None or exception	Local exception	Threshold compliance of supplied values, not semantics	Caller blocked on low activity	Explicit caller; Python call; enforcing locally; current call	SRC-1: jepa-anything-skill/assets/scaffold/src/jepa_task/activity.py.tmpl	none	none	Measurement origin and sample population unverified.
EPI-RTE-measurement-checks	check/evidence production	implemented	EPI-OBJ-audit-measurements	no content change	Full-model parameters, predictor FLOPs, steps, scopes and baseline tolerances	Integer/scope/relative-gap predicates; declared capacity comparison	Explicit audit call	None or exception	Local exception	Supplied capacity alignment, not independent counters or performance	Caller blocked on capacity mismatch	Explicit caller; Python call; enforcing locally; current call	SRC-1: jepa-anything-skill/assets/scaffold/src/jepa_task/capacity.py.tmpl	none	none	Copying estimates can satisfy predicates without genuine measurement.
EPI-RTE-numerical-audits	check/evidence production	implemented	EPI-OBJ-audit-reports	truth-apt transformation: entailed derivation	Supplied basis rank, condition number, orthogonality and completeness	Numerical linear algebra, finite values, dimensions and tolerances; supplied basis domain	Explicit basis audit call	Computed geometry report with pass/violations	Reporting, not installation gate	Numerical properties of basis, bounded by precision/tolerances	No automatic admission or parameter update	Calling program; report object; advisory; current result unless caller retains it	SRC-1: jepa-anything-core/src/jepa_anything_core/audit.py	none	none	Full-rank or orthogonal geometry cannot establish prediction quality or semantics.
EPI-RTE-numerical-audits	check/evidence production	implemented	EPI-OBJ-audit-reports	truth-apt transformation: entailed derivation	Pseudoinverse round-trip reconstruction	Compare reconstructed supplied state with original; max absolute error threshold	Explicit round-trip audit call	Error summaries and threshold verdict	Reporting, not installation gate	Reconstruction of supplied samples within numerical tolerance	No automatic admission or parameter update	Calling program; report object; advisory; current result unless caller retains it	SRC-1: jepa-anything-core/src/jepa_anything_core/audit.py	none	none	Reconstruction is not held-out prediction.
EPI-RTE-numerical-audits	check/evidence production	implemented	EPI-OBJ-audit-reports	truth-apt transformation: entailed derivation	Direct transpose-synthesis reconstruction	Compare reconstructed supplied state with original; normalized MSE threshold	Explicit transpose-synthesis audit call	Error summaries and threshold verdict	Reporting, not installation gate	Transpose reconstruction of supplied samples within numerical tolerance	No automatic admission or parameter update	Calling program; report object; advisory; current result unless caller retains it	SRC-1: jepa-anything-core/src/jepa_anything_core/audit.py	none	none	Does not license replacing pseudoinverse synthesis outside checked geometry.
EPI-RTE-variance-history	retention	implemented	EPI-OBJ-variance-tracker	truth-apt transformation: entailed derivation	Samples update sufficient statistics	Welford merge, shape/finite/overflow checks; coordinate samples	Explicit update or enabled streaming wrapper	Mutated count/mean/m2, or rejection	In-process state update	Numerical summary of admitted samples within floating precision	Changes subsequent loss inputs	Later streaming call; buffers; computational input; until reset/load/replacement	SRC-1: jepa-anything-core/src/jepa_anything_core/losses.py	none	none	No optimizer or saved-state trace.
EPI-RTE-variance-history	operational admission/selection/consumption	implemented	EPI-OBJ-variance-tracker	non-truth-apt policy/content update: history-dependent differentiable activity penalty	Current batch evaluated with historical sufficient statistics	Combined population variance, std floor, reduction	Streaming loss call; wrapper training flag controls default update	Scalar/tensor loss	Numeric input, not an optimizer decision	No world-facing claim license	Can shape gradients if an external optimizer uses the loss	Calling optimizer code, if supplied; loss tensor; advisory input; training step	SRC-1: jepa-anything-core/src/jepa_anything_core/losses.py	none	none	Potential adaptation, not observed learning or compelled optimization.
RT-RTE-core-projection	behavior/policy adaptation	implemented	RT-CMP-orthogonal-factor-projection	non-truth-apt policy/content update: externally driven basis/EMA parameter changes	Learnable geometry and target weights	Mode, step and explicit caller hooks; numerical parameter domain	External optimization/retraction/EMA calls	Changed parameters	Local numerical mutation	None for predictive truth	Later forward calls use changed values	see RT-RTE-core-projection	SRC-1: jepa-anything-core/src/jepa_anything_core/opf.py	none	none	Exact tensor values, learned content and improved capacity unestablished.
RT-RTE-state-consumption	operational admission/selection/consumption	implemented	RT-OBJ-scaffold-manifest	no content change	State delivered to readout, analysis or repeated transition	Configured mode/callability/exogenous presence; not outcome truth	Explicit consume_state call	Callback result or selected horizon states	Direct callback execution	No truth license	Permits callbacks before any separate evidence audit	see RT-RTE-state-consumption	SRC-1: jepa-anything-skill/assets/scaffold/src/jepa_task/usage.py.tmpl	none	none	Callback-generated content/update relation not determinable; earlier callbacks not rolled back on later failure.
RT-RTE-structural-checks	check/evidence production	implemented	see record	truth-apt transformation: entailed derivation	Synthetic fixture structure and oracle-coordinate activity	Deterministic fixture calculation; seeded synthetic state domain	Recipe driver CLI/API; Make/CI wiring	Pass/fail report/status	CLI nonzero blocks successful check completion	Fixture conclusions only	Can fail make check; no automatic model successor installation	Recipe caller/Make; JSON and exit status; diagnostic and enforcing check completion; current run	SRC-1: recipes/synthetic-linear-dynamics/run_recipe.py	none	none	No learned activity or empirical world-model acceptance criterion.
RT-RTE-structural-checks	check/evidence production	implemented	see record	truth-apt transformation: entailed derivation	Publication declarations and referenced-file integrity	Manifest consistency and hash predicates; metadata/file domain	Verifier CLI/API; Make/CI wiring	Pass/fail report/status	CLI nonzero blocks successful check completion	Integrity/publication-boundary conclusions only	Can fail make check; no scientific result acceptance	Verifier caller/Make; JSON and exit status; diagnostic and enforcing check completion; current run	SRC-1: checkpoints/verify_manifest.py	none	none	No model result evaluated; merge policy uninspected.

