Skip to content

PPT-LAT-Roadmap.md

Latest commit

 

History

History
121 lines (92 loc) · 13.9 KB

PPT-LAT-Roadmap.md

File metadata and controls

121 lines (92 loc) · 13.9 KB
Raw
type roadmap
title PPT-LAT-Roadmap — the forward roadmap (post-KEYSTONE)
description The clean, forward-facing roadmap for Shannon-Prime after the KEYSTONE milestone (2026-06-25): KEYSTONE is achieved (~90% of the envisioned organism), the DONE pillars are listed, and the four open edges are the next work. The full historical roadmap is archived; PPT-LAT-KEYSTONE.md is the current-state source of truth.
tags
roadmap
keystone
forward
navigation
okf
timestamp 2026-06-25 00:00:00 UTC
resource shannon-prime-lattice/papers/PPT-LAT-Roadmap.md
sp_status GREEN-LIVE
sp_gate KEYSTONE-1
sp_commit keystone-1
sp_repro see PPT-LAT-KEYSTONE.md §10 (Run it) + §9 (Gate index)

PPT-LAT-Roadmap (post-KEYSTONE)

Project: shannon-prime-lattice · Authors: Knack + Claude + Gemini · License: MIT (all repos). Public front door: Position Is Arithmetic · live site.

Read order. The current-state source of truth is PPT-LAT-KEYSTONE.md (the map) + PPT-LAT-KEYSTONE-API.md (the call surface). This roadmap is the short forward-plan that sits on top of them. The pre-KEYSTONE roadmap (the full ~8,500-line phase history) is archived at Archived/PPT-LAT-Roadmap-pre-keystone.md — read it only for per-phase gate definitions and historical context, never as current state.

0. Status — KEYSTONE is achieved (~90% of the organism)

KEYSTONE (keystone-1, 2026-06-25) is the night the arches locked together: the pieces that were proven in isolation — the byte-exact O_K forward, the two-ring/XBAR memory, the learned librarian, the diffusion judge, the agent harness — were integrated into one self-supporting organism. The served Gemma-4-12B chat now holds the conversation faithfully, learns/recalls/forgets/supersedes/ merges facts on its own judgement, calls tools and runs Python, manages its own memory on a heartbeat, and stores conversations in tiers — and the loop closes with zero manual steps.

Every mechanism is a flag that is a strict no-op when unset (the null floor); every number has a reproducing command and a gate. This roadmap supersedes the "current state" of the archived one.

1. DONE — the pillars that are built, gated, and (where applicable) live

Pillar What it is Gate(s) State
Byte-exact O_K forward The whole Gemma-4-12B forward on exact-integer O_K = Z[(1+√-163)/2] / dual-prime CRT-NTT; the 4 nonlinear islands (RMSNorm/softmax/GELU/RoPE) as exact-integer (CORDIC RoPE, no libm). Byte-exact = exact arithmetic / cross-machine determinism / AUDITABILITY, NOT compression. G-BYTEEXACT-FORWARD-12B (off=4.6665 byte-identical null floor / on=parity / run-to-run bit-identical) GREEN
Coherent served chat Real 12B chat through the full stack on a single latent entry point — coherent, reproducible, byte-exact, O(1)-context. G-CHAT-FULLSTACK GREEN-LIVE
Autonomous recall + reject The learned W_c librarian does instance-level episodic recall with clean foreign-reject, live on the served chat. G-CHAT-B3-WC-DEPLOY (matched→RECALL / foreign→NULL→clean) GREEN-LIVE
Memory agency — FORGET / DECIDE / MERGE The model owns its memory: forget on intent, supersede a changed fact, consolidate two complementary facts into one synthesized truth. Default-off = null floor. G-FORGET, G-DECIDE, G-MERGE GREEN-LIVE
Tool-calling harness (H1–H7) CosySim runtime re-hosted on sp-daemon: ephemeral text-protocol tool calling + Python exec, memory-as-tools, the agency loop, the KAIROS heartbeat tick, tiered conversation memory, the live consolidation hook. G-HARNESS-{DAEMON,TOOLCALL,MEMTOOLS,AGENCY,KAIROS-TICK,CONVMEM,HOOK}-E2E GREEN
Tiered conversation memory SHORT (live convo, re-prefilled + faithfulness prompt) → MID (extracted facts, registry + ep.k) → LONG (full+summary MEM-OKF), one sha256/C2-sig scheme linking the tiers. G-HARNESS-CONVMEM-E2E GREEN
Live consolidation loop The daemon writes each turn to disk; the KAIROS scheduler consolidates it + runs a maintenance round on its tick — zero manual steps. G-HARNESS-HOOK-E2E (H7) GREEN
Deterministic recall/reject judge The production recall gate is a deterministic token-overlap (Jaccard) verifier @~0.6, not the 26B (83%/95% on a CPU string op; the 26B cascade retired). G-JUDGE-BATTERY GREEN
Latent Interceptor (hardened heads) The finetuned draft body as a latent-native router: shared 1024-d body + tiny action/memory/tool heads, near-miss-hardened so they never fire on idle chatter (false-fire 0.000 on isolated cross-dist OOD; KEEP recall lifted 0.429→1.000). G-TH-HARD (tool OOD 1.000), G-ACT-HARD (action OOD 0.979) GREEN
Telepathy — cross-family LatentBridge Tokenizer-free latent→latent transfer between model families via a ridge affine adapter + adapter registry: gemma-3n-E2B ↔ qwen2.5-coder-0.5b — alignment + foreign-reject + generation steering. The cemented architecture is two-stage decide_route(latent) → delegate_execute(clean text) (latent fusion degrades — honest negative). TELE-14: standalone SOVEREIGN native delegate — the coder runs fully in-engine (SP_TELEPATHY_NATIVE, L1 prefill_chunk/decode_step, zero Python); CPU L1 ⇒ free live co-residency with the 12B, no g_w refactor. Honest scope: gist/intent steering, NOT verbatim symbolic forcing. Licensing (SPEC): proprietary component on the MIT substrate, fail-closed license-key + attestation, no host-external effects. PARKED. spec G-TELEPATHY-ROUNDTRIP (retr@1 1.000/rt 0.891), -REJECT (AUC 0.999), -GEN-TRIGGER (steer-acc 1.000), -TWOSTAGE, -NATIVE (engine 2f57520) GREEN (repr+steer+native delegate)
Persistent O(1) conversation KV A follow-up turn appends to the resident 12B cache instead of re-prefilling the whole conversation (SP_PERSIST_KV, default-ON): longest-common-prefix reuse of the committed sequence + suffix-only prefill + bounded-tail rewind. Excludes cache-mutating paths (replay/inject/agency-writers/speculative-recall); =0 forces the O(n) null floor. G-PERSIST-KV (engine d211fd2): 6-turn byte-identical on==off; TTFT off 7.47× growth vs on flat (~1s); default-run == baseline GREEN-LIVE

