The campaign is a series of short, receipts-first papers, each independently citable, each sharing one ledger and one discipline. This file is the control surface: the layout, the manifest, and the rule for slotting a new release in.
Scope (carried on every number): proof-of-mechanism on a single dev host (RTX 2060, 12 GB). The 0.6B model (Qwen3-0.6B) carries the two-ring memory-ladder and control experiments (paper 01); the 12B (Gemma-3-12B, QAT 4-bit — the B1 artifact) carries the XBAR and KAIROS headline results (papers 06–09). Not scale-validated, not multi-model, not independently reproduced — say exactly that, everywhere.
shannon-prime/ public repo / site root
README.md SERIES landing: what SP is + the series index (live + upcoming) + the live paper's front-door receipt
SERIES.md this file — the manifest + release control
LEDGER.md master claims ledger (every paper's receipts, paper-tagged) — single source of truth
METHODOLOGY.md shared discipline (gate vocabulary, bit-exact-when-off, "no number without a command") — written once, every paper cites it
COMPANION-THEORY.md shared cathedral pointer
CONTRIBUTING.md CITATION.cff LICENSE
papers/
01-two-ring-memory/ { README.md paper.md receipts.md repro/ } <- the module shape
02-reducing-loader/ { README.md paper.md receipts.md repro/ }
03-frobenius-quant/ (staged later, same shape)
04-oracle-teacher/ (staged 2026-06-06 — front-door mapped; verification methodology)
05-probe-suite/ (staged 2026-06-06 — front-door mapped; the testing-methodology SET)
06-dp4a-bandwidth-ladder/ (staged 2026-06-06 — front-door mapped; computing on the packed codes)
07-auditable-latent-crossbar/ (written 2026-06-14 — paper.md complete; latent KV write, X-R1)
08-o1-kv-learned-router/ (written 2026-06-14 — paper.md + repro/ complete; O(1) KV + learned router, X-R2)
09-kairos-resident-daemon/ (written 2026-06-14 — paper.md complete for the mechanism; resident daemon + rewind; release held on the in-flight soak)
10-receipts-or-it-didnt-happen/ (written 2026-06-14 — paper.md complete; the methodology as the contribution)
12-memo-curator/ (written 2026-06-17 — paper.md + README complete; autonomous discrete recall, X-C2)
13-speculate-and-undo/ (written 2026-06-17 — paper.md + README complete; O(1) bit-exact rewind, X-222)
14-parameter-free-neocortex/ (written 2026-06-17 — paper.md + README complete; VSA/HRR Ring-3 consolidation, X-R3VSA)
15-the-organism-breathes/ (written 2026-06-17 — paper.md + README complete; real audio → episodic memory, X-ORG)
16-the-unification/ (written 2026-06-18 — paper.md + README complete; the bind re-carried onto exact-integer O_K, X-OK-BIND)
17-frobenius-integer-store/ (written 2026-06-18 — paper.md + README complete; integer episode codec + the boundary of the algebra, X-OK-FROB)
18-organism-on-silicon/ (written 2026-06-18 — paper.md + README complete; the full cross-modal loop on the integer substrate, X-OK-ORG)
19-killing-the-float-islands/ (written 2026-06-18 — paper.md + README complete; the four nonlinear fp32 islands → exact-integer on the 12B forward, X-BX-ISLANDS)
20-one-substrate-every-backend/ (written 2026-06-18 — paper.md + README complete; the universal-crate reference + CRT-NTT attention + the O(1) daemon decode, X-BX-WIRE)
21-byte-exact-not-compression/ (written 2026-06-18 — paper.md + README complete; the de-conflation, the boundary thesis, the re-derivation lesson, X-BX-BOUNDARY)
(Numbering: paper 11 is reserved; the orchestration-tier set above the closed P3 substrate is published as 12–15, matching the operator's release plan; the discrete-container set — the crossbar re-carried onto the exact-integer O_K substrate — is 16–18; the byte-exact set — the forward pass itself carried onto the exact-integer substrate — is 19–21.)