There is no lifecycle-integration row for a scientifically accepted comparative claim: none is established by these inspected routes. This is not a system-complete absence assertion about uninspected callbacks, host processing or remaining source branches. Structural validity is an implemented, evidence-consuming admission decision for generation, not for scientific use.

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source ID/anchor and evidence layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support and design limits	Supported conclusion	Mismatch/unknown
EPI-CLM-world-model-compiler	see record	SRC-1: jepa-anything-skill/SKILL.md; doctrine/design	see record	EPI-RTE-design-diagnostics, RT-RTE-scaffold-admission, EPI-RTE-template-specialization	None supplied	No host conversion comparison or real training evidence	A declared model-assisted compiler with executable deterministic support, not observed reliable automatic task conversion	Direct renderer bypass qualifies the public refusal statement; no contradiction with the documented CLI recipe.
RT-CLM-validation-scope	see record	SRC-1: jepa-anything-skill/scripts/validate_design.py; implementation and doctrine/design	see record	EPI-RTE-design-diagnostics, EPI-RTE-evidence-envelope, EPI-RTE-measurement-checks	None supplied	No independent test of validator completeness; no model-quality intervention	Implemented declaration checks and explicit non-scientific boundary agree	A successful generated test can use copied metadata; passing must not be promoted to measured support.

Bounded conclusion

The compiler produces a candidate contract and executable obligations; its gate licenses skeleton generation, not a scientific conclusion. Host acquisition and proposal behavior are see RT-RTE-task-compilation. Imported causal facts have unknown warrant in this run. The source asks for hypotheses and assumptions; without a concrete task/design trace it does not establish which proposed propositions were entailed, reshaped or ampliative. No produced accepted ampliative model-quality claim is established.