| L5-cosine recall + attribute-grounding (faithfulness closed) | The recall selector is the layer-5 fact signal (SP_RECALL_L5 τ=0.30, 86.89% paraphrase). Faithfulness on zero-prior/private data = a deterministic attribute-gate + query-token guard with a ZERO-INFERENCE symbolic decline (no gemma4 forward on reject). The generative judge is PARKED (earned nothing vs τ + native robustness). Law: ADR-002; levers: FINDINGS-LEDGER. | G-L5-RECALL-LIVE, G-SNE-ATTRGATE-ZEROINF, G-HARDFOREIGN-* | GREEN(-LIVE) | | SP-SWARM private memory mesh (L0–L4) | Replication + discovery over the content-addressed MEM-OKF store (not a new store): L1 addressing (sha256+C2 classes), L2 have/want replication with verify-on-arrival, L3 Ed25519 sign/verify-vs-roster provenance, L0 QUIC (quinn/rustls) + Ed25519 mutual roster handshake, L4 C2-SimHash discovery gossip (SIM shortlist→exact-fetch). sp_swarm Rust crate, Rust↔Python byte-parity, integrated into sp-daemon (default-off swarm feature) + sp-swarm-node bin. C2-256 is a shortlist (recall@5 0.885), not top-1 (0.607) — honest-negative kept. Remaining = multi-host deploy. Call surface: PPT-LAT-MESH-API.md. | G-SWARM-{REPLICATE-CONVERGE, PROVENANCE-ED25519, RUST-PARITY, TRANSPORT-QUIC, NODE, DAEMON-WIRE, C2-INDEX, C2-SEMANTIC, GOSSIP-DISCOVERY} | GREEN (default-off) |

Foundation context (the substrate these ride on): the two-ring ARM KV memory, the XBAR auditable latent crossbar (C2 256-bit sigs, native integer Ring-3 VSA bind, Frobenius Ring-2 store), NIGHTSHIFT the offline curator, KAIROS the resident kernel, the GNA "EAR" audio sense. All closed; see PPT-LAT-KEYSTONE.md §4 and the archived roadmap for the full closure record.

Boundary thesis (kept): O_K wins on exact arithmetic (the container); every structure-on-content compression lever is measured-inert and kept as an honest negative (Dirichlet carriers, Möbius-on-M, entropy-on-codes, T2/T4-on-weights). The win is the container, not the content.

2. NEXT — the four open edges (from PPT-LAT-KEYSTONE.md §12)

These are the honest open edges, in no forced order — pick by the day's leverage.

