Skip to content

DESIGN.md

Latest commit

 

History

History
77 lines (62 loc) · 3.34 KB

DESIGN.md

File metadata and controls

77 lines (62 loc) · 3.34 KB
Raw

Hyperlex Design Principles

1. Real Over Synthetic

All analysis must be traceable to real signals. Synthetic data is only for internal unit tests and explicitly marked.

2. Provenance is Sacred

Every output carries:

  • canonical_hash
  • timestamp
  • version
  • ingest_source
  • arxiv_concepts_applied
  • integrity receipt hash

3. ArXiv-Grounded

Core algorithms distilled from:

  • Neologism pipeline (2605.06426)
  • Semantic variation drivers (2210.08635)
  • Virality/diffusion (2510.05761)
  • Memetics protocol (2407.11861)
  • Hyperstition loops (2410.23794)
  • Cultural transmission (2203.00715)

4. Strict Interfaces

Public API is minimal and stable:

  • ingest_signal(query, source)
  • detect_memetic_patterns(...)
  • mock_integrate_with_external_signal(result)
  • emit_receipt(result)
  • extract_forecasts / settle / score_series (calibration)
  • Helper scorers

5. Modularity for Integration

Analysis blocks are independent so they can be:

  • Used standalone
  • Wired into larger Hermes runes
  • Fed into forecasting or market-signal systems
  • Composed in different orders

6. Humanizer Layer

Light post-processing to remove AI-isms while preserving sharp signal.

7. Receipt-Centric Workflow

The emit_receipt pattern ensures every serious run produces an auditable artifact. Receipts are the primary output for serious use.

8. Hermes Skill + Modular Host Compat

Hyperlex is a Hermes skill implemented in this Python package repository. It has no hard dependency on Hollersports, Abraxas, Orchestra, or any domain host. Betting slang is the initial validation domain.

Relevant Abraxas capabilities (Brier ledger/score packets, claim labels, operator review, HLX rune envelopes) are implemented as Hyperlex modules under hyperlex.compat.abraxas. Hosts may import hyperlex.compat.abraxas; Hyperlex never imports Abraxas. See docs/hermes-skill.md.

9. Determinism + Graceful Degradation

  • Mock mode is fully deterministic
  • Real modes fall back gracefully
  • Errors surface explicitly in output

10. Evolution via Receipts

Future improvements are validated by comparing receipt lineages and Brier evolution over settled forecast series.

11. Lineage as First-Class Structure

Slang is documented and analyzed as historical families with emergent branches rather than flat term lists. Visual diagrams live under examples/slang-families/ and are described in docs/slang-lineages.md. Lineage confidence is the primary continuous signal eligible as a forecast probability.

12. Brier Requires Settlement

A Brier score is never emitted from an open analysis run. Analysis may emit forecast probabilities (via extract_forecasts). Only after a settlement records an outcome may score_pair / score_series compute atomic BS, series BS, BSS, Murphy, and Yates decompositions. Missing outcomes yield NOT_COMPUTABLE, never a fabricated float. See docs/brier-calibration.md.

13. Phase 5 Simulation Is Speculative Research

Cultural transmission, multi-agent memetics, hyperstition risk, and phylogeny scaffolds are research tooling. They:

  • always label provenance: SPECULATIVE (or INFERRED for structural phylogeny)
  • never emit numeric Brier (brier: null)
  • never rewrite historical receipt integrity
  • never auto-settle forecasts

Risk tiers inform operator diligence (scan frequency, archive, settlement criteria) — not automated market action. See docs/phase5.md.