06-reference/research

fastino glide vs jev contamination resistant evidence

2026-10-27·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
competitordecision-modelsfastinotypesafebenchmark-hygiene

Fastino is real, GLiDE 1.2's 71.5 reproduces from the index's own dataset, and it is cheaper than Jev - but the public-versus-private gap turns out to be a model-size artifact, not a Cloudflare tell

The question

"Who is Fastino, and does GLiDE 1.2 hold up as a decision-model alternative to Jev on contamination-resistant evidence?"

Asked because [[2026-10-24-cloudflare-clef-clairic-vs-typesafe-jev]] went hunting a Cloudflare threat and found an unfamiliar name at the top of the independent board instead. Its own closing line: the accuracy leader "is Fastino GLiDE 1.2 at 71.5 and 93ms, eleven points above Jev, and RDCO has never looked at it." A folded-in second scope asks whether the public-versus-private score gap that brief read as Clef's contamination signature is general across all 228 reproductions.

Method and evidence labels

Same four labels as the parent brief: (vendor) for a vendor's own claim, (vendor-measured) for a performance number the vendor measured, (independent) for a party selling none of these models, (U) for unverified or unpublished.

One methodological upgrade matters. Rather than reading the Decision Index's rendered leaderboard, this pass downloaded the index's own published data files (data/index.json, data/own.json, data/methodology.json, generated 2026-10-26T15:30:58Z) and computed the statistics over all 229 entries directly. Every board number below is read off that dataset, not off a screenshot, a tooltip, or a press summary. The score-field mapping was pinned by reconciliation: public_skill reproduces the parent brief's Clef 61.7 and balanced_skill reproduces its Clef 54.9, which confirms public_skill is the PUBLIC ONLY toggle and balanced_skill is the FULL SCORE private column.

What we already know (from the vault)

What the web says

The 93ms is a median on a datacenter card, and the tail is the real story

The parent brief carried "71.5 and 93ms" as a pair. The first number survives scrutiny. The second needs a correction that changes its meaning.

Source Who Median p95 Mean Rig
Decision Index independent 92.9ms 1,121.1ms 284ms 1x NVIDIA RTX PRO 6000, local process
Fastino model card vendor-measured 84ms (RTX PRO 6000), 48ms (B200), 78ms (H100) p99 2,709ms (RTX PRO 6000), 1,086ms (B200) not stated local, first 300 requests
Jev 1.14, same index independent 196ms 296.8ms not stated hosted HTTPS round trip from the lab

Three readings follow, and only the first is the one in circulation.

  1. On the median, GLiDE is about twice as fast as Jev. 92.9ms against 196ms.
  2. On the 95th percentile, Jev is about 3.8x faster. 297ms against 1,121ms. GLiDE's mean of 284ms exceeds Jev's median, which is the signature of a heavy right tail rather than a slow model.
  3. Neither comparison is like-for-like, and the index says so. GLiDE runs in-process on a graphics card; Jev is measured over the public internet, with the index's own note reading "Hosted API, network round-trip from our lab... Not comparable to the on-card single-process figures." Fastino's own card adds a 25-35s cold start and a 30-60s weight load, and reports a single 39k-token question taking about 13s warm.

The tail is not an accident, and it is not GLiDE-specific. The index entry GLiDE 1.2 replaced is named decision-27b-ckpt-nothink, and every one of the top ten models on the private column carries a p95 between 674ms and 2,245ms while Jev's hosted p95 sits at 297ms. The models that win the contamination-resistant column are adaptive-compute models that think longer when uncertain, and they pay for it in tail latency. For RDCO's /ideas batch judgments that is irrelevant. For anything on a request path it is the whole question, and it is a trade the "93ms" headline hides.

The gap question: general, but it is a size effect, not a category-wide contamination tell

This is the scope that got folded in, and it comes back with a more useful answer than expected. Measuring public_skill - balanced_skill across all 229 entries:

Statistic Value
Mean gap +4.13
Median gap +2.90
Standard deviation 6.29
Range -5.94 to +24.91
Positive (public flatters) 161 of 229 (70%)
Negative 68 of 229 (30%)
Gap above Clef's +6.79 63 of 229

So the first-order answer is yes, general: on 70% of the board the public column reads higher than the private one, by a median of nearly three points. But stratifying by parameter count changes what that means.

Size band n Mean gap Median gap Share positive
Under 2B 58 +7.24 +4.60 81%
2-6B 58 +5.71 +4.45 79%
6-15B 31 +4.89 +3.23 87%
15-40B 78 +0.44 -0.11 50%

Pearson correlation between log-parameters and gap: r = -0.417. By training method: full fine-tunes average +5.84 (n=127), LoRA adapters +2.42 (n=44), pure inference techniques -0.70 (n=26). The pattern is coherent: the entrants that lose the most ground on held-out tests are small models fully fine-tuned on the visible suite.

Three consequences, in order of how much they change RDCO's behaviour:

  1. At the size class where this decision actually lives, the gap is zero on average. In the 15-40B band that contains Jev, Clef, GLiDE and every serious contender, mean gap is +0.44 and only half the entries are positive at all. Jev's gap is +0.46 and GLiDE's is +0.47, both unremarkable for their class.
  2. Clef's +6.79 therefore survives as an outlier, but only once you control for size. It ranks 8th largest of the 78 entries in its own size band, roughly the 90th percentile, and clef-omni is the single worst in that band at +11.50. The parent brief's contamination read on Clef stands. Its implied generalization - that the gap is "the signature of public-suite contamination" as a category property - does not, because the category-wide gap is explained better by model size than by vendor behaviour.
  3. The honest reading rule is a conditional one, and it is cheap to apply. A published public-benchmark score for a sub-6B decision model should be discounted by roughly 5 to 7 points before it is compared against anything. A public score for a 27B-class model needs no size discount, so a large gap there is genuinely informative about that specific entrant. This is the reusable finding, and it is more durable than any single model ranking.