PRIMARY (re-elevated 2026-07-01): SP-SWARM / DHT memory mesh — NOW BUILT (L0–L4 GREEN, 2026-07-01). Private, content-addressed, signed replication of MEM-OKF across the operator's nodes, alongside the ADR-002 Decide→Execute spine. The full stack is built and gated: L1 addressing, L2 have/want replication, L3 Ed25519 provenance, L0 QUIC transport, L4 C2-SimHash discovery gossip — 9 gates GREEN, Rust↔Python byte-parity, integrated into sp-daemon (default-off swarm feature) + a standalone sp-swarm-node bin (see the DONE pillar below). It rides the proven byte-exact content addressing (G-BYTEEXACT-FORWARD-12B) + the receipt ledger; it is NOT the rejected C2-Kademlia/poison-pill/field-crypto proposal (banked negative). Remaining = multi-host deployment only. Blueprint (why + rejected mechanics): PPT-LAT-DESIGN-SWARM-MEMORY-MESH.md; call surface (how: crate/flags/wire/gates): PPT-LAT-MESH-API.md.

  1. Persistent O(1) conversation KV — P1 DONE (now a DONE pillar above), P2/P3 open. The append is live and default-ON: G-PERSIST-KV GREEN (engine d211fd2) proved 6-turn byte-identity (on==off) with TTFT off 7.47× growth vs on flat, and the kill-test held (persisted KV stays byte-identical to a fresh re-prefill). Remaining: P2 = larger Pmax + ring tuning for ~40K-token sessions (the SWA ring is already on at W=2048); P3 = evict the 8 GLOBAL layers (the O(n) floor) to the existing XBAR Ring-2/Optane demotion tier for unbounded context. Scope: PPT-LAT-OKV-Persistent-KV-SCOPE.md.

  2. The 2-physical-GPU byte-exact check. The one remaining EXTERNAL item for byte-exact: a bit-identical logit check across two physical GPUs (needs a 2nd machine). On-machine we have run-to-run bit-identity + integer-reduction order-immunity as the proxy; the cross-card check closes it.

  3. Deeper faithfulness via reliable tiered recall. The model still leans on parametric priors over in-context grounding; the system prompt is the patch, reliable tiered recall is the structural answer — pull conversation facts into MID/LONG so they survive window-scroll/restart, and make recall dependable enough that the model trusts stated facts over its priors. Includes a fuller MEM-OKF capabilities corpus. Kill: extracted facts don't survive a restart, or recall stays unreliable.

  4. Native-C XBAR port + T4 Frobenius of the model weights. Port the host-Python XBAR/VSA tooling into the native core/ tree (the resident loop no longer needs Python — partially done via core/ring3/), and execute T4 Frobenius π^k of the 9.4 GB model weights — the validated, untouched compression lever (distinct from the convicted structure-on-content levers). Kill: the native port regresses bit-identity, or T4 fails to net-reduce at bit-exact.

3. Standing disciplines (binding)

  • Receipts-first. No number without a reproducing command + a LEDGER.md row, with scope attached. Bit-exact-when-off. No silent gate revision — surface upstream. Honest negatives stay attached.
  • OKFS / SP-OKF + MEM-OKF anti-rebuild pre-flight. Before building ANY subsystem: python tools/okf_mem.py lookup --root memory-okf <kw> + grep the tree. A new file for an existing capability is a defect (the 20×-repeated rebuild failure). All knowledge docs carry the SP-OKF frontmatter; run python tools/okf_validate.py papers on a touched bundle before commit. Specs: SP-OKF-PROFILE.md, MEMORY-OKF-PROFILE.md.
  • Check the code + commits + git fetch before trusting memory. STATE (PPT-LAT-STATE.md) is the proven record. Git ops on these repos via native PowerShell, not the Linux mount (the mount CRLF-churns + locks).
  • Recurring lesson, banked: served-model misbehavior is almost always ours (template / decode / sampler / forward / prompt), not the weights — verify vs llama.cpp + our PPL first. For meta-cognitive model-calls: frame as detection, not decision, and force the answer prefix.

4. Where current truth lives — supersession order

PPT-LAT-KEYSTONE.md (current-state map) > PPT-LAT-STATE.md (proven ledger) > active contract run records > this roadmap > Archived/PPT-LAT-Roadmap-pre-keystone.md (historical).

Need Go to
Current-state map / API PPT-LAT-KEYSTONE.md, PPT-LAT-KEYSTONE-API.md
Proven state record PPT-LAT-STATE.md
Bootstrap / methodology / operator prompt.md, CLAUDE.md
Architecture RFC PPT-LAT-RFC-001-Universal-Discrete-Architecture.md (see the KEYSTONE addendum)
Public papers / claims ledger Position_Is_Arithmetic SERIES.md, papers/, LEDGER.md
Historical roadmap (full phase history) Archived/PPT-LAT-Roadmap-pre-keystone.md

Post-KEYSTONE roadmap, 2026-06-25. KEYSTONE achieved (~90% of the organism); four open edges remain. Receipts-first; honest negatives attached; default-off is the null floor.