docs: de-stale model card to live state (30L dual-stream, ep~6205, EP-AVG ATL 6.4339 @ ep6132, chip ATL 1.2637, 128ctx, 18 surgeries + cord)

README.md

@@ -33,8 +33,8 @@ pipeline_tag: text-generation
 **Maintainer:** RFI-IRFOS, contact@ternlang.com  
 **Repository:** https://github.com/rfi-irfos/ternary-intelligence-stack  
 **License:** LGPL-3.0-or-later (model weights, training code, inference runtime). Platform infrastructure (API server, MCP tooling, HDL) is BSL-1.1. See [README §Licensing](README.md#licensing) for the full tier breakdown.  
-**Last updated:** 2026-05-27  
-**Training status:** Paused (Modal billing ceiling, ep4234) — **26L dual-stream** · 13 depth surgeries + 1 cord surgery complete. Cord surgery fired autonomously ep4202, 2026-05-27T16:44Z — first documented single-to-dual-stream bifurcation mid-training. S13 (25L→26L) fired ep~4207. Chip ATL **8.6852** (post-S13). EP_AVG ATL **9.2847** (ep3456, 20L). fib_index=7 · window=34 · Gen3 step1/6. Resuming on Modal T4 once billing settled.
+**Last updated:** 2026-06-12  
+**Training status:** Active (ep~6205) on Modal T4 — **30L dual-stream** · 18 depth surgeries + 1 cord surgery complete. Cord surgery fired autonomously ep4202, 2026-05-27T16:44Z — first documented single-to-dual-stream bifurcation mid-training. S17 (29L→30L) fired ep5610, 2026-06-06 (checkpoint-mtime verified). Context was reduced 256→128 — memory-forced: a depth surgery's activation memory exceeded the L4's VRAM; restoring 256 (via 2× L4 / tighter memory management) is planned. Best EP-AVG ATL **6.4339** (ep6132, 30L). Chip ATL **1.2637** (best single intra-batch loss). fib_index=7 · window=34 · Gen3 step1/6. (Resumed 2026-06-12 after a ~1-week Anthropic billing-migration gap.)
 
 ---
 
@@ -50,19 +50,19 @@ weights, targeting inference on edge hardware and low-power devices.
 |----------|-------|
 | Architecture | **Dual-stream** Ternary MoE (Mixture of Experts) |
 | Streams | **2** (bifurcated via cord surgery ep4202, 2026-05-27) |
-| Layers | **26** per stream |
+| Layers | **30** per stream |
 | Hidden size | **2×256H** (256H per stream) |
 | Anastomosis gates | **6** — bidirectional F32 cross-stream fusion at Fibonacci layers [2,3,5,8,13,21] |
 | Experts | 12 per stream (Top-3 routing; shared FFN weights, independent routing gates) |
-| Context length | 256 tokens |
+| Context length | 128 tokens (reduced from 256, memory-forced; 256 restore planned) |
 | Vocabulary | 32,000 tokens (ByteLevel BPE — EN/DE/FR/ES/PT/IT/NL/PL) |
 | Weight representation | Ternary {-1, 0, +1} with STE training |
 | Gate linear | F32 |
 | Positional encoding | RoPE (rotate_half) |
 | Optimizer | AdamW, cosine LR decay, BATCH=1 (post-cord) |
-| Parameters (total) | **~194.4M** |
-| Safetensors | **2,044 tensors · 741.4 MB** |
-| Surgeries | **13 depth (S1–S13)** + **1 cord surgery** = 14 total surgical events |
+| Parameters (total) | **~221.7M** |
+| Safetensors | **2,356 tensors · ~845 MB** |
+| Surgeries | **18 depth (S1–S17)** + **1 cord surgery** = 19 total surgical events |
 
 The central technical innovation is the **@sparseskip** primitive — a
 learned sparse-skip layer that dynamically bypasses computation paths
@@ -157,6 +157,13 @@ noisy inputs.
 | **Ep4202** | **CORD surgery** | **25L → 2×25L dual-stream** | — | **2026-05-27T16:44Z — first ever autonomous single→dual-stream bifurcation** |
 | Ep~4203 | 9.3241 | ← first post-cord epoch avg | **8.7123** (chip, new ATL) | Dual-stream live |
 | Ep~4207 | S13 fired | 25L→26L surgery (both streams) | **8.6852** (chip, new ATL) | 2026-05-27T17:40Z; fib_index 6→7 |
+| ~ep4280 | S14 fired | 26L→27L (both streams) | — | 2026-05-29 |
+| ~ep4350 | S15 fired | 27L→28L (both streams) | — | 2026-05-29 |
+| ~ep4453 | backbone-trains cliff | EP-AVG **9.38 → 8.43** in <100 ep | — | DiffLayerNorm fix unfroze the transformer body — largest structural drop in the run |
+| ~ep4740 | S16 fired | 28L→29L (both streams) | — | 2026-05-31 (✓ checkpoint-mtime) |
+| ep5610 | S17 fired | 29L→30L (both streams) | — | 2026-06-06 21:08 (✓ checkpoint-mtime) |
+| ep6132 | EP-AVG ATL **6.4339** | 30L — all-time best | **1.2637** (chip) | best EP-AVG aggregate; chip = best single intra-batch loss |
+| ~ep6205 | training active | 30L dual-stream, live | — | 128CTX · BATCH=1 · Modal T4 · resumed 2026-06-12 |
 
 The benchmark suite runs 5 fixed prompts covering English, German,
 multilingual, narrative, and technical domains. Results are reproducible
@@ -176,10 +183,10 @@ The gate fires when loss plateaus below 9.8 for a 144-epoch window with `myc_sta
 
 **Known limitations:**
 
-- At current training depth (~1459 epochs), output quality is pre-fluency:
-  the model produces partially coherent text in familiar domains but lacks
-  consistent grammatical structure across longer sequences.
-- Context window of 256 tokens is shorter than contemporary LLMs; cannot
+- At current training depth (~ep6205, 30L dual-stream), output quality is
+  pre-fluency: the model produces partially coherent text in familiar domains
+  but lacks consistent grammatical structure across longer sequences.
+- Context window of 128 tokens is shorter than contemporary LLMs; cannot
   maintain coherence over longer passages.
 - Ternary quantization trades weight precision for size — at this scale,
   some representational capacity is lost relative to F32 equivalents.
@@ -225,8 +232,8 @@ oversight mechanisms are in place:
    by the RFI-IRFOS team.
 2. **Surgery governor:** Architectural growth (layer addition via net2net)
    is fully autonomous — the `EvolutionManager` fires on a Fibonacci-gated
-   plateau detector with no human intervention required. **13 depth surgeries
-   (12L→26L) + 1 cord surgery (single→dual-stream bifurcation)** have been
+   plateau detector with no human intervention required. **18 depth surgeries
+   (12L→30L) + 1 cord surgery (single→dual-stream bifurcation)** have been
    executed autonomously to date. The cord surgery (ep4202, 2026-05-27) is the
    first documented autonomous single-to-dual-stream bifurcation in a live
    ternary MoE.