Company / Evidence
Every test result behind the claims on this site, with the gaps left in.
All suites were executed on 2 August 2026 and the numbers tallied from actual output, not from documentation. Where a suite crashed, where a repository has no tests at all, and where our own falsification harness reported a false negative, it is on this page — in the same type size as the passes.
The tally
Twenty-six suites were discoverable across the four repositories in scope. Twenty-five completed. The counts below are assertions from completed suites only — the 263 assertions that passed inside the run that later crashed are deliberately excluded, because a partial run is not a result.
One column per suite, height set by its own assertion count. The tower is voxel addressing at 741, exhaustive rather than sampled.
The severed stub is the integration run that crashed at section O after 263 assertions. Two hollow rings are repositories with no test suite at all — nothing stands on them.
Six markers above are the falsification probes. The amber one is the control that is required to diverge.
Suite 06 dominates this picture because it is exhaustive — 741 assertions covering every one of the 342 addresses in both directions. Impressive height, but it proves one narrow thing very thoroughly: that the addressing map is correct.
The severed column and the two empty footprints are the parts worth your attention, and they are drawn at the same scale as everything else on purpose. A chart that renders its gaps as invisible is not a chart of the evidence, it is a chart of the good news.
| Suites that ran to completion | 25 of 26 |
|---|---|
| Suites that crashed | 1 — harness fixture defect, detailed below |
| Skipped | 8 — require live external services |
| Repositories with no test suite | 2 of the 4 surveyed — Cubex³ and OrbitGen-AI |
| Scope of the survey | Four repositories, chosen as the ones these claims rest on. Other codebases exist and were not surveyed — their coverage is unknown and is not represented anywhere on this page |
| Method | Suites executed directly; results tallied from output. Nothing on this page is quoted from documentation |
Read this first
Placed before the results rather than after them. A reader who only takes one section from this page should take this one.
run_all_tests reached section O and aborted with a
SqliteError: FOREIGN KEY constraint failed at
services/domainLibrary.ts:46. The test inserts a domain template without
first creating the parent row the foreign key requires. That reads as a fixture defect
rather than a product defect — but it is unproven either way until fixed, and
every section after O never ran.
No test suites, no test script in the manifest. It compiles clean and produces a working production build — 2,781 modules in 10.6s — and that is a statement about the build, not about coverage. A green typecheck proves internal consistency and nothing else. Seven typecheck errors were repaired on 2 Aug 2026; that changed nothing about coverage.
The key system — identifier normalisation, reference rewriting, parent-link repair, format validation — is entirely untested. Its whole purpose is enforcing invariants a language model will otherwise violate, which makes this the highest-value gap in the portfolio and also the cheapest to close: the functions are pure and take plain objects.
Two of the four surveyed repositories have no automated tests. The 1,686 figure describes the two that do — and the survey itself is not the whole portfolio. Other codebases were not examined, so nothing here should be read as portfolio-wide coverage in either direction. We would rather draw that boundary ourselves than be corrected on it later.
Byte for Byte · falsification battery
Each probe is an attack on our own claim, run against a live local model. One of the six is required to diverge: if the harness cannot detect nondeterminism when nondeterminism is present, no other row in the table means anything. The full evidence bundle for every run below is downloadable — including the run that failed.
Cache eviction unloads the model between calls and demands the same bytes back. Interleave runs a competing generation in between to disturb server state. Replay across processes requires a fresh OS process to reproduce a previous one's output exactly. Control divergence switches the guard off and requires the output to differ — it exists to catch us, not to reassure anyone.
Every generation's full raw text and sha256 is in the bundle, along with OS, CPU,
Ollama version and model digests. Recompute sha256(utf8(fullText)) for any
generation and compare it against the recorded hash; diff any two fullText
fields a check asserts are equal. A single-byte difference in an asserted-equal pair
falsifies the claim on that machine.
| 2 Aug 2026 · 20:26 UTC | PASS — 6 of 6 held · bundle (JSON, 14 KB) · transcript |
|---|---|
| 2 Aug 2026 · 01:29 UTC | PASS — 6 of 6 held · bundle (JSON, 14 KB) · transcript |
| 1 Aug 2026 · 00:19 UTC | FALSIFIED — reported our own claim broken · bundle (JSON, 14 KB) · transcript |
The run we kept
On 1 August the battery reported FALSIFIED. Two of the three replay workers had died — memory pressure on a 15.4 GB laptop — and the harness scored a worker that crashed identically to a worker that produced different bytes. An environment failure was reported as a falsified claim.