The design-check result derives from declared input and fixed predicates, not observations of the proposed world. Acceptance has an explicit criterion and use: report.valid allows the generator CLI to emit a skeleton; destination and format checks constrain emission. See EPI-RTE-design-diagnostics and RT-RTE-scaffold-admission. Direct render calls are a supported countercase to a universal validity-gate claim. The renderer specializes declarations and retains config, report and code; later imports enforce only the branches actually called. See EPI-RTE-template-specialization and RT-BAP-generated-contract.

Planned comparison scope is inspectable, and metadata mismatches can stop a caller. Activity floors, capacity tolerances and geometry bounds are likewise material local criteria. Their intended use is contract/audit checking; none authenticates the caller's measurement process or performs statistical acceptance. See EPI-RTE-evidence-envelope and EPI-RTE-measurement-checks. The generated state consumer does not require them. Operational reliance on callback state can therefore precede any such check; see RT-RTE-state-consumption.

Independent core reports compute actual numerical properties of their tensor inputs. This is stronger than checking a supplied rank assertion but still bounded to numerical geometry and sampled reconstruction, not domain truth. See EPI-RTE-numerical-audits. The synthetic driver uses oracle coordinates; manifest verification licenses only provenance/integrity conclusions. See RT-RTE-structural-checks.

The variance tracker is a separate historical-state route: sample summaries affect later differentiable penalties, with explicit reset and update controls. See EPI-RTE-variance-history. Basis updates, QR retraction and EMA are separate parametric mutation affordances; see RT-RTE-core-projection. These support possible behavior adaptation without a truth-apt acceptance route. They establish neither observed learning nor improved capacity. Method texts are addressable guidance and requested repair is not evidenced criticism of the compiler's own method. Reflection, autonomous execution and compiler self-improvement do not follow from package generation or parameter mutability; the independently bounded runtime theory assessment is see RT-RTE-task-compilation. Missing traces and causal comparisons remain local limits, not an overall epistemic score.

Shared records

Operative objects

EPI-OBJ-claims

Label: Structured planned claims Part of: RT-OBJ-design-config

Source-native identity: claims and linked experiments entries in schema 2.0. Symbolic mappings in a config file/in-process mapping. This part has different checks and authority from free-text task assertions: comparison metadata must be reciprocal and status must be planned. It is not a measured outcome. Generated copies specialize the same declarations in the package described by RT-OBJ-scaffold-manifest. SRC-1 implementation: jepa-anything-skill/scripts/validate_design.py.

if claim.get("status") != "planned": self.fail("FIELD_INVALID", f"{path}.status", "Design-time claims must have status 'planned'.") --- jepa-anything-skill/scripts/validate_design.py

EPI-OBJ-evidence-envelope

Label: Claim evidence envelope

Source-native identity: evidence argument to audit_claim_evidence_envelope. Symbolic caller mapping in memory. This is distinct from EPI-OBJ-claims: it asserts result metadata and passed audit IDs rather than a future plan. Closest supplied object RT-OBJ-scaffold-manifest contains the checking code, not this downstream caller data. No supplied object counterpart after comparison. SRC-1 implementation: jepa-anything-skill/assets/scaffold/src/jepa_task/evaluation.py.tmpl.

if set(evidence.get("passed_audit_ids", ())) != set(claim["audit_ids"]): raise ValueError("Evidence audit IDs do not match the planned claim.") --- jepa-anything-skill/assets/scaffold/src/jepa_task/evaluation.py.tmpl

EPI-OBJ-audit-measurements

Label: Backend audit measurements

Source-native identities: separate measurements arguments for geometry, activity and capacity helpers. Symbolic numeric arrays/mappings supplied by callers, held in memory, with no persistence implemented by these helpers. The three operative parts have different targets: geometry rank/condition/NMSE; factor/encoder standard deviations; reference/baseline parameter/FLOPs/step counts. RT-OBJ-scaffold-manifest contains the helper programs, not their caller-provided inputs; no supplied object counterpart after comparison. SRC-1 implementation: jepa-anything-skill/assets/scaffold/src/jepa_task/model.py.tmpl, jepa-anything-skill/assets/scaffold/src/jepa_task/activity.py.tmpl, jepa-anything-skill/assets/scaffold/src/jepa_task/capacity.py.tmpl.

