quantprobe 1.5.0__tar.gz → 1.5.2__tar.gz

{quantprobe-1.5.0/quantprobe.egg-info → quantprobe-1.5.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: quantprobe
-Version: 1.5.0
+Version: 1.5.2
 Summary: Probe-then-quantize for LLMs: measure a model's fragility curve, plan bit/tier placement by the tiered decode law, and emit ready-to-run llama.cpp recipes.
 Author: Federico Sciuca
 License: MIT
@@ -30,7 +30,7 @@ Dynamic: license-file
 > Pick any model + your machine → predicted tok/s, memory fit, quality cost, and your cheapest next upgrade — from the law below, with your config plotted against every validated measurement.
 
 
-**New here? → [QUICKSTART.md](QUICKSTART.md) gets you running in 60 seconds.** Two tiers: `plan`/`target` and the web calculator need *nothing* installed; `quantize`/`probe`/`run`/`bench`/`dashboard` drive [llama.cpp](https://github.com/ggml-org/llama.cpp/releases) (point quantprobe at it with `--llama-dir`, `QUANTPROBE_LLAMA_DIR`, or `PATH`). Preview any command without llama.cpp using `--dry`. 16 machine presets ship in (`--machine`): GTX 1060 → RTX 5090, Apple `mac-m2/m3/m4-*`, DGX Spark, Epyc — or pass raw specs. Multi-GPU / RAID? Comma lists aggregate: `--vram 24,24 --disk-bw 14,14`. Big-VRAM + disk-streaming rigs get the three-tier expert-cache row (v1.3).
+**New here? → [QUICKSTART.md](QUICKSTART.md) gets you running in 60 seconds — or just `pip install quantprobe && quantprobe auto qwen3-coder --run`.** The calculators (`hw`/`plan`/`target`/`optimize`) need nothing installed; the weight-touching commands drive [llama.cpp](https://github.com/ggml-org/llama.cpp/releases) (point quantprobe at it with `--llama-dir`, `QUANTPROBE_LLAMA_DIR`, or `PATH`; preview anything with `--dry`). 16 machine presets ship in (`--machine`): GTX 1060 → RTX 5090, Apple `mac-m2/m3/m4-*`, DGX Spark, Epyc — or pass raw specs. Multi-GPU / RAID? Comma lists aggregate: `--vram 24,24 --disk-bw 14,14`. Big-VRAM + disk-streaming rigs get the three-tier expert-cache row (v1.3).
 
 
 ---
@@ -67,7 +67,9 @@ Add to that: a **byte-identical control** (two GGUFs the same size, 2.25 ppl apa
 <p align="center"><img src="weights/data/validation_19tok/live_run_20tps.png" width="880" alt="One frame: Task Manager showing 16 GB DDR4-3000 and the GTX 1060 6GB beside llama.cpp chatting Qwen3-30B-A3B live at 20.4 tok/s generation"></p>
 <p align="center"><em>One frame, no cuts: Task Manager (16 GB @ 3000 MT/s, GTX 1060 6 GB, RAM at 91% — the hybrid placement using the whole machine) beside llama.cpp chatting Qwen3-30B-A3B at <b>20.4 tok/s generation</b> — above the pre-registered 19. Raw logs, hardware attestation + GGUF SHA256: <a href="weights/data/validation_19tok/EVIDENCE.txt">EVIDENCE.txt</a>. Third bench run: 19.26 ± 0.45 (series 19.30 → 19.55 → 19.26).</em></p>
 
-**Open pre-registrations** — predictions staked publicly *before* measurement: [colibri v1.1, five falsifiable predictions (2026-07-23)](preregistrations/2026-07-23-colibri-v1.1.md) — dual-SSD scaling, int3 speedup, lattice-vs-scalar, AVX-512 tier-scoping, MTP×MoE antagonism.
+**Open pre-registrations** — predictions staked publicly *before* measurement: [colibri v1.1, five falsifiable predictions (2026-07-23)](preregistrations/2026-07-23-colibri-v1.1.md) — dual-SSD scaling, int3 speedup, lattice-vs-scalar, AVX-512 tier-scoping, MTP×MoE antagonism — plus open bands on the CPU-only context slope and expert-pruning decode ([all verdicts](preregistrations/)).
+
+**Live research track — Law 5 (prefill):** protocol-first and scored same-day in [weights/LAW5_PROTOCOL.md](weights/LAW5_PROTOCOL.md) — 6 staked hits, 4 published misses, and one self-caught retraction corrected in public within the hour, including the measured recipe that cuts an agent's 8k-context turn cost from ~75 s to ~0.5 s. Next up, staked before launch: [Law 6 — speculation economics](preregistrations/2026-07-24-law6-speculation-economics.md), scoring in public the week of 2026-07-27.
 