The surviving worker's hash was cd3e99af6f55814b — the same hash the
passing runs produce. The claim held. The instrument did not.
The harness now separates the two: a falsified check exits 1, and an environment that could not run a check exits 2 and reports INCONCLUSIVE. That separation is load-bearing — a harness that reports a result it did not measure is worth nothing.
The bad bundle is kept in the package and published here rather than deleted. It is the only artefact on this site that demonstrates the battery is capable of returning a negative.
The suites
Two suites are worth naming individually, because the rest of this table is unremarkable and those two are not.
| 01 Pipeline profile invariants | 15 / 15 |
|---|---|
| 02 Watchdog coherence circuit breaker | 20 / 20 |
| 03 Arc-role prompt construction | 19 / 19 |
| 04 JSON repair — boundary cases | 19 / 19 |
| 05 Arc scoring mathematics | 27 / 27 |
| 06 Voxel lattice addressing (exhaustive) | 741 / 741 |
| 07 Tensor field mathematics | 24 / 24 |
| 08 Database schema integrity | 11 / 11 |
| 09 Seven-arc system invariants | 22 / 22 |
| 10 Tensor field integration (SQLite) | 13 / 13 |
| 11 Transmutation chain mathematics | 32 / 32 |
|---|---|
| 12 Alignment rescaling (parametric) | 52 / 52 |
| 13 Stopwords exhaustive | 96 / 96 |
| 14 Goal decomposer mechanics | 134 / 134 |
| 15 Capability detector | 78 / 78 |
| 16 Goal tracker · 1×3×7 tool execution | 53 / 53 |
| 17 Capability sandbox | 57 / 57 · 8 skipped |
| 18 Metabolism reinforcement | 25 / 25 |
| Engine suite · node/parent integrity | 174 / 174 |
| Reproducibility & validation ledger | 24 / 24 |
Every one of the 342 data positions is tested individually in both directions — position to coordinate, and coordinate back to position — plus portal guards, out-of-range refusal, depth boundaries and all seven layer boundaries. An off-by-one in an odometer is silent, so the suite is written to make silence impossible.
Editing a recorded hash in place breaks the chain loudly. Deleting a line breaks it loudly. Re-asserting an existing key with different bytes is an integrity error rather than a silent replacement. A torn trailing line from a crash mid-append is tolerated while earlier records still enforce. Unwitnessed bytes are checked but never become ground truth.
Provenance
The byte-for-byte claim is scoped to local inference on fixed hardware, so the hardware is part of the result. Every bundle carries this block, and a different machine is expected to produce different hashes — what it must not produce is different hashes on the same machine.
Contract tests need none of this: 50 of 50 pass against fake adapters in about a third of a second, with no model, no GPU and no network.
| OS | win32 10.0.26200 (x64) |
|---|---|
| CPU | AMD Ryzen 7 4800H with Radeon Graphics |
| Memory | 15.4 GB |
| Node | v24.18.0 |
| Serving | Ollama 0.32.5 · localhost:11434 |
| Model | qwen2.5-coder:7b · digest dae161e27b0e |
| Embedder | mxbai-embed-large · digest 468836162de7 |
| Seed | 42 · replay runs 3 |
It is a record of what was measured, on a stated date, on a stated machine. It is not a certification, not an independent audit, and not a claim that any output is correct — reproducibility and correctness are different properties, and only the first is tested here.
The next things to fix, in order of leverage: the run_all_tests fixture so
the integration run completes; tests for the OrbitGen key system, which are pure
functions and cheap; and a first suite for Cubex³, now that a green build makes the
absence of tests the only remaining gap on that repository.