To slot a release in: drop its papers/NN-*/ folder (self-contained — paper, its receipts, its repro), flip its row in the manifest below to released, surface it on the landing index, tag the repo. Nothing else moves. That's the whole point of the module shape.
| # | Paper | Front-door receipt | Status | Target |
|---|---|---|---|---|
| 01 | Two-ring memory: query-directed recall + byte-addressable KV offload | needle off NVMe (512-position, poison-gated, 7.57 µs/read), 910× resident-cache shrink, 8× sparsification at +0.69% PPL | draft complete — R9 (32k) MISSed 2026-06-06 at the 64× budget; the 32k headline is WITHDRAWN from the front door (honest-negative row stays in the ledger); release on the 512-proven claims | release first |
| 02 | The reducing loader: output-preserving transcode + zero-copy swivel | a model transcoded to a smaller artifact that loads zero-copy and runs bit-faithfully | staged, repro green (6/6 E_FMT gates; 1,439→720 MB @50%, bit-faithful on gemma-3 + qwen3 — see EXPECTED.md) | +2–4 wks after 01 |
| 03 | Frobenius calibration-free quantization | fp8 without QAT, ~zero PPL delta (verify numbers first) | candidate | after 02, if it re-gates clean |
| 04 | The Oracle & the Teacher: oracle-grounded backend verification | 35-layer GPU port matched at KL 2.663e-10 (argmax 12/12), decode teacher-forced exact, first-try 38/38 — PLUS the case study: the hand-written oracle that measured gemma-4's TRUE PPL at 4.68 and convicted the GGUF ecosystem (192–506) while exonerating llama.cpp's forward (06-R8) | written (paper.md 2026-06-08; instruments live in lattice tests/gemma4_gold/) |
with 05/06 (the gemma-4 triptych) |
| 05 | The Probe Suite: bisection, isolation & benchmark hygiene as one set | the suite's kills (12.65× phantom, 2.8e-3, 0/256, ×25 norm, rank-205596 act-quant) PLUS the suite at ecosystem scale: tensor-class swap bisection, cosine forensics, simulate-before-build matching the built artifact to FOUR decimals (06-R9) | written (paper.md 2026-06-08; instruments live) |
with 04/06 |
| 06 | Computing on the Zip File: the dp4a bandwidth ladder | Gemma-4-12B at 26.1 tok/s AND wikitext PPL 5.12 on an RTX 2060 12GB (06-R10, 24/24 gates; llama.cpp: 31.29 tok/s at PPL 192–506; SP engine bandwidth +18%) + the isolated ladder (f32 1× → int8 ~3.8× → Q4 ~7.06×) + OK_Q4B + the sovereign supply chain | complete — gated, citable (paper.md done; 06-R6's 34.2 RETIRED with its quality-failed artifact) |
release-ready |
| — | GEMMA4-QUANT-FIX.md: community tutorial — verify the GGUF breakage + the fix recipe (post text: GEMMA4-ISSUE-POST.md) | gold 4.68 vs GGUFs 192–506; recipe ladder +45%→+9.6% | written | post alongside 06 |
| 07 | The Auditable Latent Crossbar: steering a frozen 12B through its KV cache, no tokens | 12B steered by direct KV-cache transplant: 15/15 incorporation, 15/15 selectivity (2×2 double dissociation), 3.69 orders max rank pull, self-transplant null 7/7 bit-identical, gold-instrument coherence (X-R1) | written (paper.md 2026-06-14; citable via X-R1) — architecture lattice RFC-XBAR-auditable-latent-crossbar.md |
standalone-repro re-gate (needs GPU; held while soak runs) before full release |
| 08 | O(1) KV: a context-decoupled cache via a learned router | learned 512×32 LSH router selects global top-B at +0.47% PPL @8× (oracle −0.08%, frozen +4.17%); compact slab → O(1) VRAM (8k↔16k flat ~50 MiB); NIAH needle survives 10/50/90%, frozen-router control MISSES (X-R2) | written (paper.md + repro/EXPECTED.md 2026-06-14; citable via X-R2) |
standalone-repro re-gate (needs GPU; held while soak runs) before full release |