 
 ## The four placement laws
@@ -96,6 +98,7 @@ quantprobe auto qwen3-coder --tps 15 --run   # empty machine -> optimal quant ch
 
 ```bash
 quantprobe auto qwen3-30b --tps 15                       # ONE command: optimizer picks bits, closest quant fetched, run command printed
+quantprobe auto qwen3-30b --custom --run                 # THE PRODUCT: probe YOUR model, build its personalized depth-aware GGUF, launch it
 quantprobe hw                                            # what the law sees on THIS machine (every value source-tagged)
 quantprobe plan     --gguf model.gguf                    # zero-config prediction: placement + tok/s + the launch command
 quantprobe optimize --tps 20                             # CHEAPEST PATH to a target: bits x placement x hardware, Pareto-ranked
@@ -118,14 +121,15 @@ The loop is self-validating: `plan` predicted 17.5 for a file we then measured a
 
 | | |
 |---|---|
-| [QUICKSTART.md](QUICKSTART.md) | 60-second start, three levels; Ollama interop; llama.cpp version notes |
+| [QUICKSTART.md](QUICKSTART.md) | 60-second start, three levels; recipes (own fine-tune, coding agents, hardware buying); Ollama interop |
 | [LAWS.md](LAWS.md) | the four laws — statements, measurements, falsifiable predictions, the general form |
 | [docs/EXAMPLES.md](docs/EXAMPLES.md) | worked examples with real outputs: zero-config, the ×5.4 optimizer A/B, probe walkthrough, troubleshooting |
 | [docs/HARDWARE.md](docs/HARDWARE.md) | the 2016 box: exact specs, measured bandwidths, what the next euro buys |
 | [docs/DEEP-DIVE.md](docs/DEEP-DIVE.md) | what's new vs. what's built-on, parity tables, the 744B-at-home projection, repository map |
 | [preregistrations/](preregistrations/) | every staked prediction with its verdict — hits, the near-miss, and the honest miss |
+| [weights/LAW5_PROTOCOL.md](weights/LAW5_PROTOCOL.md) | the live Law-5 research ledger — conventions, stakes, scores, and the retraction |
 | [papers/arxiv/](papers/arxiv/) | the paper (submission-ready LaTeX) |
-| [CHANGELOG.md](CHANGELOG.md) | v1.0 → v1.4, every release |
+| [CHANGELOG.md](CHANGELOG.md) | v1.0 → v1.5, every release |
 
 ## Honest limitations
 

{quantprobe-1.5.0 → quantprobe-1.5.2}/README.md

@@ -12,7 +12,7 @@
 > Pick any model + your machine → predicted tok/s, memory fit, quality cost, and your cheapest next upgrade — from the law below, with your config plotted against every validated measurement.
 
 
-**New here? → [QUICKSTART.md](QUICKSTART.md) gets you running in 60 seconds.** Two tiers: `plan`/`target` and the web calculator need *nothing* installed; `quantize`/`probe`/`run`/`bench`/`dashboard` drive [llama.cpp](https://github.com/ggml-org/llama.cpp/releases) (point quantprobe at it with `--llama-dir`, `QUANTPROBE_LLAMA_DIR`, or `PATH`). Preview any command without llama.cpp using `--dry`. 16 machine presets ship in (`--machine`): GTX 1060 → RTX 5090, Apple `mac-m2/m3/m4-*`, DGX Spark, Epyc — or pass raw specs. Multi-GPU / RAID? Comma lists aggregate: `--vram 24,24 --disk-bw 14,14`. Big-VRAM + disk-streaming rigs get the three-tier expert-cache row (v1.3).
