{{ title }}

{{ s.value }}

{{ s.label }}

{{ s.note }}

SEQUENCING

{{ step.tag }} {{ step.state }}

{{ step.what }}

{{ step.detail }}

MODEL	BASE	TRAINED WEIGHTS	WHAT IT ADDS	STATUS
{{ m.name }}	{{ m.base }}	{{ m.weights }}	{{ m.adds }}	{{ m.status }}

THE FIVE CONDITIONS

WHAT EACH DATA DIRECTORY IS, AND IS NOT

PATH	N	ROLE	CORPUS?
{{ p.path }}	{{ p.n }}	{{ p.role }}	{{ p.corpus }}

A prompt set and a source pack together make an instrument. Three now exist, and each measurement names which one produced it. Numbers from two instruments are not a measurement of change — the file names are the trap here, because both noise-floor files carry a v2 that means source pack, not prompt set.

PROMPT SET	PACK	FAM.	DISTINCT	REPEATED	WHAT IT MEASURED	STATE
{{ i.set }}	{{ i.pack }}	{{ i.fam }}	{{ i.distinct }}	{{ i.repeated }}	{{ i.measured }}	{{ i.state }}

v3 keeps every v2 instruction byte-identical and swaps in each family's own document, asserted identical across all ten prompts of a kind before use. The tasks had to stay the same or nothing measured against v2 would inform anything about v3 — the change is more material, not different work. It is still a new instrument: new source pack and new prompt set, so its results share no baseline with v1 or v2 and need their own measurement before any reproducibility claim.

IDENTITY, NOT JUST STRUCTURE — src/verify_pass_provenance.py

{{ c.name }}

{{ c.what }}

Once more than one instrument exists, a v2 trace sitting in a v3 corpus passes every structural check — right families, right splits, intact provenance — and corrupts every number computed from it, because it answers a different prompt about a different document.

{{ c.value }}

{{ c.label }}

{{ c.note }}

KIND	FAM.	DOCS v2	DOCS v3	REAL SOURCE	STATE
{{ k.kind }}	{{ k.families }}	3 · shared	10 · distinct	not approved	not approved	SYNTHETIC

{{ r.host }}

{{ r.label }}

{{ r.meta }}

{{ r.state }}

THE NOISE FLOOR AGAINST THE PRE-REGISTERED BOUNDS — AMENDMENT 8

METRIC	RUN A	RUN B	|Δ|	BOUND	STATE
{{ f.metric }}	{{ f.a }}	{{ f.b }}	{{ f.delta }}	{{ f.bound }}	{{ f.state }}

Run-to-run variance alone breaches the critical tolerance on two metrics. A single-run int4-vs-FP8 comparison therefore cannot support a verdict: noise would produce a FAIL, or a spurious PASS, with no precision effect present. Thresholds are not revised — loosening them after seeing variance is exactly what pre-registration exists to prevent. The study design is what was insufficient.

THREE PROTOCOLS, NEVER MERGED — AMENDMENT 6

{{ p.path }}

{{ p.status }}

{{ p.protocol }}

{{ p.use }}

Three different measurements of the same weights, reported side by side. Merging them would average distinct protocols into a number describing none of them — and study_report.py sets arms beside each other in two columns rather than concatenating them.

{{ a.name }}

{{ a.state }}

{{ f.k }} {{ f.v }}

PRE-REGISTERED CRITERIA, WITH THE MEASURED CONTROL FLOOR BESIDE THEM

BAND	METRIC	BOUND	CONTROL FLOOR	FP8	STATE
{{ m.band }}	{{ m.metric }}	{{ m.bound }}	{{ m.floor }}	{{ m.fp8 }}	{{ m.state }}

WHAT THE 2026-08-17 RENTAL COST, IN ORDER

Compute capability 8.9 says the silicon can do FP8. It says nothing about whether the driver can run current vLLM, whose kernels are built against CUDA 13. Torch was fine throughout — uv resolved cu129 and CUDA 12.x minor-version compatibility covered it — so torch imported, allocated GPU tensors and named the device correctly. Only vLLM's own extension failed, and only at the last step.

ORDER OF OPERATIONS — CHEAPEST REJECTION FIRST

#	STEP	NEEDS	REJECTS
{{ o.n }}	{{ o.step }}	{{ o.needs }}	{{ o.rejects }}

Do not download the model until step 0.5 passes. Every failure in the 2026-08-17 rental was detectable before that point.

AND THEN, ON ANY MACHINE WITH NO GPU

{{ s.step }}

{{ s.title }}

{{ s.body }}

{{ s.note }}

A model disappeared out from under this study once already: the gateway kept advertising ollama/muse-glimmer-30b in /v1/models while returning HTTP 400 for it, and 52 calls failed. Worse than failing loudly would be a model being silently replaced — the run completes, the numbers look fine, and they describe different weights than the ones being compared. Both arms now preflight first, and a signature mismatch fails the run.

TEACHERS — configs/teachers/

{{ t.name }}

{{ t.license }}

{{ t.arch }}

Teach it for

{{ t.teach }}

{{ r.k }} {{ r.v }}

{{ t.warn }}

HOSTS — configs/hosts/

HOST	GPU	QUANT	FP8	EGRESS	ROLE
{{ h.name }} {{ h.url }}	{{ h.gpu }}	{{ h.quant }}	{{ h.fp8 }}	{{ h.egress }}	{{ h.role }}

PROBE ROUND 1 — 2026-08-17: 0 CANDIDATES, AND WHY THAT IS NOT AN ANSWER

RECORDED DECISIONS — PROMPT SET V1 EVIDENCE, VOLATILE_DECISIONS.yaml

LABEL MISMATCH — QUARANTINED, NOT DISCARDED

HOW A TRACE IS REJECTED BEFORE IT REACHES REVIEW

{{ r.name }}

{{ r.what }}

All three of the quality cases look like success at the HTTP layer. A batch that does not yield exactly one answer per item is rejected wholesale — a partial parse would misalign answers against prompts and silently mislabel training data.