SubRep Deep Dive

help edit space_dashboard
Approved by Ursula Addison on 2026-06-29
ai generated
open_in_full page
fullscreen modal
edit edit
space_dashboard advanced
human approved
open_in_full page
fullscreen modal
edit edit
space_dashboard advanced

Standalone deep-dive card (outside the curated topic Index). Cross-Org Cognitive Repo Sweeps, SC-6 (iCog-Labs) Source 2.

Verdict: [SUBREP-CERTIFIED-SKILL-DISCOVERY-HYPERON-CERTIFICATE-STORE] · implementation-backed.

SubRep (Subgoal Refinement and Representation Learning; iCog-Labs-Dev/subrep, pin d128ba91) turns skill discovery into a certificate-driven, auditable process. A neural generator proposes skills from experience; two mathematical admission tests (CDS and PDS) must pass before a skill enters the library, guaranteeing composition safety across motive shifts and preventing negative transfer. Certified skills are stored as native MeTTa Atoms in a real Hyperon space. It is the first implementation-backed SubRep in the Hyperon-ecosystem sweeps.

Architecture

  • Neural Skill Generator — a 2-head MLP (a shared trunk feeding a payoff head and a motives head), trained with TD-error signals; a Motive-Decomposition-Network (MDN) interface and runtime selector sit alongside it. (The model file is a defined skeleton; some training/runtime selection work is on feature branches.)
  • CDS / PDS certification gates — the admission core. CDS (Conditional Dominance / "Universal Benefit") requires the skill to help on the payoff and not hurt on any motive: Δr + min(Δn) ≥ 0 (worst-case vertex over a weight set). PDS-ε ("Acceptable Trade-off") relaxes this to an ε-bounded inferior: inf_w [Δr + w·Δn] ≥ -ε. Both gates validate inputs and are exercised by pass/fail/boundary tests; the skill library re-verifies the certificate math at admission time and rejects forged margins.
  • Hyperon certificate store — a CertificateStore backed by a real Hyperon MeTTa space (requirements: hyperon>=0.2.0). A bridge converts Certificate dataclasses to/from canonical (Certificate (field value) …) expression atoms (add / query / serialize), persists them to a .metta file, and round-trips them under test.
  • Validation environment — Multi-Objective LunarLander (mo-gymnasium). Integration tests run real episodes, compute payoff/motive deltas, gate them through CDS/PDS, store admitted certificates, and report admission/rejection rates.

Hyperon coupling (precise)

The Hyperon coupling is real but scoped to certificate atom storage and query — the bridge imports hyperon, instantiates MeTTa(), builds certificate Atoms, and stores/queries them in a Hyperon space, with dedicated round-trip tests. It is not a full Hyperon skill-execution runtime: skill generation, certification math, and the RL loop run in Python/PyTorch; Hyperon is the certificate persistence/query layer. The README's "for future Hyperon integration" understates the code at HEAD — the certificate store is wired now; broader Hyperon-native skill execution/reasoning is the future-facing part.

Paper alignment

SubRep is a paper-aligned implementation of the Goertzel 2025-09 SubRep concept: the CDS/PDS gates cite "SubRep Paper Section 3.2, Definitions 1 and 2" in both code and tests. It is not AIRIS-derived — there is no AIRIS import or dependency; the implementation line is MO-LunarLander + PyTorch + certificate gates + Hyperon certificate atoms. This directly resolves the AIRIS cluster-pilot narrative guard [SUBREP-AIRIS-EXTENSION-NOT-SOURCE-SPEC]: SubRep is a separate certificate/admission-gate line, not part of the Cook & Hammer 2024 AIRIS source spec. See AIRIS Deep Dive for the related-but-distinct lineage.

Absorb-value

The transferable pattern is certificate-driven skill admission: storing mathematically-verified skill-safety boundaries (does this skill help globally? does it stay within an acceptable trade-off across motives?) as native MeTTa Atoms, so a skill library is auditable and composition-safe by construction. Other Hyperon teams can express the CDS/PDS admission gates as a MeTTa-native ruleset.

Maturity caveats

  • Implementation-backed on the load-bearing points (CDS/PDS gates, tested certificate admission, Hyperon certificate store, MO-LunarLander validation) — but not fully productized: the trained generator weights (generator.pt) are gitignored and must be trained before the demo runs (the pinned tree ships only a pilot PPO checkpoint).
  • Some runtime extensions are branch-only at the pin (Zero-Shot-Reuse validation, parts of the MDN runtime selector); main already carries the CDS/PDS gates, certificate storage, and the demo path.
  • Active multi-developer iCog research line (HEAD 2026-06-21).

Provenance

Extracted under the Cross-Org Cognitive Repo Sweeps program (SC-6 iCog-Labs, Source 2), 2026-06-23, three-model triangulation (Codex citation-grade + Gemini synthesis + orchestrator spot-verification of the Hyperon-coupling and CDS/PDS claims at pin d128ba91). Indexed from Reference+GitHub Repositories (SC-6 S2 group).

Verdict tags: [SUBREP-CERTIFIED-SKILL-DISCOVERY-HYPERON-CERTIFICATE-STORE] · [CDS-PDS-IMPLEMENTATION-BACKED] · [HYPERON-CERTIFICATE-STORE-WIRED] · [SUBREP-PAPER-ALIGNED-IMPLEMENTATION-NOT-AIRIS-DERIVED]

Hyperon component card: the SubRep concept/whitepaper card is view (SubRep) not supported for Cognitive Architecture & Research+PRIMUS (formerly CogPrime)+SubRep. This Deep Dive is its implementation-backed counterpart from the Cross-Org sweeps (iCog-Labs-Dev/subrep), bridging "proposed" (whitepaper) to "built" (code).