szl-quant-bench 0.2.0__tar.gz

szl_quant_bench-0.2.0/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 SZL Holdings
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

szl_quant_bench-0.2.0/PKG-INFO

@@ -0,0 +1,81 @@
+Metadata-Version: 2.4
+Name: szl-quant-bench
+Version: 0.2.0
+Summary: Honest quantization quality-curve bench: cosine/KL/top-1 degradation vs bit-width, hash-chained receipts. Fail closed.
+License: Apache-2.0
+Requires-Python: >=3.11
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Dynamic: license-file
+
+# szl-quant-bench
+
+Honest quantization quality-curve bench for the SZL estate. Real uniform
+absmax quantization (round-to-nearest, signed ints, 2–16 bits) with quality
+metrics — cosine similarity, KL divergence over softmaxed rows, top-1
+agreement, MSE — on the supplied logits. This measures logit quantization,
+not weight quantization, GGUF behavior, inference speed, or model quality
+after a weight conversion. Stdlib only; no downloads, no network.
+
+Doctrine: empty or missing inputs return `BLOCKED` with a reason. The demo
+fixture is synthetic and labeled as such. Nothing here claims to measure a
+real model unless you pass real logits in. Run states:
+`MEASURED | BLOCKED | INVALID | FAILED`.
+
+## Install and verify
+
+```
+pip install -e . pytest
+python -m pytest tests/ -q
+python -m szl_quant_bench.harness       # deterministic fixture curve + receipt
+```
+
+## Measured fixture curve (synthetic gaussian logits, seed 42 — NOT a real model)
+
+| bits | cosine | KL (mean) | top-1 agreement | vs fp32 |
+|---|---|---|---|---|
+| 16 | 1.0000 | 0.00000 | 100.0% | 2.0x |
+| 8 | 1.0000 | 0.00003 | 100.0% | 4.0x |
+| 4 | 0.9882 | 0.01211 | 68.8% | 8.0x |
+| 3 | 0.9420 | 0.06556 | 56.2% | 10.67x |
+| 2 | 0.5715 | 0.57019 | 50.0% | 16.0x |
+
+These numbers came out of `run_curve` in this repository, executed before the
+initial push. Your model's curve will differ — run it on real logits.
+
+## Measuring a real model
+
+Capture a logit batch from the exact model revision and fixed evaluation
+prompt set. Store a JSON object with `logits` (a rectangular finite numeric
+matrix) and `provenance` containing:
+
+- `source_kind`: `REAL_MODEL_LOGITS`
+- `model_id`, `runtime`, `hardware`: nonempty identifiers
+- `model_revision`: the exact 40-character lowercase commit
+- `prompt_set_sha256`: the full lowercase SHA-256 of the saved prompt artifact
+
+```sh
+python -m szl_quant_bench.harness --input model-logits.json --output curve-receipt.json
+```
+
+The output file must not already exist. Missing provenance, malformed matrices,
+non-finite values and invalid bit widths fail closed with a nonzero exit.
+With no input the CLI runs only the explicitly `SYNTHETIC` fixture.
+The Python API accepts `run_curve(logits, chain=chain, provenance=provenance)`;
+omitting provenance labels the source `UNVERIFIED_CALLER_INPUT`. Every
+`MEASURED` API result carries a hash-chained receipt record. If no chain is
+supplied, the API creates a fresh one-record chain for that measurement; pass
+a caller-owned `ReceiptChain` to append the record to an existing chain.
+
+Receipts bind the normalized float input matrix hash, complete provenance,
+all quality records and the explicit measurement boundary. Provenance remains
+`CALLER_DECLARED`: a hash cannot establish that the caller actually ran a model.
+Preserve the raw logits and independent model execution evidence to support
+that claim. Empty or malformed chains do not verify. KL is evaluated in log
+space without silently clipping small probabilities.
+
+Compare bit widths only on the same inputs, hardware and revision. A measured
+logit curve does **not** authorize a GGUF export decision: actual quantized
+model outputs and runtime evaluation are separately required.
+
+Apache-2.0 · Doctrine v11 · SZL Holdings