+**New here? → [QUICKSTART.md](QUICKSTART.md) gets you running in 60 seconds — or just `pip install quantprobe && quantprobe auto qwen3-coder --run`.** The calculators (`hw`/`plan`/`target`/`optimize`) need nothing installed; the weight-touching commands drive [llama.cpp](https://github.com/ggml-org/llama.cpp/releases) (point quantprobe at it with `--llama-dir`, `QUANTPROBE_LLAMA_DIR`, or `PATH`; preview anything with `--dry`). 16 machine presets ship in (`--machine`): GTX 1060 → RTX 5090, Apple `mac-m2/m3/m4-*`, DGX Spark, Epyc — or pass raw specs. Multi-GPU / RAID? Comma lists aggregate: `--vram 24,24 --disk-bw 14,14`. Big-VRAM + disk-streaming rigs get the three-tier expert-cache row (v1.3).
 
 
 ---
@@ -49,7 +49,9 @@ Add to that: a **byte-identical control** (two GGUFs the same size, 2.25 ppl apa
 <p align="center"><img src="weights/data/validation_19tok/live_run_20tps.png" width="880" alt="One frame: Task Manager showing 16 GB DDR4-3000 and the GTX 1060 6GB beside llama.cpp chatting Qwen3-30B-A3B live at 20.4 tok/s generation"></p>
 <p align="center"><em>One frame, no cuts: Task Manager (16 GB @ 3000 MT/s, GTX 1060 6 GB, RAM at 91% — the hybrid placement using the whole machine) beside llama.cpp chatting Qwen3-30B-A3B at <b>20.4 tok/s generation</b> — above the pre-registered 19. Raw logs, hardware attestation + GGUF SHA256: <a href="weights/data/validation_19tok/EVIDENCE.txt">EVIDENCE.txt</a>. Third bench run: 19.26 ± 0.45 (series 19.30 → 19.55 → 19.26).</em></p>
 
-**Open pre-registrations** — predictions staked publicly *before* measurement: [colibri v1.1, five falsifiable predictions (2026-07-23)](preregistrations/2026-07-23-colibri-v1.1.md) — dual-SSD scaling, int3 speedup, lattice-vs-scalar, AVX-512 tier-scoping, MTP×MoE antagonism.
+**Open pre-registrations** — predictions staked publicly *before* measurement: [colibri v1.1, five falsifiable predictions (2026-07-23)](preregistrations/2026-07-23-colibri-v1.1.md) — dual-SSD scaling, int3 speedup, lattice-vs-scalar, AVX-512 tier-scoping, MTP×MoE antagonism — plus open bands on the CPU-only context slope and expert-pruning decode ([all verdicts](preregistrations/)).
+
+**Live research track — Law 5 (prefill):** protocol-first and scored same-day in [weights/LAW5_PROTOCOL.md](weights/LAW5_PROTOCOL.md) — 6 staked hits, 4 published misses, and one self-caught retraction corrected in public within the hour, including the measured recipe that cuts an agent's 8k-context turn cost from ~75 s to ~0.5 s. Next up, staked before launch: [Law 6 — speculation economics](preregistrations/2026-07-24-law6-speculation-economics.md), scoring in public the week of 2026-07-27.
 
 
 ## The four placement laws
@@ -78,6 +80,7 @@ quantprobe auto qwen3-coder --tps 15 --run   # empty machine -> optimal quant ch
 