| 09 | KAIROS: a resident 12B daemon that stays silent and rewinds at the metal | 24-tick crucible perfect (21/21 NO_OP, 3/3 ACTION, 0 false/0 missed/0 drift; 0.6B control collapses); O(1) byte-exact rewind (48 layers diffs=0; metal 0.0073 vs prefix-grow 0.924 s/action, 127×); wrap-aware journaled ring byte-exact (KAIROS-03) uniting O(1)-time with O(1)-space + the semantic metal loop (24 ticks 0-fault) (KAIROS-01/02/03) | written for the mechanism (paper.md 2026-06-14; citable via KAIROS-01/02) — full release HELD on the ≥24 h endurance soak, which is IN-FLIGHT (no verdict from a mid-run log) |
release on the soak receipt |
| 10 | Receipts or it didn't happen: bit-exact-or-bounded as the contribution | the four self-corrections (34.2 retired by its own PPL rule; 32k MISS kept on the front page; the small-N noise illusion; oracle-ceiling-before-training) | staged draft — evidence is the existing ledger + METHODOLOGY | with 07–09 |
| 12 | The Memo Curator: autonomous discrete recall above the crossbar | the loop that drives the closed crossbar on its own — inert when off (G-MEMO-NULL, PPL 4.6665 bit-identical), a 256-bit LSH / integer-Hamming address (reduction-order-immune; sign-binarize collapses at r=32, ship r=256), promote-matched / discard-corrupted (G-MEMO-LOOP, +0.000% / +40106%) | written, citable — X-C2 (paper.md + README 2026-06-17) |
with 13–15 |
| 13 | Speculate and Undo: O(1) bit-exact rewind of latent memory | replay into the resident cache is load-bearing (zeroed reads back all-zero) + rewind resets the prefix byte-identical (G-222, 12B+E2B; G-222-WRAP across an SWA-ring wrap) — the §4-trap guarantee made mechanical | written, citable — X-222 (paper.md + README 2026-06-17) |
with 12, 14, 15 |
| 14 | A parameter-free neocortex: VSA/HRR Ring-3 consolidation | the gist tier from discrete NTT, zero training; retrieve-and-verify (P2.b top-5 honored) — G-R3-BIND (recall@1=1.0 to N=32), G-R3-LOSS (step function: hit +0.000% / miss +8.04% caught), G-R3-DUALROUTE (decoy scan), G-R3-NIGHTSHIFT (349.8 MB → 16.3 KB) | written, citable — X-R3VSA (paper.md + README 2026-06-17) |
with 12, 13, 15 |
| 15 | The organism breathes: real audio to episodic memory | real speech → EAR on physical GNA 2.0 (0.877, KAIROS-04) → 12B pivots 7/8 → audio-conditioned KV serialized as a canonical Ring-2 episode; signature separates (self 211/256, margin +79); round-trip clean (the +1989% is foreign-by-design) | written, citable (step 1) — X-ORG (paper.md + README 2026-06-17) |
with 12–14 |
| 16 | The Unification: the latent crossbar re-carried onto the exact-integer O_K substrate | the Ring-3 bind re-carried onto the engine-native dual-prime negacyclic CRT-NTT: 256/256 bit-identical to native sp_pr_mul + reduction-order-immune (int byte-identical / float 4.44e-15); live loop on native sp_pr_mul (CAP=32 unregressed); Leg B χ_d carrier lowers coherence but is inert (the first negative) |
written, citable — X-OK-BIND (paper.md + README 2026-06-18) |
with 17, 18 |
| 17 | The Frobenius integer episode store: and the boundary of the algebra | rank-2 O_K codec a·s_a+b·s_b sub-ULP at 24 bits (a16b8 relL2 1.2e-7, 0.76× store); losslessness = reconstruction fidelity, not a fake +0.000% (the n=42 PPL gate is blind below ~1%); three boundary negatives (entropy 1.02×, Möbius sheds memories, T2-on-weights ≈ random) |
written, citable — X-OK-FROB (paper.md + README 2026-06-18) |
with 16, 18 |
| 18 | The organism breathes (full loop): a cross-modal continuous→discrete→continuous loop on silicon | the full audio-cue loop on the integer substrate — sig (audio 256 / decoys 147,129) → integer bind → retrieve top-1 (cos +0.47) → C2 cross-modal verify (accept audio / reject text) → Frobenius land (relL2 ~9e-8) → metal SP_REPLAY clean; + the period-8→6 rebase (separation cleaner, decoy 154→129) |