szl_quant_bench-0.2.0/README.md

@@ -0,0 +1,71 @@
+# szl-quant-bench
+
+Honest quantization quality-curve bench for the SZL estate. Real uniform
+absmax quantization (round-to-nearest, signed ints, 2–16 bits) with quality
+metrics — cosine similarity, KL divergence over softmaxed rows, top-1
+agreement, MSE — on the supplied logits. This measures logit quantization,
+not weight quantization, GGUF behavior, inference speed, or model quality
+after a weight conversion. Stdlib only; no downloads, no network.
+
+Doctrine: empty or missing inputs return `BLOCKED` with a reason. The demo
+fixture is synthetic and labeled as such. Nothing here claims to measure a
+real model unless you pass real logits in. Run states:
+`MEASURED | BLOCKED | INVALID | FAILED`.
+
+## Install and verify
+
+```
+pip install -e . pytest
+python -m pytest tests/ -q
+python -m szl_quant_bench.harness       # deterministic fixture curve + receipt
+```
+
+## Measured fixture curve (synthetic gaussian logits, seed 42 — NOT a real model)
+
+| bits | cosine | KL (mean) | top-1 agreement | vs fp32 |
+|---|---|---|---|---|
+| 16 | 1.0000 | 0.00000 | 100.0% | 2.0x |
+| 8 | 1.0000 | 0.00003 | 100.0% | 4.0x |
+| 4 | 0.9882 | 0.01211 | 68.8% | 8.0x |
+| 3 | 0.9420 | 0.06556 | 56.2% | 10.67x |
+| 2 | 0.5715 | 0.57019 | 50.0% | 16.0x |
+
+These numbers came out of `run_curve` in this repository, executed before the
+initial push. Your model's curve will differ — run it on real logits.
+
+## Measuring a real model
+
+Capture a logit batch from the exact model revision and fixed evaluation
+prompt set. Store a JSON object with `logits` (a rectangular finite numeric
+matrix) and `provenance` containing:
+
+- `source_kind`: `REAL_MODEL_LOGITS`
+- `model_id`, `runtime`, `hardware`: nonempty identifiers
+- `model_revision`: the exact 40-character lowercase commit
+- `prompt_set_sha256`: the full lowercase SHA-256 of the saved prompt artifact
+
+```sh
+python -m szl_quant_bench.harness --input model-logits.json --output curve-receipt.json
+```
+
+The output file must not already exist. Missing provenance, malformed matrices,
+non-finite values and invalid bit widths fail closed with a nonzero exit.
+With no input the CLI runs only the explicitly `SYNTHETIC` fixture.
+The Python API accepts `run_curve(logits, chain=chain, provenance=provenance)`;
+omitting provenance labels the source `UNVERIFIED_CALLER_INPUT`. Every
+`MEASURED` API result carries a hash-chained receipt record. If no chain is
+supplied, the API creates a fresh one-record chain for that measurement; pass
+a caller-owned `ReceiptChain` to append the record to an existing chain.
+
+Receipts bind the normalized float input matrix hash, complete provenance,
+all quality records and the explicit measurement boundary. Provenance remains
+`CALLER_DECLARED`: a hash cannot establish that the caller actually ran a model.
+Preserve the raw logits and independent model execution evidence to support
+that claim. Empty or malformed chains do not verify. KL is evaluated in log
+space without silently clipping small probabilities.
+
+Compare bit widths only on the same inputs, hardware and revision. A measured
+logit curve does **not** authorize a GGUF export decision: actual quantized
+model outputs and runtime evaluation are separately required.
+
+Apache-2.0 · Doctrine v11 · SZL Holdings

szl_quant_bench-0.2.0/pyproject.toml

@@ -0,0 +1,17 @@
+[build-system]
+requires = ["setuptools>=68"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "szl-quant-bench"
+version = "0.2.0"
+description = "Honest quantization quality-curve bench: cosine/KL/top-1 degradation vs bit-width, hash-chained receipts. Fail closed."
+readme = "README.md"
+requires-python = ">=3.11"
+license = {text = "Apache-2.0"}
+
+[tool.setuptools.packages.find]
+where = ["src"]
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]

szl_quant_bench-0.2.0/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

szl_quant_bench-0.2.0/src/szl_quant_bench/__init__.py

@@ -0,0 +1,2 @@
+"""szl-quant-bench — honest quantization quality curves. Fail closed."""
+__version__ = "0.2.0"