 ```bash
 quantprobe auto qwen3-30b --tps 15                       # ONE command: optimizer picks bits, closest quant fetched, run command printed
+quantprobe auto qwen3-30b --custom --run                 # THE PRODUCT: probe YOUR model, build its personalized depth-aware GGUF, launch it
 quantprobe hw                                            # what the law sees on THIS machine (every value source-tagged)
 quantprobe plan     --gguf model.gguf                    # zero-config prediction: placement + tok/s + the launch command
 quantprobe optimize --tps 20                             # CHEAPEST PATH to a target: bits x placement x hardware, Pareto-ranked
@@ -100,14 +103,15 @@ The loop is self-validating: `plan` predicted 17.5 for a file we then measured a
 
 | | |
 |---|---|
-| [QUICKSTART.md](QUICKSTART.md) | 60-second start, three levels; Ollama interop; llama.cpp version notes |
+| [QUICKSTART.md](QUICKSTART.md) | 60-second start, three levels; recipes (own fine-tune, coding agents, hardware buying); Ollama interop |
 | [LAWS.md](LAWS.md) | the four laws — statements, measurements, falsifiable predictions, the general form |
 | [docs/EXAMPLES.md](docs/EXAMPLES.md) | worked examples with real outputs: zero-config, the ×5.4 optimizer A/B, probe walkthrough, troubleshooting |
 | [docs/HARDWARE.md](docs/HARDWARE.md) | the 2016 box: exact specs, measured bandwidths, what the next euro buys |
 | [docs/DEEP-DIVE.md](docs/DEEP-DIVE.md) | what's new vs. what's built-on, parity tables, the 744B-at-home projection, repository map |
 | [preregistrations/](preregistrations/) | every staked prediction with its verdict — hits, the near-miss, and the honest miss |
+| [weights/LAW5_PROTOCOL.md](weights/LAW5_PROTOCOL.md) | the live Law-5 research ledger — conventions, stakes, scores, and the retraction |
 | [papers/arxiv/](papers/arxiv/) | the paper (submission-ready LaTeX) |
-| [CHANGELOG.md](CHANGELOG.md) | v1.0 → v1.4, every release |
+| [CHANGELOG.md](CHANGELOG.md) | v1.0 → v1.5, every release |
 
 ## Honest limitations
 

{quantprobe-1.5.0 → quantprobe-1.5.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "quantprobe"
-version = "1.5.0"
+version = "1.5.2"
 description = "Probe-then-quantize for LLMs: measure a model's fragility curve, plan bit/tier placement by the tiered decode law, and emit ready-to-run llama.cpp recipes."
 readme = "README.md"
 license = { text = "MIT" }

{quantprobe-1.5.0 → quantprobe-1.5.2}/quantprobe/__init__.py

@@ -8,4 +8,4 @@ Three commands, all grounded in measured laws (see LAWS.md in the repo):
 Every constant in `plan` is fitted from measurements published in the repo, including two
 pre-registered hardware predictions confirmed within 8%.
 """
-__version__ = "1.5.0"
+__version__ = "1.5.2"

{quantprobe-1.5.0 → quantprobe-1.5.2}/quantprobe/auto.py

@@ -37,6 +37,37 @@ def list_ggufs(repo):
                 and "mmproj" not in f["path"].lower() and "draft" not in f["path"].lower()]
 
