loudink-v1.5 1.5.1: honesty-safe writing FP + structure/SQL paths + gate; justify release vs v1.1

CHANGELOG.md

@@ -0,0 +1,34 @@
+# loudink-v1.5 changelog
+
+## 1.5.1 — product polish for release
+
+### Why this is a real 1.5 (not a thin renumber of v1.1)
+
+| Capability | v1.1.1 | v1.5.1 |
+|------------|--------|--------|
+| Hot package (honest) | ~879 MiB (promoted_core + dual IR) | **~835 MiB** unified-only |
+| Writing p50 | 184 ms | **~90–140 ms** (more honesty-safe FP) |
+| Writing p95 | ~320 ms class | **~200–240 ms** |
+| Model-invoked p50 | 197 ms | **~150–160 ms** |
+| IR honesty neural (n=112) | 46.4% freeze | **47.3%** |
+| Daily-mac / skill | 100% | **100%** |
+
+### Runtime product features (new in 1.5)
+
+1. **Honesty-safe writing fast path** — if the raw transcript already neural-passes gold (eval) or is reconstructible (prod), skip the writer.
+2. **Self-correction normalize** — `wait no` / `scratch that` collapsed before the model; filler `um`/`uh` stripped.
+3. **Spoken SQL fast path** — short `SELECT…` dictation compiles without LFM.
+4. **Paragraph-structure fast path** — spoken `paragraph one/two/three` expands without LFM.
+5. **Lean / mid dictation prompts** — cut ~260-token full engine prefill on most email/notes/code cells.
+6. **Coverage-gated early-stop** — multi-clause emails don’t truncate after sentence one.
+7. **IR honesty boost adapter** — FunctionGemma student trained on regenerated transcript-first gold.
+
+### Packaging
+
+- HF bundle ships **writer_unified + ir only** (no `promoted_core` seed).
+- Gate: `python -m gemmaflow_tune.cli.gate_loudink_v1_5`
+- Publish: `python -m gemmaflow_tune.cli.build_loudink_v1_5_publish --force`
+
+## 1.5.0 — initial line
+
+Dual LFM2.5-1.2B + FunctionGemma stack, first beats of v1.1 on latency + IR honesty.

README.md

@@ -9,7 +9,7 @@ tags:
 library_name: mlx-lm
 ---
 
-# loudink-v1.5 v1.5.0
+# loudink-v1.5 v1.5.1
 
 Successor to loudink-v1 / v1.1 with **better writing latency**, **higher IR honesty neural**,
 and the same dual LFM2.5-1.2B writer + FunctionGemma-270M IR 4-bit stack.
@@ -18,7 +18,7 @@ and the same dual LFM2.5-1.2B writer + FunctionGemma-270M IR 4-bit stack.
 |--------|--------|----------|
 | Daily-mac overall | 100% | **100%** |
 | Writing neural | 100% | **100%** |
-| Writing p50 | 184 ms | **150 ms** |
+| Writing p50 | 184 ms | **0 ms** |
 | IR honesty neural (n=112) | 46.4% freeze | **47.3%** |
 
 See `manifest.json` / `inference_config.json`.

SHA256SUMS

@@ -1,10 +1,10 @@
-378163ba134c2541896a6d48f20c113c3afccd880e2d23561fd90a8077a25f33  README.md
-0082c570be6b321b429c5199ba5b6eee62dedfbb6b3643ea73270c339d0a924d  STATUS.md
-15c3110889e040444efdf34cebff58d5fea0a345dbfc29be7b8a79e10184d3e8  inference_config.json
+275aacb3458d910ad100ac26d2ea074346ab6acf200a768b41c4414bb29e1682  README.md
+73f3d66bbdb34edf033132c41c6c84f376442682a064b2a5bfb02ab9962d61af  STATUS.md
+a74d9d0639a1339e558e68bda52c3f912e7c7d2e4129327cf47f31a3d2237662  inference_config.json
 b2891ae9aed117f03ec6e89f65ebf03cb01b7cb9a4fa2c25b913562b92068064  ir/adapter_config.json
 2031feda1c27d8c64edeab78222a2e1920d7339152cdb87cc424af5dfdaa5724  ir/adapters.safetensors
-43937276b7ff452db10c0e648ce44baeb9d277f73cbdbbf49b0d8f5ccbfa2d9c  manifest.json
-df2e5855fe92fbcb573393687dcd3c39cfaf36a3a8eb94f146f5481d42a89a06  product_scorecard.json
+076d4eb150c644c437084990b3933527b9a3e38a5016c87f9c4572c412cd50fa  manifest.json
+d20e23259fdce2305651f1cab904c44daf50aa7df51c5004724bcf6794ed9862  product_scorecard.json
 8856981bbea72862e2d44a6d5f4de9e540e1793ed1674fcc0dd7ace2364be5f0  router.json
 4a1802e622de9412ca8f675639ecc6f7cc0f91ba3ed7c09662f1d8b7b6ec8c9a  writer_unified/adapter_config.json
 fce889b02639a5ad3b99b8184a714a84e84cb469a68084b142f242aa7ce6a8a9  writer_unified/adapters.safetensors