szl_quant_bench-0.2.0/src/szl_quant_bench/harness.py

@@ -0,0 +1,138 @@
+"""Quantization quality-curve harness. MEASURED only from real math; else BLOCKED.
+
+Run states: MEASURED | BLOCKED | INVALID | FAILED. Never fabricate a number.
+"""
+import argparse
+import hashlib
+import json
+import math
+import re
+from pathlib import Path
+
+from .metrics import cosine, kl_divergence, mse, top1_agreement
+from .quant import dequantize, quantize
+from .receipts import ReceiptChain
+
+STATES = ("MEASURED", "BLOCKED", "INVALID", "FAILED")
+DEFAULT_BITS = (16, 8, 4, 3, 2)
+
+
+def _canonical(value):
+    return json.dumps(value, sort_keys=True, separators=(",", ":"),
+                      ensure_ascii=False, allow_nan=False)
+
+
+def _provenance(value):
+    if value is None:
+        return {"source_kind": "UNVERIFIED_CALLER_INPUT"}
+    if not isinstance(value, dict):
+        raise ValueError("provenance must be an object")
+    result = json.loads(_canonical(value))
+    if result.get("source_kind") not in ("SYNTHETIC", "REAL_MODEL_LOGITS"):
+        raise ValueError("source_kind must be SYNTHETIC or REAL_MODEL_LOGITS")
+    if result["source_kind"] == "REAL_MODEL_LOGITS":
+        for key in ("model_id", "runtime", "hardware"):
+            if not isinstance(result.get(key), str) or not result[key].strip():
+                raise ValueError(f"real-model provenance requires {key}")
+        if not re.fullmatch(r"[0-9a-f]{40}", result.get("model_revision", "")):
+            raise ValueError("model_revision must be an exact 40-character commit")
+        if not re.fullmatch(r"[0-9a-f]{64}", result.get("prompt_set_sha256", "")):
+            raise ValueError("prompt_set_sha256 must be a full SHA-256")
+    return result
+
+
+def run_curve(logits, bits=DEFAULT_BITS, chain=None, *, provenance=None):
+    """logits: list of rows (list of float), e.g. output logits for a token batch.
+
+    Returns per-bit quality records computed from actual quantize/dequantize.
+    Synthetic-fixture results are labeled as such by the caller; nothing here
+    claims to measure a real model unless real logits were passed in.
+    """
+    if logits is None or (isinstance(logits, (list, tuple)) and
+                          (not logits or all(isinstance(row, (list, tuple)) and not row
+                                             for row in logits))):
+        return {"state": "BLOCKED", "reason": "empty logits matrix"}
+    try:
+        if not isinstance(logits, (list, tuple)) or not logits:
+            raise ValueError("logits must be a nonempty rectangular matrix")
+        if any(not isinstance(row, (list, tuple)) or not row for row in logits):
+            raise ValueError("each logits row must be a nonempty array")
+        n_cols = len(logits[0])
+        if any(len(row) != n_cols for row in logits):
+            raise ValueError("logits rows must have equal length")
+        if any(isinstance(x, bool) or not isinstance(x, (int, float))
+               or not math.isfinite(float(x)) for row in logits for x in row):
+            raise ValueError("logits must contain only finite real numbers")
+        if not isinstance(bits, (list, tuple)) or not bits:
+            raise ValueError("bits must be a nonempty array")
+        if any(type(b) is not int or not 2 <= b <= 16 for b in bits):
+            raise ValueError("each bit width must be an integer in 2..16")
+        if len(set(bits)) != len(bits):
+            raise ValueError("duplicate bit widths are not allowed")
+        origin = _provenance(provenance)
+        logits = [[float(x) for x in row] for row in logits]
+        input_hash = hashlib.sha256(_canonical(logits).encode()).hexdigest()
+        flat = [x for row in logits for x in row]
+        records = []
+        for b in bits:
+            ints, scale = quantize(flat, b)
+            deq = dequantize(ints, scale)
+            rows_dq = [deq[i * n_cols:(i + 1) * n_cols] for i in range(len(logits))]
+            records.append({
+                "bits": b,
+                "mse": round(mse(flat, deq), 8),
+                "cosine": round(sum(cosine(a, d) for a, d in zip(logits, rows_dq)) / len(logits), 6),
+                "kl_mean": round(sum(kl_divergence(a, d) for a, d in zip(logits, rows_dq)) / len(logits), 8),
+                "top1_agreement": round(top1_agreement(logits, rows_dq), 6),
+                "compression_vs_fp32": round(32.0 / b, 2),
+            })
+        _canonical(records)  # reject overflow or any non-finite computed metric
+    except (ValueError, TypeError, OverflowError, ZeroDivisionError) as exc:
+        return {"state": "INVALID", "reason": str(exc)}
+    result = {"state": "MEASURED", "lane": "quant-curve", "rows": len(logits),
+              "cols": n_cols, "curve": records, "input_sha256": input_hash,
+              "provenance": origin, "provenance_authority": "CALLER_DECLARED",
+              "measurement": "UNIFORM_LOGIT_QUANTIZATION",
+              "weight_quantization_measured": False,
+              "gguf_export_authorized": False}
+    receipt_chain = chain if chain is not None else ReceiptChain()
+    result["receipt"] = receipt_chain.emit({"type": "quant_curve", **result})
+    return result
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--input", type=Path, help="JSON object with logits and provenance")
+    parser.add_argument("--output", type=Path, help="new receipt JSON file; never overwritten")
+    args = parser.parse_args()
+    from random import Random
+    rng = Random(42)  # deterministic synthetic fixture, honestly labeled
+    logits = [[rng.gauss(0, 1) for _ in range(64)] for _ in range(16)]
+    origin = {"source_kind": "SYNTHETIC", "fixture": "gaussian seed 42"}
+    chain = ReceiptChain()
+    try:
+        if args.input:
+            if args.input.stat().st_size > 128 * 1024 * 1024:
+                raise ValueError("input exceeds 128 MiB")
+            envelope = json.loads(args.input.read_text(encoding="utf-8"))
+            if not isinstance(envelope, dict) or "provenance" not in envelope:
+                raise ValueError("input requires logits and provenance")
+            logits, origin = envelope.get("logits"), envelope["provenance"]
+        out = run_curve(logits, chain=chain, provenance=origin)
+    except (OSError, ValueError, TypeError) as exc:
+        out = {"state": "INVALID", "reason": str(exc)}
+    out["chain"] = chain.chain
+    out["chain_valid"] = chain.verify()
+    encoded = json.dumps(out, indent=2, ensure_ascii=False, allow_nan=False) + "\n"
+    if args.output:
+        try:
+            with args.output.open("x", encoding="utf-8") as destination:
+                destination.write(encoded)
+        except OSError as exc:
+            parser.error(str(exc))
+    print(encoded, end="")
+    return 0 if out["state"] == "MEASURED" else 2
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

