Skip to content

ROADMAP.md

Latest commit

 

History

History
87 lines (59 loc) · 6.28 KB

ROADMAP.md

File metadata and controls

87 lines (59 loc) · 6.28 KB

OMC Roadmap

Current chapter: v0.6-fibtier-memory (shipped 2026-05-17). Next chapter: GPU Prometheus scaffold (in flight). The six-chapter symbolic-context arc (v0.3 → v0.6) has landed.

See CHANGELOG.md and GitHub Releases for the chapter-by-chapter history of how OMC got here. This file describes what's on the path going forward.

Post-v0.5 candidates (none committed yet)

v0.6 candidate A — fibtier-bounded memory

v0.5 ships substrate-keyed memory but the store grows unbounded. Long-running agents need pruning. Wire fibtier's tier-bounded eviction into MemoryStore:

  • Each namespace gets a tier-state file alongside the index
  • Stores cascade into higher tiers via the fibtier fold mechanism
  • Old entries get summarized/aggregated as they fold upward
  • Bounded total entries across all tiers (default ~4180 = Fib(18))

v0.6 candidate B — Prometheus rerank pass

The substrate-ranked predict candidates can be reranked by a learned probability overlay. Train a small Prometheus model on the corpus, score top-k candidates' next-token probabilities, blend with the substrate distance.

v0.6 candidate C — substrate-attention follow-ups

  • Substrate-modulated Q projection. Q hasn't been swapped yet; the V resample recipe (post-projection modulation) may generalize.
  • Substrate FF: dampen off-attractor activations in the feed-forward residual.
  • Substrate LayerNorm: substrate-distance-weighted variance computation.
  • Larger-scale validation: every substrate-attention claim was made at TinyShakespeare scale (1.1MB). Need to verify the stack holds at 10-100MB corpora.

Other deferred items

  • Stateful corpus API — omc_corpus_build returns a handle, omc_predict_from(handle, prefix, top_k) reuses it. Saves the corpus-rebuild cost on repeated queries.
  • Streaming queries — incremental updates as the prefix grows token-by-token.
  • Cross-corpus weighted blending — give different paths different priority in the ranking.
  • Conversation-aware predict — omc_predict(..., context_hash=H) where H references prior reasoning state, biasing the ranking by which fns the agent has already touched.

v0.7+ candidates

Substrate-attention follow-ups

  • Substrate-modulated Q projection. Q hasn't been swapped yet; the V resample recipe (post-projection modulation) may generalize.
  • Substrate FF: dampen off-attractor activations in the feed-forward residual.
  • Substrate LayerNorm: substrate-distance-weighted variance computation.
  • Larger-scale validation: every substrate-attention claim was made at TinyShakespeare scale (1.1MB). Need to verify the stack holds at 10-100MB corpora.

Beyond (rough)

Transformerless LLM

The substrate-attention components stack to −8.94% inside one block. The path forward is a top-to-bottom harmonic-only architecture trained competitively. Open: how to handle non-integer-coherent quantities at this scale (the substrate metric only applies to integer-valued quantities, per the rule derived from the HBit-gate falsification).

JIT path expansion

  • AVX-512 widening — blocked on array-processing OMC fns to fill the wider lanes.
  • JIT for float-returning harmonic primitives — returns_float dispatch flag mirroring returns_array_int.
  • JIT for dict ops — currently pure tree-walk for string-keyed data; the L1 array-of-hashed-int rewrite avoided this for hot paths.

Tooling polish

  • Improved formatter (--fmt) — preserve comments, configurable line width.
  • LSP improvements: completion (uses the v0.3 predict engine), hover with substrate signature.
  • VS Code extension: snippet library, inline hint UI for the heal pass.

Done (linked to chapter releases)

Chapter Key shipped items
v0.5-substrate-memory omc_memory_store/recall/list/stats + filesystem persistence + 10.61× LLM context-budget reduction measured on a 20-turn agent task
v0.4-substrate-context omc_compress_context / omc_decompress tools + format=codec thumbnails + directory ingest + measured 1.85×-2.81× LLM context-budget reduction
v0.3.1-symbolic-compression omc_predict gains format=hash/signature/full (3.8× compression default) + omc_fetch_by_hash for on-demand recovery
v0.3-symbolic-prediction omc_predict_files(paths, prefix, top_k) returns ranked provenance-tracked continuations from a content-addressed corpus
v0.2-ergonomics += / -= / *= / /= / %=, len/range/getenv/to_hex/parse_int, negative array indexing, did-you-mean, traced errors, 11 heal classes
v0.1-substrate-attention Substrate-K + S-MOD softmax + substrate-V resample → −8.94% val on TinyShakespeare
v0.0.6-prometheus Tape autograd, AdamW, Embedding, LayerNorm, multi-block transformer, first substrate-K wins
v0.0.5-codec-kernel-protocol Substrate codec, omc-kernel, omc-grep, OMC-PROTOCOL v1, substrate-aware tokenizer
v0.0.4-jit-and-dual-band LLVM JIT, dual-band SSE2 codegen, harmony-gated branch elision, array support
v0.0.3-substrate-and-stdlib Heal pass, substrate-routed search family, stdlib expansion, --check / --fmt
v0.0.2-language-core Parser, two-engine interpreter, HInt, bytecode VM, self-hosting fixpoint
V0.0.1 Genesis: circuit evolution engine, FFI, Unity/Unreal bindings

ROADMAP.json is preserved for archaeology — it captured the state through v0.0.4. This file supersedes it as the canonical forward plan.