written, citable — X-OK-ORG (paper.md + README 2026-06-18) |
with 16, 17 |
| 19 | Killing the float islands: an exact-integer forward pass on a 12B | the four nonlinear fp32 islands (RMSNorm/softmax/GELU/RoPE) → deterministic exact-integer, no libm (RoPE by device CORDIC), no __int128 (M ≈ 2^60 fits u64); G-BYTEEXACT-FORWARD-12B OFF 4.6665 byte-identical to bf16 gold (null floor) / ON 4.6569 parity (−0.21% @ n=42) / ON run-to-run bit-identical |
written, citable — X-BX-ISLANDS (paper.md + README 2026-06-18; the 2-physical-GPU check is the open external step) |
with 20, 21 |
| 20 | One substrate, every backend: the universal crate, the CRT-NTT attention, an O(1) daemon decode | the L2 Rust crate is the bit-exact reference the C/CUDA/HVX backends gate to; attention ⟨q,k⟩/p·V → exact-integer dual-prime convolution (p·V ~2^46 ≪ M ⇒ no third prime); G-WIRE-CUDA-DECODE-GEMMA4 32/32 tokens bit-identical to the oracle, VRAM flat (O(1)) |
written, citable — X-BX-WIRE (paper.md + README 2026-06-18) |
with 19, 21 |
| 21 | Byte-exact, not compression: the boundary thesis + the honest negatives | the de-conflation (auditability ≠ compression); compression levers convicted (incoherence 1.37× / column reorder 1.05×, redundant vs OK_Q4B); the boundary thesis on the forward (T2-on-real-weights recon ≈ random); the re-derivation kept on the record |
written, citable — X-BX-BOUNDARY (paper.md + README 2026-06-18; the reflective record — no new performance number) |
with 19, 20 |
| — | multi-device CRT sharding · MTP rollback | — | not yet (no standalone receipt) | when gated |
| — | the algebraic framework (CM elliptic curve) | — | companion only, never a receipts paper | — |
- Lead with the strongest, ship it complete. 01 (memory) is the loudest receipt and it's drafted — it goes first, finished, before 02 starts its release clock. One complete paper beats six started.
- Stagger 2–4 weeks. Each release gets its own moment; cross-link back (to released) and forward (to upcoming) so the series compounds attention instead of splitting it.
- Every paper carries its own repro before it releases. 02's headline (a smaller artifact, bit-faithful) needs a one-command repro the way 01 has
run_r9. No release without it. - Re-gate before claiming. 02 and 03 rest on prior results (C1 transcode, Frobenius fp8) measured in earlier work; re-run their gates and build the repro before the paper goes public — same standard as 01, no grandfathering.
- The ledger is shared and append-only. New paper = new rows tagged to it; no number appears anywhere that isn't a ledger row with a command.
PAPER-DRAFT.md→papers/01-two-ring-memory/paper.mdCLAIMS-LEDGER.md→LEDGER.md(paper-tag the rows: R1–R9 =[01])repro/→papers/01-two-ring-memory/repro/repo-skeleton/README.md→ split: the series-level intro becomes rootREADME.md; paper-01 front-door becomespapers/01-two-ring-memory/README.md
- 22 — The honest-negatives wall (
X-B3-NEGATIVES): nine hand-designed recall signals, all refuted open-world. The negatives that justify the learned head. - 23 — Parametric steel & the teacher-forced ablation knockout (
X-B3-ABLATION): novel-needle causal ablation makes episodic dependency measurable (−33.56 vs −0.15); the oracle that is also the labeler. - 24 — The learned librarian (
X-B3-WC): a W_c head, diversity (34%→100%), logsumexp-mean (int16==f32), 360/361 recall + 50/50 reject, DEPLOYED LIVE on the 12B.
The KEYSTONE milestone (lattice papers/PPT-LAT-KEYSTONE.md): the pieces lock into one self-curating organism — the model owns its memory, acts through tools, and tidies its own state between turns. All default-off (byte-identical null floor); host-side, no frozen-ABI / .sp-model change.
