OSH-Sense

Sparse Horizon Sense Inference — multi-step soft-state over large discrete inventories when only a few skills are live.

connecting…

Live demo · one client at a time

Inventory Barrier Live

You play Blue against a locally trained AlphaGo-lite opponent. Each move still fires a sparse skill from an N≥8192 bank — OSH keeps inference ms-class (k ≪ N).

Contact / pilot — · Not in queue

Scale: where Sense pays

Low N: dense wins (OSH overhead). High N: dense grows steeply; OSH stays near-flat — the tool to scale soft-state.

Time · median evolve (ms) vs N

Measured · realtime + context-scale · log-y · OSH / CSR / dense

Memory · peak RSS (MB) vs N

Measured · footprint bench · log-y · dense / CSR / OSH

Problem

Soft-state evolution over large discrete inventories is expensive when only 1–2 skills are live. Buyers either pay full dense multi-step cost, or drop to non-evolving heuristics and lose the multi-step path.

Solution

OSH-Sense keeps a low-density live register and runs expand → evolve → unique support. Same family as multi-step soft-state — sparse when support is small.

Lean proofs · lossless vs choice

  • Lossless regime — Lean OSHoracle + SparseSimulationDensityCrossover: when live support density is low, sparse expand→evolve matches dense amplitudes (error bound in the digital model) and is the cheaper representation.
  • When lossless isn’t possible — algorithmic choice: stay sparse on the live register; route outliers / high-density mass to dense (Hybrid / H-SBD). Clifford distill Lean scaffold: partition skimmable vs outlier blocks with bounded reconstruction error — not a fake global lossless factorization.
  • Charts below are the engineering crossover; Lean is the when is sparse exact / when do we switch certificate.

Sell headlines (measured)

  • Crossover ~N=1–2k — below that dense can win; above, OSH latency stays ~flat while dense climbs
  • ≥3× median vs dense inventory (N=2048 suite) — measured ~180×; ~7× vs CSR
  • Scale: ~268× @ N=8k latency · dense peak RSS ~244× OSH @ N=32k
  • SemEval ALL: OSH 58.2% vs Lesk high bar 60.6% (non-evolving)

See time + memory charts ↑

1v1 brain

Train offline (local self-play). Live is you vs AI — you pick doctrine + hex; the checkpoint answers as Red. OSH-Sense still names the live skill from the 8k inventory and reports measured evolve_ms.

Partner surfaces

  • Coding agents — tool/mode soft-state; gate expensive model calls
  • Agent side-controllers — skill/policy inventories on CPU
  • Edge skill/mode registers — large named catalogs under RAM pressure

This demo’s AI design

N skills—
Core tactics—
Per-tile option pads—
Board—

Speed is measured only in the live session — not claimed on this page.

This box · runtime

Pinned CPUs—
Host cores—
Host RAM—
Demo RSS—
Worker1 dedicated thread

Live HUD shows measured evolve_ms vs dense wall each move — not marketing copy.

Families


        

        

          
Sample skills

          

        

        

          
Talk to us

          
Design-partner pilot, NDA evidence pack, or exclusive license / IP. Leave a note — or email directly.

          

            steven@disregardfiat.tech