szl_quant_bench-0.2.0/src/szl_quant_bench/metrics.py

@@ -0,0 +1,40 @@
+"""Quality metrics between original and dequantized vectors/matrices."""
+import math
+
+
+def mse(a, b):
+    return sum((x - y) ** 2 for x, y in zip(a, b)) / max(1, len(a))
+
+
+def cosine(a, b):
+    dot = sum(x * y for x, y in zip(a, b))
+    na = math.sqrt(sum(x * x for x in a)) or 1e-12
+    nb = math.sqrt(sum(y * y for y in b)) or 1e-12
+    return dot / (na * nb)
+
+
+def _softmax(row):
+    m = max(row)
+    e = [math.exp(x - m) for x in row]
+    s = sum(e)
+    return [x / s for x in e]
+
+
+def kl_divergence(p_row, q_row):
+    """KL(P || Q) over softmaxed rows. >= 0; 0 means identical distributions."""
+    if len(p_row) != len(q_row) or not p_row:
+        raise ValueError("KL rows must have equal nonzero length")
+    # Work in log space: clipping Q would silently cap large divergences.
+    pm, qm = max(p_row), max(q_row)
+    pz = math.log(sum(math.exp(x - pm) for x in p_row))
+    qz = math.log(sum(math.exp(x - qm) for x in q_row))
+    log_p = [x - pm - pz for x in p_row]
+    log_q = [x - qm - qz for x in q_row]
+    return max(0.0, sum(math.exp(p) * (p - q)
+                        for p, q in zip(log_p, log_q) if math.exp(p) > 0))
+
+
+def top1_agreement(rows_a, rows_b):
+    agree = sum(1 for a, b in zip(rows_a, rows_b)
+                if a.index(max(a)) == b.index(max(b)))
+    return agree / max(1, len(rows_a))