measurements must be produced by the implementation's parameter/FLOPs counters; copying design estimates is not empirical capacity evidence. --- jepa-anything-skill/assets/scaffold/src/jepa_task/capacity.py.tmpl

EPI-OBJ-audit-reports

Label: Numerical geometry reports

Source-native identities: BasisGeometryReport, RoundTripReport, TransposeSynthesisReport, and containing OPFGeometryReport. Symbolic numeric report fields in Python objects, computed from supplied dense tensors. Distinct from RT-CMP-orthogonal-factor-projection, the numerical model being assessed, and from RT-RTE-structural-checks, which covers fixture/manifest reports. No supplied object counterpart after comparison. SRC-1 implementation: jepa-anything-core/src/jepa_anything_core/audit.py.

rank = int(torch.linalg.matrix_rank(work).item()) singular_values = torch.linalg.svdvals(work) minimum_singular_value = float(singular_values.min().item()) --- jepa-anything-core/src/jepa_anything_core/audit.py

EPI-OBJ-variance-tracker

Label: OnlineVarianceTracker sufficient statistics

Source-native identity: OnlineVarianceTracker.count, .mean, .m2. Symbolic numerical sufficient statistics indexed by factor/coordinate, not learned semantic weights; PyTorch registered buffers in process and checkpointable through module state. They retain information from prior samples. Distinct from the basis parameters in RT-CMP-orthogonal-factor-projection; no supplied counterpart after comparison. SRC-1 implementation: jepa-anything-core/src/jepa_anything_core/losses.py.

self.mean.copy_(mean_for_storage) self.m2.copy_(m2_for_storage) self.count.fill_(int(total_count)) --- jepa-anything-core/src/jepa_anything_core/losses.py

Routes

EPI-RTE-design-diagnostics

Label: Design diagnostic production

Endpoints/progression: parsed mapping → DesignValidator.validate branch checks → report mapping → standalone CLI JSON/optional file, or generator CLI. Owner: deterministic validator. Context is the declared configuration; it emits facts about coded constraints, not independent world facts. Distinct checking function overlapping RT-RTE-scaffold-admission, whose identity includes admission and emission; the standalone validator also works without generation. No supersession of that valid container is needed. SRC-1 implementation: jepa-anything-skill/scripts/validate_design.py.

• Immediate return: Diagnostic mapping, valid bit, counts; standalone CLI status 0/1, or load error status 2.
• Later read-back: Optional report file can be read by caller; no automatic later reload or expiry policy is implemented here.
• Delegated visibility: inapplicable — no worker dispatch.
• Selection predicate: Explicit config argument; all dispatched validator groups run for a mapping, root-type failure returns early.
• Invalidation or expiry: Recompute for changed input; no report invalidation service.
• Activation or effect: Diagnostics are wired to generator main's validity branch; host repair is instructed, not observed.
• Evidence limits: Predicate dispatch and claim/output branches inspected; not an exhaustive audit of every schema constraint.
• implementation conclusion status: wired
• operation conclusion status: uninspected

Proposer is the host/caller; code judges declaration validity; host chooses revisions. No mandatory human per-field approval. The oracle is the coded constraint set, not an expected successful design. Mode is open task conversion with deterministic repair feedback. The check admits no change by itself; the separate generator gate consumes it.