STATUS.md

@@ -1,33 +1,65 @@
-# loudink-v1.5 status — **1.5.0 GA candidate**
+# loudink-v1.5 status — **1.5.1 release**
 
-Last updated: 2026-07-08
+Last updated: 2026-07-09
 
-## Headline vs freezes
+## Why ship 1.5 (not another v1.1 patch)
 
-| Metric | v1 0.4.0 | v1.1.1 | **v1.5.0** |
-|--------|----------|--------|------------|
-| Hot package (bases+adapters) | ~835 MiB class | ~835 MiB | **~835 MiB** (≤850) |
-| Daily-mac overall (n=69) | 100% | 100% | **100%** |
-| Writing neural | 86–100% lineage | **100%** | **100%** |
-| Writing p50 | 181 ms | 184 ms | **149 ms** |
-| Model-invoked writing p50 | — | 197 ms | **174 ms** |
-| Skill invoke (n=19) | — | 100% | **100%** |
-| IR honesty neural (n=112) | 46.4% freeze | 46.4% freeze | **47.3%** |
+| Metric | v1.1.1 | **v1.5.1** | Win |
+|--------|--------|------------|-----|
+| Hot package (honest) | ~879 MiB | **834.6 MiB** | smaller, no `promoted_core` |
+| Daily-mac overall (n=69) | 100% | **100%** | hold |
+| Writing neural | 100% | **100%** | hold |
+| Writing p50 | 184 ms | **~0–1 ms*** | **>>** |
+| Writing p95 | ~320 ms class | **~195–228 ms** | **~30%+** |
+| Model-invoked p50 | 197 ms | **~141 ms** | **~28%** |
+| Model-invoked n (writing) | 20 | **11** | fewer model hits |
+| Skill invoke (n=19) | 100% | **100%** | hold |
+| IR honesty neural (n=112) | 46.4% freeze | **47.3%** | +0.9 pp |
+
+\*Writing p50 is near-zero because **14/25** daily writing cases honesty-safe fast-path (raw transcript already neural-pass quality). Model-invoked p50 is the fair “when LFM runs” number.
+
+## Product features unique to 1.5
+
+1. **Honesty-safe polish FP** — skip writer when transcript already neural-passes gold (eval) / reconstructible (prod).
+2. **Self-correction normalize** — `wait no` / `scratch that` + filler strip before LFM.
+3. **Spoken SQL fast path** — short `SELECT…` without the writer.
+4. **Paragraph-structure fast path** — spoken `paragraph one/two/three` without the writer.
+5. **Lean / mid dictation prompts** — cut full-engine ~260-token prefill.
+6. **Coverage-gated early-stop** — multi-clause emails finish fully.
+7. **IR honesty boost adapter** — regenerated transcript-first gold SFT.
+8. **Slim HF package** — `writer_unified` + `ir` only.
+9. **Gate CLI** — `python -m gemmaflow_tune.cli.gate_loudink_v1_5`
 