szl_quant_bench-0.2.0/src/szl_quant_bench/quant.py

@@ -0,0 +1,21 @@
+"""Uniform symmetric quantization: quantize/dequantize with absmax scales.
+
+Real integer quantization math, stdlib only. Per-tensor absmax scale,
+round-to-nearest, signed integers at the requested bit width.
+"""
+
+
+def quantize(values, bits):
+    """values: flat list of floats. Returns (ints, scale). bits in 2..16."""
+    if type(bits) is not int or not 2 <= bits <= 16:
+        raise ValueError("bits must be in 2..16")
+    if not values:
+        return [], 1.0
+    qmax = (1 << (bits - 1)) - 1
+    absmax = max(abs(v) for v in values) or 1.0
+    scale = absmax / qmax
+    return [int(round(v / scale)) for v in values], scale
+
+
+def dequantize(ints, scale):
+    return [i * scale for i in ints]

szl_quant_bench-0.2.0/src/szl_quant_bench/receipts.py

@@ -0,0 +1,39 @@
+"""Hash-chained run receipts. UNSIGNED_HONEST: proves integrity + order, not identity."""
+import hashlib
+import json
+import time
+
+GENESIS = "0" * 64
+
+
+def _canonical(obj):
+    return json.dumps(obj, sort_keys=True, separators=(",", ":"), ensure_ascii=False,
+                      allow_nan=False)
+
+
+class ReceiptChain:
+    def __init__(self):
+        self.chain = []
+
+    def emit(self, record):
+        prev = self.chain[-1]["self_hash"] if self.chain else GENESIS
+        body = {"prev_hash": prev, "ts": time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime()),
+                "signature": "UNSIGNED_HONEST", "run": json.loads(_canonical(record))}
+        body["self_hash"] = hashlib.sha256(_canonical(body).encode()).hexdigest()
+        self.chain.append(body)
+        return body
+
+    def verify(self):
+        prev = GENESIS
+        try:
+            for r in self.chain:
+                if not isinstance(r, dict) or r["prev_hash"] != prev:
+                    return False
+                check = dict(r)
+                h = check.pop("self_hash")
+                if hashlib.sha256(_canonical(check).encode()).hexdigest() != h:
+                    return False
+                prev = h
+        except (KeyError, TypeError, ValueError):
+            return False
+        return bool(self.chain)