Concrete implementation deduction, without execution: a parsed JSON [] is not a mapping. It receives SCHEMA_REQUIRED at $, one failed check, valid: false; standalone main returns 1 and generator main returns 1 without rendering. This licenses the proposition that this input violates the implemented root contract, not that the validator knows whether a real task is causally correct.

if not isinstance(self.config, Mapping): self.fail("SCHEMA_REQUIRED", "$", "The root value must be an object.") return self.report() --- jepa-anything-skill/scripts/validate_design.py

"valid": self.error_count == 0, --- jepa-anything-skill/scripts/validate_design.py

EPI-RTE-template-specialization

Label: Template specialization Part of: RT-RTE-scaffold-admission

Endpoints: config fields → _template_values / _python_repr → placeholder replacement → generated modules and normalized config. Owner: renderer. This part changes representational surface, not scientific support. SRC-1 implementation: jepa-anything-skill/scripts/generate_scaffold.py.

• Immediate return: Rendered text internally; enclosing renderer returns generated file list.
• Later read-back: Generated literal constants are imported by downstream modules/tests; config and report copies remain on disk until caller edits/removes them.
• Delegated visibility: inapplicable — no delegated execution.
• Selection predicate: Fixed template keys select config values; every .tmpl under the shipped template root is rendered.
• Invalidation or expiry: Regeneration is explicit; missing placeholders raise; no automatic synchronization after edits.
• Activation or effect: Constants wire downstream checks and dispatcher modes; actual generated package adoption unobserved.
• Evidence limits: Non-ampliative classification applies to literal declaration transfer, not arbitrary downstream behavior of callbacks.
• implementation conclusion status: wired
• operation conclusion status: uninspected

Concrete text trace: for an otherwise renderer-shaped configuration with task.id equal to pendulum, _python_repr(task["id"]) yields the Python string literal 'pendulum'; TASK_ID_REPR replaces the corresponding placeholder. For usage.mode equal to repeated_transition, USAGE_MODE_REPR similarly becomes 'repeated_transition', which the dispatcher consumes through generated contracts. The emitted literal asserts no new empirical property. This deduction uses inspected text, not target execution. Generation admission and recovery are see RT-RTE-scaffold-admission.

"TASK_ID_REPR": _python_repr(task["id"]), --- jepa-anything-skill/scripts/generate_scaffold.py

TASK_ID = {{TASK_ID_REPR}} --- jepa-anything-skill/assets/scaffold/src/jepa_task/contracts.py.tmpl

for key, value in values.items(): rendered = rendered.replace("{{" + key + "}}", value) --- jepa-anything-skill/scripts/generate_scaffold.py

EPI-RTE-evidence-envelope

Label: Planned coverage and evidence-envelope checks

Endpoints: generated claims/experiments → assert_planned_claim_coverage; separately caller metadata → audit_claim_evidence_envelope → return/exception. Owner: fixed Python comparisons. Two checking functions, separately assessed in the ledger. These are material evaluator parts inside the package in RT-OBJ-scaffold-manifest, distinct from the generation/admission path RT-RTE-scaffold-admission. No supplied route counterpart after comparison. SRC-1 implementation: jepa-anything-skill/assets/scaffold/src/jepa_task/evaluation.py.tmpl.

• Immediate return: None on matching declarations/metadata, otherwise exception.
• Later read-back: inapplicable — helper neither retains results nor updates claim status.
• Delegated visibility: inapplicable — no delegation.
• Selection predicate: All planned claims for static coverage; explicit claim_id and experiment_id for supplied result metadata.
• Invalidation or expiry: Each call checks current imported declarations; no expiry or withdrawal record.
• Activation or effect: Generated contract test calls both; consume_state does not call either.
• Evidence limits: No score/significance test; passed audit IDs are compared as sets, not authenticated against audit results.
• implementation conclusion status: wired
• operation conclusion status: uninspected

Caller proposes evidence; deterministic code can veto metadata mismatch by exception. There is no scientific answer oracle or statistical acceptance criterion, and no mandatory human approval. Mode is structural contract testing, not an experimental curriculum. The source's own test constructs evidence by copying claim fields, including audit IDs. Thus a successful call can occur without any measured outcome; this concrete countercase defeats an interpretation that passing means scientific support.

