Harden binomial tails and complete scientific-result validation

.github/workflows/tests.yml

@@ -3,23 +3,33 @@ name: corrected-analysis-tests
 on:
   push:
   pull_request:
+  workflow_dispatch:
+
+permissions:
+  contents: read
 
 jobs:
   test:
     runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.10", "3.11", "3.12", "3.13"]
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-python@v5
         with:
-          python-version: "3.11"
+          python-version: ${{ matrix.python-version }}
           cache: pip
       - name: Install corrected package
-        run: python -m pip install -e .[test]
-      - name: Run exact and numerical regression tests
-        run: pytest -q
+        run: python -m pip install -e '.[test]'
+      - name: Verify complete checksum manifest
+        run: python scripts/verify_checksums.py
+      - name: Run exact, numerical, and corruption regression tests
+        run: python -m pytest -q
       - name: Verify historical root scripts
         run: python scripts/verify_historical_files.py
-      - name: Validate committed scientific results
+      - name: Validate every committed scientific result field
         run: python scripts/validate_committed_results.py
       - name: Reproduce decisive exact counterexample
         run: python examples/reproduce_counterexample.py
@@ -29,3 +39,5 @@ jobs:
         run: python examples/reproduce_figure5.py
       - name: Check minimal four-qubit circuit
         run: python examples/simulate_minimal_four_qubit_circuit.py
+      - name: Confirm validation did not modify tracked files
+        run: git diff --exit-code

CHANGELOG.md

@@ -1,5 +1,14 @@
 # Changelog
 
+## 1.1.1-author-correction - 2026-09-07
+
+- Replaced overflow-prone binomial evaluation with scaled recurrence and log-tail APIs; positive underflow raises explicitly.
+- Expanded result validation to all five result files and 458 fields, including schemas, archived Monte Carlo metadata, and Wilson arithmetic.
+- Added tests corrupting every result field, checksum integrity tests, and exact rational statistical references.
+- Added complete manifest verification and Python 3.10-3.13 CI coverage.
+- Kept the August correction PDF, mathematics, results, historical scripts, and original release unchanged.
+- Fast checks explicitly distinguish archived Monte Carlo records from fresh simulation.
+
 ## 1.1.0-author-correction - 2026-08-17
 