szl_quant_bench-0.2.0/src/szl_quant_bench.egg-info/PKG-INFO

@@ -0,0 +1,81 @@
+Metadata-Version: 2.4
+Name: szl-quant-bench
+Version: 0.2.0
+Summary: Honest quantization quality-curve bench: cosine/KL/top-1 degradation vs bit-width, hash-chained receipts. Fail closed.
+License: Apache-2.0
+Requires-Python: >=3.11
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Dynamic: license-file
+
+# szl-quant-bench
+
+Honest quantization quality-curve bench for the SZL estate. Real uniform
+absmax quantization (round-to-nearest, signed ints, 2–16 bits) with quality
+metrics — cosine similarity, KL divergence over softmaxed rows, top-1
+agreement, MSE — on the supplied logits. This measures logit quantization,
+not weight quantization, GGUF behavior, inference speed, or model quality
+after a weight conversion. Stdlib only; no downloads, no network.
+
+Doctrine: empty or missing inputs return `BLOCKED` with a reason. The demo
+fixture is synthetic and labeled as such. Nothing here claims to measure a
+real model unless you pass real logits in. Run states:
+`MEASURED | BLOCKED | INVALID | FAILED`.
+
+## Install and verify
+
+```
+pip install -e . pytest
+python -m pytest tests/ -q
+python -m szl_quant_bench.harness       # deterministic fixture curve + receipt
+```
+
+## Measured fixture curve (synthetic gaussian logits, seed 42 — NOT a real model)
+
+| bits | cosine | KL (mean) | top-1 agreement | vs fp32 |
+|---|---|---|---|---|
+| 16 | 1.0000 | 0.00000 | 100.0% | 2.0x |
+| 8 | 1.0000 | 0.00003 | 100.0% | 4.0x |
+| 4 | 0.9882 | 0.01211 | 68.8% | 8.0x |
+| 3 | 0.9420 | 0.06556 | 56.2% | 10.67x |
+| 2 | 0.5715 | 0.57019 | 50.0% | 16.0x |
+
+These numbers came out of `run_curve` in this repository, executed before the
+initial push. Your model's curve will differ — run it on real logits.
+
+## Measuring a real model
+
+Capture a logit batch from the exact model revision and fixed evaluation
+prompt set. Store a JSON object with `logits` (a rectangular finite numeric
+matrix) and `provenance` containing:
+
+- `source_kind`: `REAL_MODEL_LOGITS`
+- `model_id`, `runtime`, `hardware`: nonempty identifiers
+- `model_revision`: the exact 40-character lowercase commit
+- `prompt_set_sha256`: the full lowercase SHA-256 of the saved prompt artifact
+
+```sh
+python -m szl_quant_bench.harness --input model-logits.json --output curve-receipt.json
+```
+
+The output file must not already exist. Missing provenance, malformed matrices,
+non-finite values and invalid bit widths fail closed with a nonzero exit.
+With no input the CLI runs only the explicitly `SYNTHETIC` fixture.
+The Python API accepts `run_curve(logits, chain=chain, provenance=provenance)`;
+omitting provenance labels the source `UNVERIFIED_CALLER_INPUT`. Every
+`MEASURED` API result carries a hash-chained receipt record. If no chain is
+supplied, the API creates a fresh one-record chain for that measurement; pass
+a caller-owned `ReceiptChain` to append the record to an existing chain.
+
+Receipts bind the normalized float input matrix hash, complete provenance,
+all quality records and the explicit measurement boundary. Provenance remains
+`CALLER_DECLARED`: a hash cannot establish that the caller actually ran a model.
+Preserve the raw logits and independent model execution evidence to support
+that claim. Empty or malformed chains do not verify. KL is evaluated in log
+space without silently clipping small probabilities.
+
+Compare bit widths only on the same inputs, hardware and revision. A measured
+logit curve does **not** authorize a GGUF export decision: actual quantized
+model outputs and runtime evaluation are separately required.
+
+Apache-2.0 · Doctrine v11 · SZL Holdings

szl_quant_bench-0.2.0/src/szl_quant_bench.egg-info/SOURCES.txt

@@ -0,0 +1,15 @@
+LICENSE
+README.md
+pyproject.toml
+src/szl_quant_bench/__init__.py
+src/szl_quant_bench/harness.py
+src/szl_quant_bench/metrics.py
+src/szl_quant_bench/quant.py
+src/szl_quant_bench/receipts.py
+src/szl_quant_bench.egg-info/PKG-INFO
+src/szl_quant_bench.egg-info/SOURCES.txt
+src/szl_quant_bench.egg-info/dependency_links.txt
+src/szl_quant_bench.egg-info/top_level.txt
+tests/test_provenance.py
+tests/test_quant_bench.py
+tests/test_receipt_required.py

szl_quant_bench-0.2.0/src/szl_quant_bench.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