This stub intentionally does not decide statistical support. It rejects a result measured with a different metric, direction, split, horizon, model, use mode, baseline set, or audit set. --- jepa-anything-skill/assets/scaffold/src/jepa_task/evaluation.py.tmpl

"baseline_ids": claim["baseline_ids"], "passed_audit_ids": claim["audit_ids"], --- jepa-anything-skill/assets/scaffold/tests/test_contract.py.tmpl

EPI-RTE-measurement-checks

Label: Backend measurement envelope checks

Endpoints: three independent caller mappings → geometry/activity/capacity helpers → return or exception. Owners are fixed Python predicates. Each target and operational path is split in the ledger. Closest supplied RT-RTE-structural-checks concerns repository fixture/provenance drivers, not these generated downstream helpers. RT-OBJ-scaffold-manifest is their program container, not a duplicate route. SRC-1 implementation: jepa-anything-skill/assets/scaffold/src/jepa_task/model.py.tmpl, jepa-anything-skill/assets/scaffold/src/jepa_task/activity.py.tmpl, jepa-anything-skill/assets/scaffold/src/jepa_task/capacity.py.tmpl.

• Immediate return: None or target-specific ValueError.
• Later read-back: inapplicable — no result retention or claim update.
• Delegated visibility: inapplicable — no delegation.
• Selection predicate: Explicit caller chooses which audit to invoke and supplies values; configured AUDITS and capacity contracts set thresholds.
• Invalidation or expiry: Current call only; no link to changed backend weights or automatic staleness detection.
• Activation or effect: Helper exceptions constrain their explicit callers; no automatic requirement before callback consumption.
• Evidence limits: These helpers validate supplied values, not the process that measured them; no supplied measurement trace.
• implementation conclusion status: wired
• operation conclusion status: uninspected

Caller/backend proposes measurements; code decides local threshold compliance; there is no human approval branch. Answer references are declared floors, ceilings, scopes and tolerances, not expected task predictions. Operating mode is bounded numerical envelope checking. No accepted scientific claim or model successor is installed, and rollback of external backend state is not implemented.

if condition_number > geometry["max_condition_number"]: raise ValueError("Analysis map exceeds the declared condition-number ceiling.") --- jepa-anything-skill/assets/scaffold/src/jepa_task/model.py.tmpl

if inactive_factors: raise ValueError(f"Inactive projected factor coordinates: {inactive_factors}.") if inactive_encoder: raise ValueError(f"Inactive online-encoder coordinates: {inactive_encoder}.") --- jepa-anything-skill/assets/scaffold/src/jepa_task/activity.py.tmpl

if steps != reference["training_steps"] or steps != contract["training_steps"]: raise ValueError("Baseline and reference training steps must match exactly.") --- jepa-anything-skill/assets/scaffold/src/jepa_task/capacity.py.tmpl

EPI-RTE-numerical-audits

Label: Numerical tensor audits

Endpoints: basis/state/projection → detached CPU float64 geometry computations and reconstruction → reports. Owner: numerical library. Context is supplied tensors, precision and tolerances. Distinct from RT-RTE-core-projection, which computes/transforms state and mutates parameters; distinct from fixture/provenance checks in RT-RTE-structural-checks. SRC-1 implementation: jepa-anything-core/src/jepa_anything_core/audit.py.

• Immediate return: Numerical reports with pass/violations; malformed/nonfinite inputs can raise.
• Later read-back: uninspected — caller may retain reports, but no later consumer of accumulated reports is established here.
• Delegated visibility: inapplicable — no worker dispatch.
• Selection predicate: Caller chooses audit and tolerances; aggregate audit invokes basis, factor, round-trip and transpose checks.
• Invalidation or expiry: Report covers supplied inputs; caller must rerun for changed tensors.
• Activation or effect: Returns diagnostics; does not alter projection, accept scientific claims or enforce a caller response.
• Evidence limits: Actual sampled numerical inputs and runs not supplied; factor-audit internals not assessed here.
• implementation conclusion status: wired
• operation conclusion status: uninspected

For basis checking the target is a finite floating tensor with three positive dimensions; numerical rank and orthogonality/completeness errors are compared with dimensions/tolerance. For round trips the target is reconstruction of the supplied state, not prediction of an unseen world. Caller proposes tensors and criteria; program computes verdicts; no human decision is encoded. Numerical reference is the same input state/dimensional geometry, not domain labels. Reporting success has advisory force only.