 
+WIKI_URL = "https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip"
+
+
+def ensure_eval(dest_dir):
+    """The held-out corpus for probing, fetched once (1.3 MB)."""
+    import os, zipfile, io
+    path = os.path.join(dest_dir, "wiki.test.raw")
+    if os.path.isfile(path):
+        return path
+    print("[quantprobe auto] fetching the WikiText-2 eval corpus (1.3 MB, one time)...")
+    with urllib.request.urlopen(WIKI_URL, timeout=60) as r:
+        z = zipfile.ZipFile(io.BytesIO(r.read()))
+    with z.open("wikitext-2-raw/wiki.test.raw") as f, open(path, "wb") as out:
+        out.write(f.read())
+    return path
+
+
+def pick_source(files, t):
+    """High-precision source for the custom build: prefer Q8-class (smallest requantizable), else max bits."""
+    cands = []
+    for path, size in files:
+        if "-of-" in path.lower():
+            continue                                # split multi-part files: out of scope for v1
+        bits = size * 8 / (t * 1e9)
+        cands.append((bits, size, path))
+    if not cands:
+        return None
+    hi = sorted([c for c in cands if c[0] >= 7.5], key=lambda c: c[1])
+    return hi[0] if hi else max(cands)
+
+
 def run(a):
     target = a.target
     if target in MODEL_REPOS:
@@ -79,6 +110,40 @@ def run(a):
         raise SystemExit("no usable quant in that repo for this machine")
     scored.sort()
     _, _, path, size, bits, best = scored[0]
+
+    if getattr(a, "custom", False):
+        src = pick_source(files, t)
+        if not src:
+            raise SystemExit("no usable high-precision source in that repo (split files unsupported)")
+        sbits, ssize, spath = src
+        import os
+        dest = getattr(a, "dir", None) or "./models"
+        os.makedirs(dest, exist_ok=True)
+        print("\n[quantprobe auto --custom] source: " + spath + f" ({ssize/1e9:.1f} GB, {sbits:.1f}-bit)")
+        print("  pipeline: fetch source -> probe the fragile band (~30-60 min) -> build the")
+        print("  depth-aware GGUF. Interrupt anytime; the fetch resumes.")
+        if getattr(a, "dry", False):
+            print("  (--dry: nothing downloaded)")
+            return spath
+        from . import fetch as fetchmod
+        if not fetchmod.fetch(repo, dest, spath, fetchmod.token()):
+            raise SystemExit("source download failed (re-run: it resumes)")
+        srcfull = os.path.join(dest, os.path.basename(spath))
+        evalf = ensure_eval(dest)
+        from . import probe as probemod
+        import argparse
+        out = os.path.join(dest, os.path.basename(spath).rsplit(".gguf", 1)[0] + "-depthaware.gguf")
+        pa = argparse.Namespace(gguf=srcfull, bands=4, chunks=32, eval=evalf, ngl=99,
+                                workdir=dest, llama_dir=getattr(a, "llama_dir", None),
+                                apply=True, out=out, dry_run=False)
+        probemod.run(pa)
+        print("\n[quantprobe auto --custom] your personalized model:")
+        print("  quantprobe run --gguf " + out)
+        if getattr(a, "run", False):
+            from . import runtime
+            a.gguf = out; a.bits = None
+            runtime.run(a)
+        return out
     print(f"\n[quantprobe auto] optimizer wants ~{want_bits:g}-bit; closest file in {repo}:")
     print(f"  {path}  ({size/1e9:.1f} GB, {bits:.2f} effective bits)")
     print(f"  predicted on this machine: {best[1]:.1f} tok/s  ({best[0]})")
@@ -96,8 +161,8 @@ def run(a):
     full = os.path.join(dest, os.path.basename(path))
     print(f"\n[quantprobe auto] ready. Run it:")
     print(f"  quantprobe run --gguf {full}")
-    print(f"\n  Want better quality at the SAME size? That's the custom path (the actual product):")
-    print(f"  fetch the f16, then:  quantprobe probe --gguf <f16>.gguf --eval wiki.test.raw --apply")
+    print("\n  Better quality at the SAME size: rerun with --custom - it probes YOUR model's")
+    print("  fragile band (~30-60 min) and builds a depth-aware GGUF personalized to it.")
     if getattr(a, "run", False):
         from . import runtime
         a.gguf = full

{quantprobe-1.5.0 → quantprobe-1.5.2}/quantprobe/cli.py

@@ -99,6 +99,7 @@ def main():
     au.add_argument("--tps", type=float, default=None, help="target tok/s for the optimizer")
     au.add_argument("--max-quality", type=float, default=None)
     au.add_argument("--allow-prune", action="store_true"); au.add_argument("--any-runtime", action="store_true")
+    au.add_argument("--custom", action="store_true", help="THE PRODUCT: fetch a high-precision source, probe YOUR model (~30-60 min), build a depth-aware GGUF personalized to it")
     au.add_argument("--serve", action="store_true", help="with --run: llama-server instead of chat")
     au.add_argument("--extra", default=None)
 

{quantprobe-1.5.0 → quantprobe-1.5.2}/quantprobe/optimize.py

@@ -35,6 +35,7 @@ HW_DELTAS = [
 
 
 def resolve(a):
+    planmod.check_presets(a)
     """Model + machine resolution, same semantics as plan.run (autospec + autodetect included)."""
     from . import spec as specmod
     specmod.apply(a, quiet=True)

{quantprobe-1.5.0 → quantprobe-1.5.2}/quantprobe/plan.py

@@ -144,9 +144,27 @@ def qual_of(moe, bits):
     return QUAL[moe][min(keys, key=lambda k: abs(k - bits))]
 
