Kai9987kai/Supermix-Poseidon

Supermix Poseidon v3.12.0: Local Research Workbench

Version 3.12.0 packages a local experimental workbench around the frozen Tidal core, with a separate pretrained conversation backend, optional synthetic controllers, and mathematical research tools. Runtime availability, replay integrity, and measured model performance are separate checks. Additional modules do not imply that the selected core was retrained or that a candidate was promoted.

TRIDENT remains available as an opt-in identity-gated controller around the unchanged Tidal checkpoint. It corrects one-step predictions with recent observed prediction errors, admits directed travel only after enough local visits unless that transition has been observed, and requires an override to beat erased and time-shifted residual controls by configured margins. TRIDENT is experimental software for synthetic TidePool; it does not modify or train model weights.

The six-arm paired benchmark compares the frozen core, TRIDENT, erased residual, shifted residual, directed-support-gate-off, and residual-gate-off controls on identical seeds. The receipt stores actions and transitions and independently replays controller decisions and world outcomes. Its four-seed results are descriptive synthetic evidence, not proof of a general performance gain; they do not promote a model.

Additional modules are software simulations or heuristics: Archimedes computes synthetic fluid telemetry; Genesis runs an auxiliary toy ecology; S-Video encodes telemetry; Universal Modder applies static Python hooks; QuantumBot uses a classical state-vector simulator. They do not provide physical RF harvesting, NTAG hardware, analog video, hot-patching, quantum hardware, biological evidence or real-world survival capability. Causeway uses deterministic maximum-probability selection in this controller.

Source commit · TRIDENT design · TRIDENT results · TITAN design · TITAN results · HYPERION results · Morpheus design · Prometheus design · Intermittent design · CHIMERA results · Causeway design · Release guide