reconstructed = compose_state(decompose_state(state, basis), basis) max_error, rms_error, normalized_mse = _reconstruction_statistics(state, reconstructed) --- jepa-anything-core/src/jepa_anything_core/audit.py

passed=( basis_report.passed and factor_report.passed and round_trip_report.passed and transpose_report.passed ), --- jepa-anything-core/src/jepa_anything_core/audit.py

EPI-RTE-variance-history

Label: Variance history and streaming penalty

Endpoints: factor samples → finite/shape checks → Welford update of count/mean/m2; later samples plus stored statistics → differentiable combined variance → std-floor penalty. Owner: caller-invoked tracker/loss functions. Distinct from RT-RTE-core-projection: these are sample-summary buffers and losses, not basis/target parameter mutations. No supplied counterpart after comparison. SRC-1 implementation: jepa-anything-core/src/jepa_anything_core/losses.py.

• Immediate return: update mutates buffers; streaming function returns a loss and optionally updates history.
• Later read-back: Subsequent streaming calls combine current samples with historical count/mean/m2; registered buffers make checkpoint retention possible but no actual save/reload is evidenced.
• Delegated visibility: inapplicable — no delegation.
• Selection predicate: Explicit update argument; OnlineVarianceLoss defaults to updating in training mode only.
• Invalidation or expiry: reset zeroes all accumulated statistics; invalid/nonfinite/overflowing updates reject before buffer writes; no age-based expiry.
• Activation or effect: Retained statistics directly enter later combined variance and loss; no optimizer is invoked by this route.
• Evidence limits: Floating-point summary accuracy and actual caller usage are not empirically measured here; improved future action and learning remain unestablished.
• implementation conclusion status: wired
• operation conclusion status: uninspected

The target is per-coordinate activity over historical plus current samples. Evaluator is a numeric std-floor penalty, not a scientific accept/reject decision. Caller controls update/reset and optimizer use; no human content approval or successor selection is encoded. The separate composite objective explicitly performs no optimizer/EMA update, so differentiability is not evidence of an executed policy change. This is trace-derived numerical history with an implemented later consumer, not automatically an accepted theory or knowledge claim.

coordinate_variance = tracker.combined_variance(factors) standard_deviation = torch.sqrt(coordinate_variance + eps) penalties = torch.relu(min_std - standard_deviation) loss = _reduce(penalties, reduction) if update: tracker.update(factors) return loss --- jepa-anything-core/src/jepa_anything_core/losses.py

self.count.zero_() self.mean.zero_() self.m2.zero_() --- jepa-anything-core/src/jepa_anything_core/losses.py

This pure function intentionally performs no optimizer or EMA update. It is suitable for generated task skeletons while keeping training lifecycle and target-encoder stop-gradient decisions explicit at the call site. --- jepa-anything-core/src/jepa_anything_core/losses.py

Claims

EPI-CLM-world-model-compiler

Label: World-model compiler with bounded validation

Claimed operation: natural-language world-model task → validated design and inspectable no-training skeleton; invalid designs and nonempty destinations are refused along the documented workflow. Claimed warrant is expressly internal completeness, not empirical success. This compiler-facing claim is distinct from, and bounded by, validator-specific RT-CLM-validation-scope. Source: SRC-1 doctrine/design, jepa-anything-skill/SKILL.md; implementation support and direct-render countercase are see RT-RTE-scaffold-admission.

Treat all design choices as hypotheses. --- jepa-anything-skill/SKILL.md

The generator refuses invalid designs and non-empty destinations. Its adapter, OPF state, geometry/activity audits, mode consumer, capacity audit, claim audit, and contract tests contain no optimizer or training loop. --- jepa-anything-skill/SKILL.md

A passing validator proves only that the declared design and evidence plan are internally complete. It does not prove activity, synthesis quality, semantics, model quality, or claim support. --- jepa-anything-skill/SKILL.md