 
+def check_presets(args):
+    """Refuse unknown preset names LOUDLY instead of silently falling back to defaults."""
+    mdl = getattr(args, "model", None)
+    if mdl and mdl not in MODELS and not getattr(args, "total", None):
+        raise SystemExit(
+            "unknown --model '%s' (and no --total to describe it).\n"
+            "  presets: %s\n"
+            "  or describe it:      --total <B> --active <B> [--always-active <B>]\n"
+            "  or point at a file:  --gguf model.gguf   (exact spec read from the GGUF)" % (mdl, ", ".join(sorted(MODELS))))
+    mac = getattr(args, "machine", None)
+    if mac and mac not in MACHINES:
+        raise SystemExit(
+            "unknown --machine '%s'.\n"
+            "  presets: %s\n"
+            "  or pass flags: --vram/--vram-bw/--ram/--ram-bw/--disk-bw   (no flags = auto-detect this box)" % (mac, ", ".join(sorted(MACHINES))))
+
+
 def run(args):
     from . import spec as specmod
     specmod.apply(args)
+    check_presets(args)
     if getattr(args, "bits", None) is None:
         args.bits = 2.5
     m = dict(MODELS[args.model]) if args.model in MODELS else {}

{quantprobe-1.5.0 → quantprobe-1.5.2}/quantprobe/target.py

@@ -52,6 +52,8 @@ def feasible(a, tps_target):
 
 
 def run(a):
+    from . import plan as planmod
+    planmod.check_presets(a)
     hw, vc, vb, rc, rb, db, geta, gl = hw_of(a)
     rows = feasible(a, a.tps)
     print(f"\nquantprobe target - smartest model meeting >= {a.tps:g} tok/s "

{quantprobe-1.5.0 → quantprobe-1.5.2/quantprobe.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: quantprobe
-Version: 1.5.0
+Version: 1.5.2
 Summary: Probe-then-quantize for LLMs: measure a model's fragility curve, plan bit/tier placement by the tiered decode law, and emit ready-to-run llama.cpp recipes.
 Author: Federico Sciuca
 License: MIT
@@ -30,7 +30,7 @@ Dynamic: license-file
 > Pick any model + your machine → predicted tok/s, memory fit, quality cost, and your cheapest next upgrade — from the law below, with your config plotted against every validated measurement.
 
 
-**New here? → [QUICKSTART.md](QUICKSTART.md) gets you running in 60 seconds.** Two tiers: `plan`/`target` and the web calculator need *nothing* installed; `quantize`/`probe`/`run`/`bench`/`dashboard` drive [llama.cpp](https://github.com/ggml-org/llama.cpp/releases) (point quantprobe at it with `--llama-dir`, `QUANTPROBE_LLAMA_DIR`, or `PATH`). Preview any command without llama.cpp using `--dry`. 16 machine presets ship in (`--machine`): GTX 1060 → RTX 5090, Apple `mac-m2/m3/m4-*`, DGX Spark, Epyc — or pass raw specs. Multi-GPU / RAID? Comma lists aggregate: `--vram 24,24 --disk-bw 14,14`. Big-VRAM + disk-streaming rigs get the three-tier expert-cache row (v1.3).
+**New here? → [QUICKSTART.md](QUICKSTART.md) gets you running in 60 seconds — or just `pip install quantprobe && quantprobe auto qwen3-coder --run`.** The calculators (`hw`/`plan`/`target`/`optimize`) need nothing installed; the weight-touching commands drive [llama.cpp](https://github.com/ggml-org/llama.cpp/releases) (point quantprobe at it with `--llama-dir`, `QUANTPROBE_LLAMA_DIR`, or `PATH`; preview anything with `--dry`). 16 machine presets ship in (`--machine`): GTX 1060 → RTX 5090, Apple `mac-m2/m3/m4-*`, DGX Spark, Epyc — or pass raw specs. Multi-GPU / RAID? Comma lists aggregate: `--vram 24,24 --disk-bw 14,14`. Big-VRAM + disk-streaming rigs get the three-tier expert-cache row (v1.3).
 