Components and activation

Component	Artifact	Status
Tidal core: 327,230 parameters	runs/tidal_dagger/core.pt	Existing selected DAgger policy, default scene/control backend
Supervised baseline	runs/tidal/core.pt	Comparison checkpoint, original embedded training state
SmolLM2: 134,515,008 stored parameters	models/language/model.safetensors	Separate unchanged pretrained conversation backend
LoRA: 61,440 parameters	runs/language/adapter/	Experimental candidate, inactive by default
Poseidon-X	poseidon/poseidon_x.py	Experimental geometric and homeostatic controller; opt-in
Problem Forge	poseidon/problem_solver.py	Bounded exact and numerical tools; 95 verified domains
Project NEUROGRAFT-II	poseidon/neurograft_engine.py, poseidon/research/successor_memory_adapter.py	Successor representations (SR-TD), zero-shot reward revaluation, SWR reverse replay, Graph Laplacian spatial embeddings & trainable graph circuits
Project CROSSWIRE	poseidon/crosswire_engine.py, poseidon/research/crosswire_adapter.py	Version space tracking, active diagnostic probing, all-hypothesis contract verification & closed-loop runtime shielding
Project FAULTLINE	poseidon/faultline_engine.py, poseidon/research/faultline_adapter.py	Condition-stratified joint failure selection, 3-bit parity hyperedge audit, continuous error covariance & minimax panel triage
Project TETHER	poseidon/tether_engine.py, poseidon/research/tether_gate.py	Continuous reliability weighting, 4-view causal invariance, input x gradient saliency, consensus isolation & fallback gate
CHRONOS, NOETHER, and CARTAN	poseidon/chronos.py, poseidon/noether.py, poseidon/cartan.py	Experimental mathematical constructions and synthetic telemetry
Project HYPERBOREA	poseidon/hyperborea.py	Cellular sheaf cohomology, Port-Hamiltonian Dirac structures, Wilsonian RG & Fibonacci anyon braiding
Project HEPHAESTUS	poseidon/hephaestus.py	Wake-sleep library synthesis, GFlowNet DAG sampling, Koopman EDMD & persistent homology
Project ATHENAEUM	poseidon/athenaeum.py	Equality saturation e-graphs, Process-Reward MCTS & SINDy equation discovery
Project ARCHIMEDES & DA VINCI	poseidon/proof_engine.py, poseidon/creativity_engine.py	Formal natural-deduction proof certificates & Koestler/TRIZ bisociative invention
Project OLYMPUS & PANTHEON	poseidon/hypothesis_engine.py, poseidon/pantheon.py	Popperian crucial experiment falsification & 24-pillar universal synthesis
Clifford Geometric Algebra	poseidon/geometric_algebra.py	Complete Cl(3,0) multivector engine, Rotor3D orientations & fluid vorticity
Autopoietic Affective System	poseidon/autopoiesis.py	Autopoietic homeostasis, PAD affective appraisal & micro-sleep self-repair
Hypergraph Theorem Discovery	poseidon/hypergraph_discovery.py	Neural-symbolic inductive logic programming with Laplace confidence
Continuous 3D OmniEcosystem	poseidon/ecosystem_3d.py	Multi-agent trophic dynamics (grazers, predators, symbionts) & duty-cycled pulse signaling
Tessera LTL Prover	poseidon/tessera_prover.py	Bounded synthetic contract and invariant checks
Universal Modder & Reactive Actors	poseidon/universal_reactive.py	Static Python hooks with rollback and reactive message passing
Active Inference & Free Energy	poseidon/active_inference.py	Expected free energy minimization, epistemic curiosity & reserve override
Metacognitive Monitor	poseidon/metacognition.py	Metacognitive error calibration & dynamic confidence monitoring
TRIDENT Identity-Gated Controller	poseidon/trident.py	Identity-gated, observed-topology causal residual controller around frozen core; opt-in
TRIDENT Paired Replay Receipt	outputs/trident_experiments/RECEIPT.json	Six-arm matched-seed controller and world replay; no promotion
TITAN Composite Controller	poseidon/titan.py	Selected project mechanisms around the frozen Tidal core; opt-in
TITAN Paired Replay Receipt	outputs/titan_experiments/RECEIPT.json	Six-arm matched-seed controller and world replay; no promotion
HYPERION Super-Controller	poseidon/hyperion.py	Separate opt-in experimental controller
Morpheus Consolidation & Replay	poseidon/morpheus.py	Simulated trace compression and replay summaries
Prometheus Meta-Plasticity	poseidon/prometheus.py	Rolling Shannon entropy meta-plasticity & 4-drive homeostatic allostasis
Intermittent Computing	poseidon/intermittent.py	Software model of harvested-energy state and compact frames
NexusSearch Federated Index	poseidon/nexus_search.py	Dual HRR associative vector & 3D diamond lattice spatial search
HYPERION Receipt	outputs/hyperion_experiments/RECEIPT.json	Paired multi-arm evaluation benchmark receipt
CHIMERA Super-Controller	poseidon/chimera.py	Unified quantum superposition, holographic memory & diamond lattice, opt-in
Causeway Decision Engine	poseidon/causeway.py	6D complex superposition, non-Hermitian gain & Born collapse
Holographic Concept Memory	poseidon/hologram.py	256-d HRR circular convolution binding & associative clean-up
Diamond Lattice Neuropil	poseidon/lattice.py	3D FCC tetrahedral geodesic multi-scale grid cells
CHIMERA Receipt	outputs/chimera_experiments/RECEIPT.json	Paired multi-arm evaluation benchmark receipt
METAMORPH Controller	poseidon/metamorph.py	Unified developmental ecdysis, causal flux & cyclic sync controller, opt-in
MOLT Morphogenesis	poseidon/molt.py	Larval/pupa/imago developmental instars & exuvia hashing
NexusFlow Causal Flux	poseidon/nexusflow.py	Hydraulic replenishment gradients & superposition interference checking
Chronos & Ghost Auditor	poseidon/chronos.py	Simulated periodic ticks and counterfactual SDI tracking
METAMORPH Receipt	outputs/metamorph_experiments/RECEIPT.json	Paired multi-arm evaluation benchmark receipt
AURA Biomimetic Controller	poseidon/aura.py	16-wedge CX-ring attractor & 128-KC sparse neuropil arbiter, opt-in controller
Tessera Macro-Commons	outputs/tessera_experiments/RECEIPT.json	Independent dual-lineage ratified macro-actions, opt-in commons
Mnemorph Memory Assays	outputs/mnemorph/RECEIPT.json	Typed structural lesion and associative regrowth receipts
Odysseus Navigator	outputs/odysseus/atlas.json	Empirical macro-action transitions & Bayesian replenishment LCB, opt-in candidate
Helm Critic v0.6	outputs/helm/atlas.json	Fitted multi-horizon critic with observed history, opt-in candidate
Odyssey Atlas v5	outputs/odyssey/atlas.json	Spatial cognitive map and replenishment-decay waypoint planner, opt-in candidate
Horizon Atlas v4	outputs/horizon/atlas.json	Fitted multi-horizon return advantage, opt-in planner
Contrast Atlas	outputs/contrast/atlas.json	Matched-action residual memory, opt-in candidate
Counterfactual Atlas	outputs/atlas/atlas.json	Fitted residual memory, opt-in candidate