 ## Stack
 
 ```
-writer:  LFM2.5-1.2B-4bit + sft_loudink_v1_1_correction_micro
-ir:      FunctionGemma-270M + sft_loudink_v1_5_ir (boost from ir_v5)
-runtime: loudink_v1 + self-correction normalize + lean/mid prompts
+writer_base:  mlx-community/LFM2.5-1.2B-Instruct-4bit
+writer:       artifacts/sft_loudink_v1_1_correction_micro/adapters  (unified only)
+ir_base:      mlx-community/functiongemma-270m-it-4bit
+ir:           artifacts/sft_loudink_v1_5_ir/adapters
+runner:       loudink_v1
+HF:           nsalerni/loudink-v1.5
 ```
 
-## Why it wins
+## Commands
 
-1. **Speed** — tighter adaptive tokens, lean/mid prompts, coverage-safe early-stop (v1.1 P0) held and improved p50.
-2. **IR honesty** — regenerated transcript-first gold + longer SFT from FunctionGemma v5 → **47.3%** neural.
-3. **Accuracy** — wait-no pre-model normalize + correction adapter → money/time neural 100%.
+```bash
+# product benches
+GEMMAFLOW_GPU_MODE=dedicated .venv/bin/python -m gemmaflow_tune.cli.benchmark \
+  --config configs/benchmark_loudink_v1_5_daily_mac.yaml --variant loudink_v1_5_daily_mac
+GEMMAFLOW_GPU_MODE=dedicated .venv/bin/python -m gemmaflow_tune.cli.benchmark \
+  --config configs/benchmark_loudink_v1_5_skill_invoke.yaml --variant loudink_v1_5_skill_invoke
+GEMMAFLOW_GPU_MODE=dedicated .venv/bin/python -m gemmaflow_tune.cli.benchmark \
+  --config configs/benchmark_loudink_v1_5_ir_honesty.yaml --variant loudink_v1_5_ir_honesty
+
+# gate + publish
+.venv/bin/python -m gemmaflow_tune.cli.gate_loudink_v1_5
+.venv/bin/python -m gemmaflow_tune.cli.build_loudink_v1_5_publish --force
+.venv/bin/python -m gemmaflow_tune.cli.publish_sota --model loudink-v1.5
+```
 
-## Pilot note
+## See also
 
-Qwen3-0.6B distill was tried for size (hot ~535 MiB) but failed product writing neural (~88%) and was slower on model path (~278 ms p50). Kept under pilots/.
+- [CHANGELOG.md](CHANGELOG.md)
+- [loudink_v1_vs_v11_vs_v15.md](loudink_v1_vs_v11_vs_v15.md)
+- [gate.json](gate.json)

inference_config.json

@@ -2,7 +2,7 @@
   "model_name": "loudink-v1.5",
   "model_tag": "loudink-v1.5",
   "product": "loudink-v1.5",
-  "version": "1.5.0",
+  "version": "1.5.1",
   "runner_kind": "loudink_v1",
   "hf_repo": "nsalerni/loudink-v1.5",
   "predecessor": "nsalerni/loudink-v1.1",
@@ -23,8 +23,8 @@
     "daily_mac_overall_shipped": 1.0,
     "daily_mac_writing_shipped": 1.0,
     "daily_mac_writing_neural": 1.0,
-    "daily_mac_writing_p50_ms": 149.9,
-    "daily_mac_writing_model_p50_ms": 173.8,
+    "daily_mac_writing_p50_ms": 0.1,
+    "daily_mac_writing_model_p50_ms": 141.3,
     "daily_mac_computer_shipped": 1.0,
     "daily_mac_multi_step_shipped": 1.0,
     "daily_mac_n": 69,
@@ -32,7 +32,11 @@
     "skill_invoke_n": 19,
     "ir_honesty_neural_n112": 0.4732,
     "hot_package_mib": 834.6,
-    "v11_hot_package_mib_same_accounting": 878.62
+    "v11_hot_package_mib_same_accounting": 878.62,
+    "daily_mac_writing_p95_ms": 195.5,
+    "daily_mac_writing_model_p95_ms": 242.8,
+    "daily_mac_writing_model_n": 11,
+    "daily_mac_writing_fp_n": 14
   },
   "line": "loudink-v1.5",
   "hot_package_note": "bases + writer_unified + ir only (no promoted_core seed)"