- 25 — The memory agency (
X-AGENCY): the model forgets, supersedes, and merges its own facts. FORGET (SP_FORGET, token-overlap match → drop from live set + rewrite the persisted registry), DECIDE (SP_DECIDE, the model answersCHANGED=nto supersede, gated by the "cannot both be true at once" test), MERGE (the model returns one combined fact; the daemon drops both originals + captures the synthesis). Emergent self-curation — the verdict is the model's, not a dedup rule. Gates G-FORGET / G-DECIDE / G-MERGE. - 26 — Ephemeral tool calling (
X-HARNESS-TOOLS): tool calling + sandboxed Python on a text-only backend. The model emits<tool name="…">{json}</tool>in plain text;run_with_toolsparses + executes + feeds back (ReAct);ToolSpec.from_callablederives the schema from a Python signature. CosySim's runtime re-hosted on sp-daemon (lmstudio stripped). Honest negative: multi-line indented code in a JSON string is unreliable (one-liners clean; the model saw its own stderr and retried). Gates G-HARNESS-DAEMON-E2E (H1) / G-HARNESS-TOOLCALL-E2E (H2). - 27 — The agency loop (
X-AGENCY-LOOP): the model does memory work between turns, not just stop. An agency round shows the model its own memory + the curation tools and lets it decide what to forget/consolidate; the KAIROS scheduler fires it on a heartbeat, idle-gated so it never starves a live turn. The "auto rounds" become where the organism acts. Gates G-HARNESS-MEMTOOLS-E2E (H3) / G-HARNESS-AGENCY-E2E (H4).
The memory model, the sharp grounding lesson, and the integration that locks the whole series under one stone. All default-off (byte-identical null floor); host-side, no frozen-ABI / .sp-model change.
- 28 — Tiered memory: short, mid, and long on one content-addressed store (
X-CONVMEM): three memory tiers on the MEM-OKF store — SHORT = the live conversation (re-prefilled; a planted name survives 20 turns), MID = durable facts extracted into the recall registry, LONG = the whole conversation stored complete (full/) and summarized (sum/) under one sha256 address (gist by default viarecall_conversations, verbatim on demand viaread_conversation); three disclosure tiers LUT →sum/→full/; plus the init priming (system prompt + recallable capabilities corpus + non-parametric seed facts). Gates G-HARNESS-CONVMEM-E2E / G-HARNESS-CONSOLIDATE / G-HARNESS-HOOK-E2E. - 29 — Conversational faithfulness: the chat didn't restart, it confabulated (
X-FAITHFUL): a sharp lesson — the "restarts each turn" feeling was a misdiagnosis (the daemon carries the full thread; a planted name survives, recall on and off). The real failure was faithfulness — the model confabulated an earlier stated fact (said "dog" over the user's "octopus"), leaning on parametric priors over grounding. Fix: a default system prompt ("never substitute a stated fact") →FAITHFUL=True(engine88d924e). Meta-lesson banked: served-model misbehavior is almost always ours (template / decode / sampler / forward / prompt), not the weights — verify vs llama.cpp + our PPL first (the EOT companion case blamed the quant twice, wrongly;SP_EOT_BIAS, engine9e4b40f). The structural answer is the tiered memory of paper 28; the prompt is the patch. - 30 — KEYSTONE: the night the arches locked into one organism (
X-KEYSTONE) — THE CAPSTONE: the integration as the result. The isolated proofs (byte-exact forward 19–21, two-ring/integer memory 13–18, the learned librarian 24, the diffusion judge) lock into one self-supporting organism at milestone keystone-1 — faithful conversation + memory agency (store/forget/decide/merge) + tool calling + Python + the agency heartbeat + tiered conversation memory + self-priming, all on the byte-exactO_Ksubstrate, all default-off = null floor, all receipted, the loop closing in zero manual steps. ~90% built; open edges named (O(1) conversation KV via L1 §6b, the 2-physical-GPU byte-exact check, deeper faithfulness, native-C XBAR + T4 weights). The container is exact; the structure is learned. The foundation to build forward on.