Tidal uses 512 hashed lemma/bigram features, a 128-wide state, four gated experts, and two recurrent updates within each decision. It predicts five controlled scene attributes, six simulation actions and 16 state deltas. Its trained routing is dense across four experts; experimental sparse/ensemble modules are separate. The default protocol resets hidden state at each decision. Fluent conversation, procedural geometry/rendering, external carrier storage and exact maths use separate components. No unified multimodal, real-world survival, biological connectome, consciousness or general autonomous-learning capability is claimed.

Download and run

python -m pip install huggingface-hub
hf download Kai9987kai/Supermix-Poseidon --revision v3.12.0 --local-dir Poseidon-release
cd Poseidon-release
python -m pip install -e .
python tools/verify_huggingface_release.py . --smoke
python -m poseidon serve --port 8787

Open http://127.0.0.1:8787. Use Python 3.10+ with a compatible CPU PyTorch 2.6+ installation. The snapshot includes weights and tokenizer; no additional model download is needed for the workbench. Native checkpoints use torch.load(..., weights_only=True) and language loading disables remote code. AutoModel.from_pretrained('Kai9987kai/Supermix-Poseidon') is unsupported: use the packaged runtime. For immutable provenance, pin the Hub commit SHA returned by the release upload.

from poseidon.runtime import Poseidon
runtime = Poseidon('.')  # run from the downloaded release directory
print(runtime.respond('Survive', mode='world', seed=42, planner='policy'))
print(runtime.respond('Survive', mode='world', seed=42, planner='odyssey'))
print(runtime.respond('Survive', mode='world', seed=42, planner='horizon'))

Default language uses pinned HuggingFaceTB/SmolLM2-135M-Instruct at 12fd25f77366fa6b3b4b768ec3050bf629380bac. Explicitly test the candidate with python -m poseidon chat 'Explain why seasons change.' --adapter runs/language/adapter. Generation uses at most a 1,024-token local input budget and 256 generated tokens. This release does not automatically activate the adapter, Atlas, Contrast, Horizon or Odyssey.

Training and measured limits

The supervised Tidal run used 120,000 synthetic examples and 360,000 exposures. DAgger selected round two using validation seeds; 769,600 exposures is the aggregate over the four-round run, not a precise exposure count of the selected weights. Stored evaluation reports describe controlled scene vocabulary and synthetic TidePool survival. Reported 100% results do not imply real-world generalization.

Odyssey Atlas v0.5 constructs an episodic spatial and topological cognitive map strictly from 16-d observation streams without simulator cheats or global coordinate leakage. It tracks environmental replenishment dynamics ($\lambda_{food} = 0.007 / scarcity, \lambda_{water} = 0.014 / scarcity$), plans multi-hop navigational routes via breadth-first search with physical travel cost penalties, enforces pre-transit rest guards when stamina < 0.20, and prioritizes shelters during impending storms. The map's scarcity now matches the selected world settings. The v0.5.0 scarcity-4 suite used the default map setting of 2.5 in 12 episodes; its preserved measurements have that limitation.

Two fresh four-seed, ten-arm, 64-step suites used scarcity 2.5 and 4.0 with no tuning between them. Calibrated Odyssey matched policy with zero overrides. Ungated Odyssey's mean episode reward differences were +0.002331 and +0.041827, but the stress result was dominated by one seed. The separate H=16 auditor replayed 91 anchors, 546 actual multi-step branches and 8,712 branch transitions. Six of 11 ungated overrides had negative H-step return advantages. These small synthetic development suites do not establish superiority or a safety guarantee. The candidate remains opt-in and the core policy remains the default.