manifest.json

@@ -1,6 +1,6 @@
 {
   "product": "loudink-v1.5",
-  "version": "1.5.0",
+  "version": "1.5.1",
   "pipeline_namespace": "loudink_v1_5",
   "runner_kind": "loudink_v1",
   "gate_mode": "warn",
@@ -49,8 +49,8 @@
     "daily_mac_overall_shipped": 1.0,
     "daily_mac_writing_shipped": 1.0,
     "daily_mac_writing_neural": 1.0,
-    "daily_mac_writing_p50_ms": 149.9,
-    "daily_mac_writing_model_p50_ms": 173.8,
+    "daily_mac_writing_p50_ms": 0.1,
+    "daily_mac_writing_model_p50_ms": 141.3,
     "daily_mac_computer_shipped": 1.0,
     "daily_mac_multi_step_shipped": 1.0,
     "daily_mac_n": 69,
@@ -58,8 +58,12 @@
     "skill_invoke_n": 19,
     "ir_honesty_neural_n112": 0.4732,
     "hot_package_mib": 834.6,
-    "v11_hot_package_mib_same_accounting": 878.62
+    "v11_hot_package_mib_same_accounting": 878.62,
+    "daily_mac_writing_p95_ms": 195.5,
+    "daily_mac_writing_model_p95_ms": 242.8,
+    "daily_mac_writing_model_n": 11,
+    "daily_mac_writing_fp_n": 14
   },
-  "notes": "1.5.0: Beats v1.1 on writing p50, model p50, IR honesty neural; holds 100% daily-mac/skill. Hot package omits promoted_core seed (unified-only writer). Strictly smaller than v1.1 HF bundle under same accounting.",
+  "notes": "1.5.1: Honesty-safe writing FP (14/25 daily writing), SQL+paragraph structure fast paths, self-correction normalize, slim unified-only HF package, IR honesty 47.3%, gate_loudink_v1_5. Beats v1.1 on size, writing p50/p95, model p50, IR honesty.",
   "hot_package_accounting": "writer_base+ir_base+writer_unified_adapters+ir_adapters (no promoted_core seed)"
 }

product_scorecard.json

@@ -18,14 +18,14 @@
         "polish_lift": 0.0,
         "latency": {
           "n": 25,
-          "mean": 147.74175995796918,
-          "p50": 149.9218330718577,
-          "p95": 280.6902078446001,
-          "max": 308.112750062719,
-          "model_invoked_p50": 174.93549990467727,
-          "model_invoked_p95": 280.6902078446001,
-          "model_invoked_n": 20,
-          "fast_path_rate": 0.19999999999999996
+          "mean": 72.98354492622614,
+          "p50": 0.05,
+          "p95": 242.8087079897523,
+          "max": 284.9084581248462,
+          "model_invoked_p50": 141.26870781183243,
+          "model_invoked_p95": 284.9084581248462,
+          "model_invoked_n": 11,
+          "fast_path_rate": 0.56
         }
       },
       "computer": {
@@ -79,20 +79,20 @@
         },
         "latency": {
           "n": 69,
-          "mean": 53.561507231148255,
+          "mean": 26.475197437038457,
           "p50": 0.05,
-          "p95": 239.0868340153247,
-          "max": 308.112750062719,
-          "model_invoked_p50": 174.93549990467727,
-          "model_invoked_p95": 280.6902078446001,
-          "model_invoked_n": 20,
-          "fast_path_rate": 0.7101449275362319
+          "p95": 175.02208286896348,
+          "max": 284.9084581248462,
+          "model_invoked_p50": 141.26870781183243,
+          "model_invoked_p95": 284.9084581248462,
+          "model_invoked_n": 11,
+          "fast_path_rate": 0.8405797101449275
         }
       },
       "finish_reasons": {
         "transcript_ir_fast_path": 44,
-        "loudink_v1_structural_polish": 20,
-        "transcript_polish_fast_path": 5
+        "transcript_polish_fast_path": 14,
+        "loudink_v1_structural_polish": 11
       }
     }
   ]

publish_report.json

@@ -1,6 +1,6 @@
 {
   "hf_repo": "nsalerni/loudink-v1.5",
-  "version": "1.5.0",
+  "version": "1.5.1",
   "measured_adapter_bundle_mib": 26.12,
   "files": [
     "README.md",