 
 ---
@@ -67,7 +67,9 @@ Add to that: a **byte-identical control** (two GGUFs the same size, 2.25 ppl apa
 <p align="center"><img src="weights/data/validation_19tok/live_run_20tps.png" width="880" alt="One frame: Task Manager showing 16 GB DDR4-3000 and the GTX 1060 6GB beside llama.cpp chatting Qwen3-30B-A3B live at 20.4 tok/s generation"></p>
 <p align="center"><em>One frame, no cuts: Task Manager (16 GB @ 3000 MT/s, GTX 1060 6 GB, RAM at 91% — the hybrid placement using the whole machine) beside llama.cpp chatting Qwen3-30B-A3B at <b>20.4 tok/s generation</b> — above the pre-registered 19. Raw logs, hardware attestation + GGUF SHA256: <a href="weights/data/validation_19tok/EVIDENCE.txt">EVIDENCE.txt</a>. Third bench run: 19.26 ± 0.45 (series 19.30 → 19.55 → 19.26).</em></p>
 
-**Open pre-registrations** — predictions staked publicly *before* measurement: [colibri v1.1, five falsifiable predictions (2026-07-23)](preregistrations/2026-07-23-colibri-v1.1.md) — dual-SSD scaling, int3 speedup, lattice-vs-scalar, AVX-512 tier-scoping, MTP×MoE antagonism.
+**Open pre-registrations** — predictions staked publicly *before* measurement: [colibri v1.1, five falsifiable predictions (2026-07-23)](preregistrations/2026-07-23-colibri-v1.1.md) — dual-SSD scaling, int3 speedup, lattice-vs-scalar, AVX-512 tier-scoping, MTP×MoE antagonism — plus open bands on the CPU-only context slope and expert-pruning decode ([all verdicts](preregistrations/)).
+
+**Live research track — Law 5 (prefill):** protocol-first and scored same-day in [weights/LAW5_PROTOCOL.md](weights/LAW5_PROTOCOL.md) — 6 staked hits, 4 published misses, and one self-caught retraction corrected in public within the hour, including the measured recipe that cuts an agent's 8k-context turn cost from ~75 s to ~0.5 s. Next up, staked before launch: [Law 6 — speculation economics](preregistrations/2026-07-24-law6-speculation-economics.md), scoring in public the week of 2026-07-27.
 
 
 ## The four placement laws
@@ -96,6 +98,7 @@ quantprobe auto qwen3-coder --tps 15 --run   # empty machine -> optimal quant ch
 
 ```bash
 quantprobe auto qwen3-30b --tps 15                       # ONE command: optimizer picks bits, closest quant fetched, run command printed
+quantprobe auto qwen3-30b --custom --run                 # THE PRODUCT: probe YOUR model, build its personalized depth-aware GGUF, launch it
 quantprobe hw                                            # what the law sees on THIS machine (every value source-tagged)
 quantprobe plan     --gguf model.gguf                    # zero-config prediction: placement + tok/s + the launch command
 quantprobe optimize --tps 20                             # CHEAPEST PATH to a target: bits x placement x hardware, Pareto-ranked
@@ -118,14 +121,15 @@ The loop is self-validating: `plan` predicted 17.5 for a file we then measured a
 
 | | |
 |---|---|
-| [QUICKSTART.md](QUICKSTART.md) | 60-second start, three levels; Ollama interop; llama.cpp version notes |
+| [QUICKSTART.md](QUICKSTART.md) | 60-second start, three levels; recipes (own fine-tune, coding agents, hardware buying); Ollama interop |
 | [LAWS.md](LAWS.md) | the four laws — statements, measurements, falsifiable predictions, the general form |
 | [docs/EXAMPLES.md](docs/EXAMPLES.md) | worked examples with real outputs: zero-config, the ×5.4 optimizer A/B, probe walkthrough, troubleshooting |
 | [docs/HARDWARE.md](docs/HARDWARE.md) | the 2016 box: exact specs, measured bandwidths, what the next euro buys |
 | [docs/DEEP-DIVE.md](docs/DEEP-DIVE.md) | what's new vs. what's built-on, parity tables, the 744B-at-home projection, repository map |
 | [preregistrations/](preregistrations/) | every staked prediction with its verdict — hits, the near-miss, and the honest miss |
+| [weights/LAW5_PROTOCOL.md](weights/LAW5_PROTOCOL.md) | the live Law-5 research ledger — conventions, stakes, scores, and the retraction |
 | [papers/arxiv/](papers/arxiv/) | the paper (submission-ready LaTeX) |
-| [CHANGELOG.md](CHANGELOG.md) | v1.0 → v1.4, every release |
+| [CHANGELOG.md](CHANGELOG.md) | v1.0 → v1.5, every release |
 
 ## Honest limitations