szl_quant_bench-0.2.0/src/szl_quant_bench.egg-info/top_level.txt

@@ -0,0 +1 @@
+szl_quant_bench

szl_quant_bench-0.2.0/tests/test_provenance.py

@@ -0,0 +1,115 @@
+import hashlib
+import json
+import subprocess
+import sys
+
+import pytest
+
+from szl_quant_bench.harness import run_curve
+from szl_quant_bench.metrics import kl_divergence
+from szl_quant_bench.quant import quantize
+from szl_quant_bench.receipts import ReceiptChain
+
+
+@pytest.mark.parametrize("logits", ["bad", [1], [[1], [2, 3]],
+                                  [[True]], [[float("nan")]],
+                                  [[float("inf")]], [["1"]],
+                                  [[1e308, -1e308]]])
+def test_invalid_matrix_never_emits_receipt(logits):
+    chain = ReceiptChain()
+    assert run_curve(logits, chain=chain)["state"] == "INVALID"
+    assert chain.chain == []
+
+
+@pytest.mark.parametrize("logits", [None, [], [[]], [[], []], ((), ())])
+def test_missing_or_wholly_empty_logits_are_blocked(logits):
+    chain = ReceiptChain()
+    assert run_curve(logits, chain=chain)["state"] == "BLOCKED"
+    assert chain.chain == []
+
+
+@pytest.mark.parametrize("bits", [[], [True], [4.0], [1], [17], [4, 4], "4"])
+def test_invalid_bits_never_emit_receipt(bits):
+    chain = ReceiptChain()
+    assert run_curve([[1.0, 2.0]], bits=bits, chain=chain)["state"] == "INVALID"
+    assert chain.chain == []
+
+
+def test_provenance_and_input_are_bound_to_receipt():
+    provenance = {
+        "source_kind": "REAL_MODEL_LOGITS", "model_id": "owner/model",
+        "model_revision": "a" * 40, "prompt_set_sha256": "b" * 64,
+        "runtime": "test-runtime", "hardware": "test-host",
+    }
+    logits = [[1.0, 2.0], [3.0, 4.0]]
+    chain = ReceiptChain()
+    out = run_curve(logits, chain=chain, provenance=provenance)
+    assert out["state"] == "MEASURED"
+    canonical = json.dumps(logits, sort_keys=True, separators=(",", ":"),
+                           ensure_ascii=False, allow_nan=False)
+    assert out["input_sha256"] == hashlib.sha256(canonical.encode()).hexdigest()
+    assert out["provenance_authority"] == "CALLER_DECLARED"
+    assert out["weight_quantization_measured"] is False
+    assert out["gguf_export_authorized"] is False
+    provenance["model_id"] = "mutated"
+    out["curve"][0]["mse"] = 999
+    out["provenance"]["model_id"] = "mutated-too"
+    assert chain.verify()
+    chain.chain[0]["run"]["input_sha256"] = "0" * 64
+    assert not chain.verify()
+
+
+def test_missing_real_provenance_is_rejected():
+    out = run_curve([[1.0]], provenance={"source_kind": "REAL_MODEL_LOGITS"})
+    assert out["state"] == "INVALID"
+
+
+def test_unverified_caller_is_not_labeled_real():
+    assert run_curve([[1.0]])["provenance"]["source_kind"] == "UNVERIFIED_CALLER_INPUT"
+
+
+def test_kl_does_not_clip_large_divergence():
+    assert kl_divergence([1000, 0], [0, 1000]) == pytest.approx(1000)
+    with pytest.raises(ValueError):
+        kl_divergence([1], [1, 2])
+
+
+@pytest.mark.parametrize("bits", [True, 4.0, "4"])
+def test_quantize_rejects_noninteger_bits(bits):
+    with pytest.raises(ValueError):
+        quantize([1.0], bits)
+
+
+def test_empty_and_malformed_chains_fail_closed():
+    chain = ReceiptChain()
+    assert not chain.verify()
+    for malformed in ([{}], [None], [{"prev_hash": "0" * 64}]):
+        chain.chain = malformed
+        assert not chain.verify()
+
+
+def test_cli_labels_fixture_and_does_not_overwrite(tmp_path):
+    destination = tmp_path / "receipt.json"
+    command = [sys.executable, "-B", "-m", "szl_quant_bench.harness",
+               "--output", str(destination)]
+    first = subprocess.run(command, capture_output=True, text=True, check=True)
+    result = json.loads(first.stdout)
+    assert result["chain_valid"] is True
+    assert result["provenance"]["source_kind"] == "SYNTHETIC"
+    before = destination.read_bytes()
+    second = subprocess.run(command, capture_output=True, text=True)
+    assert second.returncode != 0
+    assert destination.read_bytes() == before
+
+
+def test_cli_invalid_input_has_nonzero_exit_and_no_receipt(tmp_path):
+    source = tmp_path / "input.json"
+    source.write_text(json.dumps({"logits": [[True]],
+                                  "provenance": {"source_kind": "SYNTHETIC"}}))
+    result = subprocess.run([sys.executable, "-B", "-m", "szl_quant_bench.harness",
+                             "--input", str(source)], capture_output=True, text=True)
+    assert result.returncode == 2
+    body = json.loads(result.stdout)
+    assert body["state"] == "INVALID"
+    assert body["chain"] == []
+    assert body["chain_valid"] is False

