This repository is the definitive public evidence and analysis artifact for the Cross-Vendor Semantic Void Matrix. It publishes the complete analysis, complete raw evidence, cryptographic inventories, provenance, and verification tooling for the finished 31,430-trial experiment.
Browse the complete analysis directory
Download the complete raw evidence archive — 170 MB compressed, 1.27 GB extracted
The raw evidence archive contains all raw attempt records, retries, failed attempts, event-chain records, schedules, frozen configuration, provider inventory, preflight records, reconstructed requests, provider identifiers, termination metadata, and retained reasoning metadata. It contains 31,523 public files, including 31,484 verified raw records and 62,968 replayed event hashes. No scientific outcome category was removed.
For byte-reproducible packaging, all ZIP member modification timestamps are normalized to 1980-01-01, the earliest conventional ZIP timestamp. These are archive metadata, not experiment timestamps; authoritative times are preserved inside the event records, manifests, raw responses, Git history, and release metadata.
Download the deterministic analysis ZIP
Open the complete GitHub release
python3 scripts/scan_for_secrets.py .
python3 scripts/verify_release.py --root . --assets-dir release_assets --deep-evidenceThe runner is separately frozen at theonlypal/void-matrix:
- Tag:
v1.0.0-definitive-runner - Commit:
2dfe895bdafd35d23d6dab5f307ed560214a3843 - Source state:
SOURCE_SEALED
Complete result: 11,658 total Voids = V0 + V1 + V2 + VU.
A Void is a model execution returning a successful provider response with exactly zero visible UTF-8 output bytes. Provider termination metadata determines its subtype. Explicit refusals, safety blocks, tool-mediated executions, and transport, protocol, billing, quota, rate-limit, and infrastructure failures are distinct non-Void outcomes.
| Class | Required observation |
|---|---|
V0 |
Successful provider response; text container present; zero visible UTF-8 bytes; recognized normal stop |
V1 |
Successful provider response; zero visible UTF-8 bytes; recognized output-budget stop |
V2 |
Successful provider response; text container absent; zero visible UTF-8 bytes; recognized normal stop |
VU |
Successful provider response; zero visible UTF-8 bytes; termination state not mapped by the frozen protocol |
All four classes retain Void status. Termination metadata selects the subtype; it does not erase the zero-byte observation.
Near-Voids, including whitespace, invisible Unicode, punctuation-only output, ellipses, and every other nonempty output, are non-Voids.
V0+V2 are the normal-stop Void subtypes. Their presence shows that recognized output-budget termination does not explain all Voids.
The frozen matrix completed 31,430 of 31,430 scheduled trials across 11 models from 4 providers. Offline verification replayed 62,968 event hashes and verified 31,484 raw attempt records, with zero pending trials, incomplete attempts, orphan raw responses, event-chain errors, classifier disagreements, or raw-audit anomalies.
The exact model set was:
gpt-4-0613gpt-5.2-2025-12-11gpt-5.5-2026-04-23gpt-5.6-lunagpt-5.6-solgpt-5.6-terraclaude-opus-4-6claude-fable-5claude-opus-5gemini-3.5-flashkimi-k3
The authoritative trial-level record is analysis/tables/trial_level.csv; the complete scope ledger is analysis/coverage_ledger.csv.
Across all final trials, the analysis classified:
- 11,658 total Voids
- 9,093 normal-stop Voids (
V0orV2) - 2,565 output-budget-stop Voids (
V1) - 0 unmapped-termination Voids (
VU)
These totals are generated from analysis/tables/trial_level.csv, reconciled in analysis/outcome_flow.csv, and summarized in analysis/normalized_headline_snapshot.json.
The primary matched comparison contains 4,290 null/control pairs:
| Outcome | Null conditions | Matched controls |
|---|---|---|
| All Void subtypes | 2,505 / 4,290 | 0 / 4,290 |
| Normal-stop Voids | 2,504 / 4,290 | 0 / 4,290 |
Every model had 0 / 390 strict matched-control Voids of every subtype. The pair-level source is analysis/matched_controls.csv; model-level profiles are in analysis/model_profiles/.
Strict matched controls are not interchangeable with the wider set of output-licensed controls. Across the pooled control set, 823 / 12,340 trials were Voids of any subtype and 3 / 12,340 were normal-stop Voids. Those control outcomes are retained, reported, and not removed from the dataset.
Gemini 3.5 Flash produced 69 / 1,250 Voids among its broader pooled controls, including all three pooled-control normal-stop Voids, while producing 0 / 390 strict matched-control Voids. The complete distinction is documented in analysis/contradictory_findings.csv and analysis/top_10_complicating_findings.md.
The zero-byte observation appeared through provider-specific response structures:
- OpenAI models predominantly produced present empty text containers with recognized
stopmetadata (V0), while GPT-5.5 and GPT-5.6 variants also produced budget-stop Voids (V1). - Claude Opus 4.6 produced absent text containers under recognized
end_turnmetadata (V2); Claude Fable 5 and Claude Opus 5 split between present-empty normal stops (V0) and output-budget stops (V1). - Gemini 3.5 Flash produced absent-container normal-stop Voids (
V2) and output-budget-stop Voids (V1). - Kimi K3 produced present-empty normal stops (
V0) and output-budget stops (V1).
Counts for every model are in analysis/tables/by_model.csv. Concrete provider-response morphology is documented in analysis/provider_morphology_examples.md and analysis/tables/provider_morphology_examples.csv.
- Original evidence-tree checksum-list SHA-256:
309cd9a77cdd6a6bf4ff733ca446163d170cb65e4ebfaac70d34f5c293481cc9 - Generated-analysis manifest SHA-256:
1fc3e04733a49a2584f496dd1f449a8d0796af71dc3542348a615b21e3e238f5 - Evidence archive SHA-256:
04d79ddb0f961729369da8ae46c58f7293e42dfd278d2e9b703edc0cec6455d2 - Analysis archive SHA-256:
f9cca29af7c8be387e1658f508ff068db14e1db9ad77bdda09ac6d5b399b2076
See the release asset RELEASE_MANIFEST.json, SHA256SUMS.txt, evidence/EVIDENCE_MANIFEST.json, and evidence/EVIDENCE_SHA256SUMS.txt.
The release manifest is generated after the annotated tag exists. Its publication_content_commit_sha identifies the complete publication content, and its metadata_seal_commit_sha identifies the exact final commit resolved by the release tag.
The release verifier checks release checksums, the frozen runner identity, both archives, every manifest-controlled analysis file, all headline counts, the exact 11-model set, the strict matched-control result, and, under --deep-evidence, the complete frozen evidence verifier.
After extracting void-matrix-analysis-v1.0.0.zip, follow analysis/reproducibility_manifest.md and run:
cd analysis
./verify_analysis.sh
python3 reproduce_all.pyThe baseline results remain independently reproducible without optional robustness dependencies. The original and hardened environments, affected outputs, and cross-environment equivalence checks are documented in analysis/environment_comparison.md and analysis/environment_comparison.csv.