The historical experiment audit forked six one-step reserve-utility branches; it did not validate multi-step reward margins. The new audit records every continuation action and replays returns without neural weights. It verifies execution and arithmetic, not neural authorship. Its strict parent checks also reconstruct maps, validate fitting partitions, paired cardinality, forecasts and terminal states. See docs/TRAJECTORY_AUDIT.md and docs/trajectory-evaluation.json.

Horizon Atlas v0.4 indexes multi-horizon returns over H=16 lookahead steps across 12 training episodes (1,728 transitions), evaluates physical depletion guards, and calibrates empirical advantage error radii across 8 disjoint calibration episodes (1,152 transitions). Returns use undiscounted rewards. Empirical radii describe the sampled fitting data and are not calibrated failure probabilities. Historical episode measurements and their one-step audit limits are in docs/HORIZON_RESULTS.md.

The candidate LoRA ran 256 steps, saw 512 examples and 42,104 tokens, and changed development loss from 0.960879 to 0.958487. That does not establish better conversation or reasoning; the dataset overlaps the upstream model's instruction-training distribution. The candidate remains inactive. The conversation model can hallucinate.

Provenance and file integrity

Source revision: 37a62518f8e364e39fa1dbb3c8b85262ead142c7. Active core file SHA-256: a053104ae63dc7f86baa15b4b34e46730e1ccb92117f221801fb21f90d2d6c27. Atlas canonical payload SHA-256: db9a4bc78e738ca55e61ec0511ea48814b13ed0ba48f8264f8c43cfda7205dc8. Contrast canonical payload SHA-256: 7ddbb687b5638c52b1f9cfa27fcace6ce801a549ba5227515f87bfa5b4f45002. Horizon canonical payload SHA-256: 44eb5abcf7bd2c4cd206645b53721f69ee6405e924794826f25c5fb2f5a90e69. Odyssey canonical payload SHA-256: 6c0285ac052a047bf961f43babf1709b8f698563bcc6790cac2a83a2bde85f5b. Helm canonical payload SHA-256: a4b7dcdb3579d2931350073e4195a1e5879b2e802f81e5b69af7152138066f5c. TITAN receipt SHA-256: 34ec27bcae3c602aa938c4c34731e733de26d20e668b3357953b69f47f19692e. TRIDENT receipt SHA-256: 5d91e7016c4ee10b3217249fb605a13f6a4b390804c653141b88847b8b3b6027. manifest.json and SHA256SUMS.txt inventory the staged files. The manifest binds source, weights, upstream provenance and the exact experiment receipts. Checksums diagnose integrity; they do not prove scientific validity or authorship. SOURCE_README.md is the project's operational README. Root core.pt, core_base.pt, and adapter/ are compatibility copies of the native layout. User memory, raw language datasets, caches, logs, rejected Atlas artifacts and language optimizer continuation state are excluded.

Superseded historical release

The old ensemble_adapted.pt is archived as legacy/ensemble_adapted.pt and retained in Hub history. Its archived bytes are inventoried for integrity, without any capability validation, and are not loaded by the default runtime. Its former promotion claim is invalidated. The old demonstration video is also archived. beyond_audit_receipt.json now records supersession, while original evidence and the previous model card are preserved under legacy/. The earlier card's claims of authentic fly-connectome wiring, calibrated failure probabilities, learned delayed-memory performance, zero-width survival intervals and verified ensemble promotion are not supported by this release. See the repaired source documentation.

License and attribution

Original Poseidon code, Tidal checkpoints and Atlas are MIT (LICENSE). Bundled SmolLM2 weights/tokenizer, source card and related adapter retain Apache-2.0 terms and attribution (licenses/APACHE-2.0.txt, models/language/README.md). The local adapter used HuggingFaceTB/smol-smoltalk at f73fe857d519ff6ac5af2ea67c4d3834da7b8bcc; its source card and selection manifest are included, without raw dataset redistribution. Dependencies retain their own licenses. See THIRD_PARTY.md. The Hub metadata uses a composite license label to avoid relicensing upstream artifacts as MIT.