szl_quant_bench-0.2.0/tests/test_quant_bench.py

@@ -0,0 +1,70 @@
+import math
+
+import pytest
+
+from szl_quant_bench.harness import run_curve
+from szl_quant_bench.metrics import cosine, kl_divergence, top1_agreement
+from szl_quant_bench.quant import dequantize, quantize
+from szl_quant_bench.receipts import ReceiptChain
+from random import Random
+
+
+def logits_fixture():
+    rng = Random(42)
+    return [[rng.gauss(0, 1) for _ in range(64)] for _ in range(16)]
+
+
+def test_roundtrip_16bit_near_exact():
+    vals = [0.1, -0.5, 3.7, 0.0, -2.2]
+    ints, scale = quantize(vals, 16)
+    deq = dequantize(ints, scale)
+    for a, b in zip(vals, deq):
+        assert abs(a - b) < 1e-3
+
+
+def test_2bit_coarser_than_8bit():
+    vals = [i * 0.01 for i in range(-50, 50)]
+    ints2, s2 = quantize(vals, 2)
+    ints8, s8 = quantize(vals, 8)
+    assert len(set(ints2)) <= 4
+    assert len(set(ints8)) > 50
+
+
+def test_quantize_rejects_bad_bits():
+    with pytest.raises(ValueError):
+        quantize([1.0], 1)
+
+
+def test_kl_nonnegative_and_zero_when_identical():
+    row = [0.1, 0.9, -0.4, 2.0]
+    assert kl_divergence(row, row) == pytest.approx(0.0, abs=1e-9)
+    assert kl_divergence(row, [2.0, -0.4, 0.9, 0.1]) > 0
+
+
+def test_top1_agreement_perfect_when_identical():
+    rows = [[1, 2, 3], [3, 2, 1]]
+    assert top1_agreement(rows, rows) == 1.0
+
+
+def test_curve_measured_and_quality_drops_at_low_bits():
+    out = run_curve(logits_fixture())
+    assert out["state"] == "MEASURED"
+    curve = {r["bits"]: r for r in out["curve"]}
+    assert curve[16]["cosine"] > 0.999
+    assert curve[16]["top1_agreement"] == 1.0
+    assert curve[2]["cosine"] < curve[8]["cosine"]
+    assert curve[2]["kl_mean"] > curve[8]["kl_mean"]
+    assert curve[4]["compression_vs_fp32"] == 8.0
+
+
+def test_curve_blocked_on_empty():
+    assert run_curve([])["state"] == "BLOCKED"
+    assert run_curve([[]])["state"] == "BLOCKED"
+
+
+def test_curve_receipt_chain_verifies():
+    chain = ReceiptChain()
+    out = run_curve(logits_fixture(), chain=chain)
+    assert "receipt" in out and chain.verify()
+    chain.chain[0]["run"]["curve"][0]["cosine"] = 0.0
+    assert not chain.verify()

szl_quant_bench-0.2.0/tests/test_receipt_required.py

@@ -0,0 +1,20 @@
+from szl_quant_bench.harness import run_curve
+from szl_quant_bench.receipts import ReceiptChain
+
+
+def test_every_measured_api_result_carries_a_verifiable_receipt():
+    out = run_curve([[1.0, 2.0]], provenance={"source_kind": "SYNTHETIC"})
+    assert out["state"] == "MEASURED"
+    receipt = out["receipt"]
+    assert receipt["prev_hash"] == "0" * 64
+    chain = ReceiptChain()
+    chain.chain = [receipt]
+    assert chain.verify()
+
+
+def test_caller_chain_still_receives_the_measured_record():
+    chain = ReceiptChain()
+    out = run_curve([[1.0, 2.0]], chain=chain,
+                    provenance={"source_kind": "SYNTHETIC"})
+    assert out["receipt"] is chain.chain[-1]
+    assert chain.verify()