 ### Referee-level revision

CHECKSUMS.sha256

@@ -1,14 +1,14 @@
-f6d1edcdabe82a8c9a6b8676470dca4755bd72abc667f9a542c0996794c8cef0  ./.github/workflows/tests.yml
+3f8c91e32f752f38c51834eeb0a38abfd791921d0be7e247ac840bb48a93ee26  ./.github/workflows/tests.yml
 f0e43839ec4fe3f086a0997f1aa238429543c8784776b05e019a719629ee1d90  ./.gitignore
 03353aa4d79773163b3c22bacc52f368149ef0b37abf8858a904169ebc7455c0  ./AI_ASSISTED_REANALYSIS.md
-73e68d44953e332f88f65ac908a872fad3fe8e39473c5c6fbbcc470af25e31d7  ./CHANGELOG.md
-7aaede28f7b53fdfa7f997bdb208fadb1c2339f3908812edbe9553003446d1d0  ./CITATION.cff
+e9e71e870915420013b31dd5f23d5c1c41ba87cbdca3eb4f752a50a91c5cb747  ./CHANGELOG.md
+ca826a2121e718472fff0f5d148f246467a519a7e4aa1d52bebe0ca7845a2cc6  ./CITATION.cff
 0c4959711013692777b7e98a35a0c8aa9e6fb85dc2583d617b303ed0cf3b0bfb  ./IBM.py
 aec03ca89a1ad1db2feba926dc30a6414e297a245987b3a77fdfa0af5efa24dc  ./ORIGINAL_2022_CODE.md
-8b2ce91ce9eb49e17cb281a9231aadbea1194c7cf6b71429631be4f156b6b430  ./README.md
-0fa3c0c2a8785e07f02f6898180e58e7e3abe2b1a1bbffed2791c5c4f3a2c9da  ./REPRODUCIBILITY.md
+fcf40b8f974ba65a7ec932643886a951d0860c10656934fec757ffb03ded0261  ./README.md
+7ad37e80a44290ad50c57c1aaa8479e52da51ccb952eea20cda8ee53ad1eb065  ./REPRODUCIBILITY.md
 52a7415b170f3ce96863c792633132fdfc2cf69fbd78bddd7eff38c5553b0b7d  ./SCIENTIFIC_STATUS.md
-e949ae83ccc8224a136145832fed7d91a161a8e2cb13853e69225812498a5baf  ./VALIDATION_RECORD.md
+c432acd797b3f422af4a1101c90d4b6c8b5cce94a05492cc8e14fc142a1a7bd8  ./VALIDATION_RECORD.md
 5a9b4e0129e44dbb52d368f9939a12cdedbd58e07d8e7bcd0945555377c8008d  ./benchmarking.py
 98b3c914732d1ff2b626e41de64619f058463956b997a7fad0eac6d41afb831d  ./circuit.py
 b368fa72d84696531f234263a75b4f7fa4de528885f91cd205a679d7e61682d3  ./compare.py
@@ -23,7 +23,7 @@ ccf0d83d58cf3645d17d8eef9b928471ee07958c1831a2a2fe47375dad7f4b2d  ./examples/rep
 e9d3750e58d51dff274a505f82f24fa9f3b4540de3e696053443868f28c12f83  ./examples/reproduce_one_sample_optimality.py
 b0e02e239a277b63f7a4cd64182aea155052359e708d095a9be30e90bb872bdd  ./examples/simulate_minimal_four_qubit_circuit.py
 657a73bbc6425559834b35866ae103f8fe57f01417fd37ef7876fa0de6da01ed  ./proba.py
-0eea87da078b39bde7ebf921849fc4d0bd8d87f3a9062f1f0e13b730a03a4c7a  ./pyproject.toml
+a843ca2838df028e56d6eb19dfcdb9c3e8712f11a74fb3bf97e8998ea1bca729  ./pyproject.toml
 ece8e18e5ab0401e138785179855e47b4ac46560c4e578f6e1a5d44c45eb539f  ./requirements-legacy.txt
 06c182a1b9b0a31420d26fba2d45a224dc71a1c84bcc1fd55784830b89c9164f  ./requirements.txt
 c4f52c8bf20bb65d0eba320aa1aef5d4b889e4c438156a42fc47307f251965f1  ./results/README.md
@@ -32,21 +32,28 @@ fac5d6248ca774cb4d3fe506518bb36594bbdbaf31dc0c46aca0a0f3524a576a  ./results/figu
 6b366e9a1765ffa38e259c78729c80855e6162dcfa6fa5b258b10e18951b72e3  ./results/heralded_witness_results.json
 02cb0b5ee8b241a556ce8c35a13829555aeeea14a0275f0bf869c06c2f58eeb6  ./results/one_sample_optimality.json
 da1935b2bd325cb523f313fa19a02270aa59e6acbcda00579a73a422b978294c  ./results/validation_crosschecks.json
-dec1ba0d29bc57db7507220340a34cf5ebe626212e8a670d80819604aee83e2f  ./scripts/generate_results.py
-3dfe8dc79a664ee9e8c11ae658981467fb54b8857080a8c66985e4d643fb1d2f  ./scripts/validate_committed_results.py
+1df70b11fc9a97b13aa9a91415508b0a0e40deaec8dc0f4fbfac0dd55125b458  ./scripts/generate_results.py
+7f75e8208141a8f5e17418991900d55599dea78d123e5fe86687046d73e1c6f1  ./scripts/validate_committed_results.py
+3b361b0be66dcda83398b443a4b39128548d5e9b3e2cb5623e2247cf1018ce70  ./scripts/verify_checksums.py
 636c2fe2c246e6806e21982e1696675f4dadd295d46d00269bce26f7dafac211  ./scripts/verify_historical_files.py
-134afb8cb64247983777b2698a5485686e88a2591936b55e3afb5945f05928d0  ./src/dcp_challenge/__init__.py
+86b2a62c29e84eea3bf15f6476f81883a7ccf43952e65261d96535aadc414e3a  ./src/dcp_challenge/__init__.py
 f520a4c550afa95ca11cf5a9b3cf656e4b0433c99b6f84a4088877d597062a35  ./src/dcp_challenge/exact_probabilities.py
 0e5b18c70d8c71ebf1ee90b37e8970dba76658ec6d10507dc2cf1b0e5f573f16  ./src/dcp_challenge/heralded_witness.py
+a6d6bfe51cc8d75a919487458e1ad622cd8ed5f2cb61559136d7318140e0ad13  ./src/dcp_challenge/integrity.py
 aa47265cc5fd3c4a2ca8f96b5b8c607ae35f74ec7db54d965c3dbfa7ea9e5f91  ./src/dcp_challenge/likelihood_decoder.py
 4de797931f4767f99b2bce6f09d46ea115e337180465287dec66fec062f02293  ./src/dcp_challenge/minimal_circuit.py
 adc35aed3c8c0d8d905cb35e2afc4fb9200057bb7567c81cdcae74a8b223e1e9  ./src/dcp_challenge/original_protocol.py
-d05eeb96dfb95747c49135ea9940ea31eb32e54db2d414723b91dbc9afd25f27  ./src/dcp_challenge/statistics.py
+a8858a11540d6b009184a0511e3e5bbda6d04626373c8f900233bbd44e5408b7  ./src/dcp_challenge/reference_results.py
+f92e394437cad537f44ef857f66b869bafa136fa66e0fb3f157d02e9c2731d13  ./src/dcp_challenge/result_validation.py
+bd90bfbd09791bc87c218ff1ad2fa00b765c4d738198b5b66d1577b1a3161981  ./src/dcp_challenge/statistics.py
 392a038abfb7ef5d14c445e11af42a942d97e75e322544ce16b045a9f372a5eb  ./tests/test_exact_probabilities.py
 61f2a3ef4c199710199f3dd378d85ba98e84ddaf5cca1017f7b9e0d86e1af36e  ./tests/test_heralded_witness.py
 862a3b357b0a2b9787274ee4186b060cbc7fe3d45c70ad510b46844138a98f12  ./tests/test_historical_files.py
+1e02e495cfa86b53854279bd644382c1aa631a11dec92d691e1bf3e9b1dacaae  ./tests/test_integrity.py
 d0202d343cb21e85e98b36241b2af4e261fd67c91603876a72f08404017b5f78  ./tests/test_likelihood_decoder.py
 ac3e495151c74a79eeb6a4cbcd135b6bddf7fd70e38ad6b1e3abd0bbb9a08075  ./tests/test_original_protocol.py
+32d3c06a6756f27efc13f6cc25f386f75ba4b7ae2638a7e21dae66b93f0249a3  ./tests/test_result_validation.py
+66b56beeb81973a82068e6c5f129e8a2f1f97d116ab90ef74167468174e9db54  ./tests/test_stable_statistics.py
 d482ef8a39d42dbfa8485abd79bb939699383928b544639abe5767697d83bd88  ./tests/test_statistics.py
 1bccb732feb1a5403164d875e304e13446ebc7162cf86634141b59b11dc3c33c  ./verification.py
 c4376bfd9dc74b1a1d781c062aab8d6a4c6150f89464e552138face53c1a654d  ./witness/CIRCUIT.md

CITATION.cff

@@ -5,8 +5,8 @@ message: >-
   corrected probability, soundness, or statistical analysis.
 title: "DCP Challenge: Author Correction and Follow-Up Analysis"
 type: software
-version: 1.1.0-author-correction
-date-released: 2026-08-17
+version: 1.1.1-author-correction
+date-released: 2026-09-07
 repository-code: "https://github.com/GoGoKo699/DCP-challenge"
 authors:
   - family-names: Lin

README.md

@@ -76,6 +76,7 @@ python examples/reproduce_counterexample.py
 python examples/reproduce_one_sample_optimality.py
 python examples/reproduce_figure5.py
 python scripts/verify_historical_files.py
+python scripts/verify_checksums.py
 python scripts/validate_committed_results.py
 pytest -q
 ```
@@ -89,6 +90,14 @@ Same measurements, better decoder: 25/32 = 0.781250
 
 See [REPRODUCIBILITY.md](REPRODUCIBILITY.md) for fixed seeds, exact-versus-numerical distinctions, and validation tolerances. The completed checks are summarized in [VALIDATION_RECORD.md](VALIDATION_RECORD.md).
 
+### Numerical maintenance - 7 September 2026
+
+Software version 1.1.1 adds stable binomial and log-probability calculations,
+validation of every committed result field, deliberate-corruption tests, and
+complete checksum verification. The August scientific correction, PDF, and
+historical scripts are unchanged. See the [reproducibility guide](REPRODUCIBILITY.md)
+for the distinction between recomputation and archived Monte Carlo checks.
+
 ## Separate follow-up: phase-twirled heralded DCP witness
 
 The directory [witness/](witness/) contains a new 2026 follow-up. It is not part of the 2022 article and is not a general proof of quantum computation.

REPRODUCIBILITY.md

@@ -117,22 +117,84 @@ The tests cover:
 
 ```bash
 python scripts/validate_committed_results.py
+python scripts/verify_checksums.py
 ```
 
-Validation policy:
+The result validator checks all five JSON/CSV records, every field, duplicate
+representations in the core JSON, schemas, and finite-number requirements.
+It recomputes exact counterexamples, one-sample records, all Figure 5 analytical
+probabilities, the exact Figure 5(a) likelihood result, witness summaries,
+dense-matrix diagnostics, the 20,000-product-effect diagnostic, and tail examples.
+Expected values are built without reading the files under validation.
+
+Integer and rational fields must agree exactly. Ordinary nonzero floating
+quantities use relative tolerance `1e-12` and zero absolute tolerance. Only the
+dense-matrix marginal residual permits absolute tolerance `1e-12`; this is not
+used to round small probabilities to zero. Byte-identical floating-point
+portability is not claimed. Validation uses explicit exceptions, so it also
+runs under `python -O`.
+
+**Monte Carlo boundary:** fast validation does not repeat the million-trial
+Figure 5(b) or the 60,000-trial Figure 5(c) experiment. It checks archived success
+counts, seeds, trial counts, estimates, and recomputed Wilson intervals, including
+the independent-seed cross-check records. These are archived evidence, not fresh
+simulation results. Deliberate-corruption tests alter each of the 458 result
+fields and require rejection, including changes made to duplicate records together.
+
+The checksum verifier additionally checks all tracked file bytes and complete
+manifest coverage. For an extracted source ZIP without Git, it checks the file
+tree excluding recognized local caches/build directories. Checksums detect
+snapshot changes, not scientific correctness or authenticity.
+
+To regenerate outputs without overwriting the archived files:
 
-- exact rational, integer, and combinatorial quantities are compared exactly;
-- fixed Monte Carlo records include seeds, trial counts, estimates, and Wilson intervals;
-- floating-point linear-algebra diagnostics are compared within the numerical tolerances encoded by the tests;
-- byte-identical portability of floating-point JSON across all numerical libraries and processors is not claimed.
+```bash
+python scripts/generate_results.py --output /tmp/dcp-reference-results
+```
 
-To regenerate the quick committed outputs:
+To explicitly rerun the larger Figure 5 simulations, use:
 
 ```bash
-python scripts/generate_results.py
+python scripts/generate_results.py --monte-carlo --output /tmp/dcp-fresh-monte-carlo
+```
+
+The latter still retains the separate historical independent-seed cross-check
+records. Fresh Monte Carlo output need not be byte-identical across numerical
+environments; it must not silently replace the archived record or be confused
+with the fast validator. Regeneration may change harmless last-bit diagnostics.
+
+## 8. Stable statistical evaluation
+
+The binomial routines use a scaled probability-mass recurrence and logarithms
+instead of multiplying huge coefficients by tiny powers. They require no new
+runtime dependency. For example, `binomial_lower_tail(2000, 400)` evaluates to
+approximately `7.183777008e-8` instead of overflowing.
+
+```python
+from dcp_challenge.statistics import (
+    binomial_lower_tail,
+    binomial_log_lower_tail,
+    adaptive_sequential_log_lower_tail_bound,
+)
+
+probability = binomial_lower_tail(2000, 400)
+log_probability = binomial_log_lower_tail(10000, 0)
+# log_probability is approximately -2876.8207245, not negative infinity.
+log_bound = adaptive_sequential_log_lower_tail_bound(10000, 0)
 ```
 
-## 8. Historical software
+All logarithms are natural. When a positive tail would underflow to zero, the
+ordinary probability function raises `FloatingPointError` directing callers to
+the log API. A zero return is reserved for an exact-zero endpoint. Counts must
+be integers and probabilities must be finite and in `[0, 1]`.
+
+These are floating-point evaluations of the analytical bound, not certified
+interval-arithmetic enclosures. They do not validate physical trust assumptions,
+experimental stopping rules, or repeated selective restarts. Work and integer
+storage increase with experiment size. Tests include exact rational references,
+endpoint checks, tail monotonicity, and the previously overflowing cases.
+
+## 9. Historical software
 
 The original scripts require their 2022 Qibo environment and are not imported by the corrected test suite. `requirements-legacy.txt` records the historical dependency request. The proof-of-concept IBM experiment is preserved but is outside the present correction.
 

VALIDATION_RECORD.md

@@ -1,5 +1,18 @@
 # Validation Record
 
+## Software hardening - 7 September 2026
+
+- Baseline tree: `93caf96908663f256de153feee54afccb358d342`, matching published `main` at `449548426b83fc22a9939fb2c2f6dede74dcd6f4`.
+- Local suite: **59 tests passed**, including 458 individual result-field mutations.
+- Previously overflowing binomial cases now agree with independent exact rational sums.
+- Positive probability underflow is explicit; finite log tails remain available.
+- All five committed result records and complete SHA-256 manifest coverage validate.
+- Historical scripts and the August scientific documents/PDF are unchanged.
+- The fast validator reuses archived large Monte Carlo counts and recomputes their arithmetic; it does not claim a fresh large simulation.
+- The automated workflow repeats tests and integrity checks on Python 3.10, 3.11, 3.12, and 3.13. Its live execution status is recorded in GitHub Actions, not inferred from this document.
+
+## August scientific validation
+
 **Validation date:** 17 August 2026  
 **Target:** author-correction repository state
 

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "dcp-challenge-correction"
-version = "1.1.0"
+version = "1.1.1"
 description = "Corrected analytical tools and follow-up checks for the 2022 DCP challenge"
 readme = "README.md"
 requires-python = ">=3.10"