One caution this brief cannot resolve: a size-dependent gap is consistent with contamination and also consistent with small models simply generalizing worse. The index's own framing is contamination ("Anyone can see them, and train on them, so they no longer count toward the Full score"), and it has removed at least one entrant for provably training on test items (reflex-4b, dropped because its adapter "was trained on 800 MMLU-Pro test items"). But capacity and contamination predict the same correlation, and nothing in the published data separates them. Calling the gap a contamination signature is (U) as a mechanism, even though it is settled as a measurement.

Convergences and contradictions

Synthesis for RDCO

Fastino clears the reality bar comfortably, and GLiDE 1.2 clears the evidence bar on every axis this brief could measure. The company has a published rate card, a self-serve key flow, versioned docs with cookbooks, a licence file with a real revenue threshold, and institutional money behind it. The model is first of 229 on the index's contamination-resistant column by 1.3 points over the runner-up and 11.3 over Jev, it wins all five categories and both private halves rather than spiking on one, it is fourth of 229 on calibration where Jev is 127th, and it costs $0.04 per million input tokens against Jev's $0.042. The licence matters less than it looks: production commercial use needs a separate agreement only above $10M annual revenue, and RDCO is nowhere near that, so under [[2026-05-05-personal-license-boundary]] the weights are usable today and the threshold is a problem worth having. The practical catch is smaller than the licence: downloading the weights needs a token with read access to the Fastino organisation, so this is community-licensed-behind-a-token rather than Clef's unconditional Apache 2.0.

The evidence does not, however, settle that GLiDE beats Jev on RDCO's workload, and the brief should not pretend otherwise. Everything above rests on one leaderboard, one pull-request-submitted entry, and no published note that the index re-ran it. There is no independent replication of Fastino's run by a second party, no hosted-to-hosted latency comparison against Jev, no published free tier, rate limits or service-level agreement, and no training-data disclosure. The measured latency profile is genuinely worse than Jev's where it counts for anything interactive: GLiDE's p95 of 1,121ms against Jev's 297ms, with Fastino's own card reporting a p99 of 2,709ms and a 25-35 second cold start. "Rank 1 on a leaderboard at a 5% lower list price" is a strong reason to run a test. It is not a product recommendation, and the gap between those two things is exactly what the parent brief's pre-registered test exists to close.

The gap scope returns the most reusable result of the night, and it partly corrects the parent brief. The public-versus-private spread is general across the board - 70% of entries, median +2.9 points - but it is driven by model size, not by vendor conduct: mean +7.24 under 2B falling to +0.44 at 15-40B, with r = -0.417 against log-parameters and full fine-tunes averaging more than twice the gap of LoRA adapters. So the reading rule RDCO should adopt is conditional rather than blanket. Discount a published public-benchmark score for a small decision model by roughly five to seven points before comparing it to anything; apply no size discount at 27B class, which makes a large gap there informative about that entrant specifically. On that rule Clef's +6.79 is still an outlier for its class at the 90th percentile and the parent brief's contamination read on Cloudflare survives, while its wider implication that the whole category is contaminated does not. And the mechanism itself stays open: capacity limits and benchmark contamination predict the same size correlation, and the published data cannot separate them.

For the standing default, the recommendation is unchanged but the reasoning has to be rewritten, and the test just got cheaper to run. Jev should stay the default under [[feedback_jev_for_typed_set_problems]] for three reasons that are all still true: it is wired and working, its latency tail is the tightest of any serious contender, and no alternative has been measured on RDCO's own labelled sample. The reason that must be retired is the one the parent brief leaned on, that Jev leads the contamination-resistant evidence. It does not; it sits 42nd. What has changed materially is the cost of finding out. Fastino's endpoint is POST /v1/systemone taking state plus typed questions, exposing Noul, Choice and Score, which is the same shape TypeSafe serves at api.typesafe.ai/v1/systemone. The --provider {jev,clef,clef-flash,glide} flag the parent brief pre-registered is still a two-constant change plus a credential for GLiDE, and GLiDE should now be the first alternative arm run rather than the last, because it is the only one that leads Jev on held-out data, on calibration, and on price simultaneously. The acceptance bar from the parent brief holds verbatim: no published benchmark counts, including the Decision Index, and GLiDE replaces Jev only if it wins on agreement against RDCO's own withheld labels at acceptable measured latency from the Mac mini.

Why this is in the vault

[[2026-10-24-cloudflare-clef-clairic-vs-typesafe-jev]] kept Jev as RDCO's default for typed-set judgments on the explicit ground that Jev led the contamination-resistant evidence, and that ground is now measurably wrong (Jev is 42nd of 229 on the private column; 41 entries beat it). This brief supplies the corrected basis for the [[feedback_jev_for_typed_set_problems]] default, names GLiDE as the first arm the pre-registered jev_judge.py provider-flag test should run, and gives RDCO a size-conditional rule for reading every published decision-model benchmark from here on, replacing the blanket "discount the public column" heuristic the parent brief implied.

Open follow-ups

Related

Sources

Independent measurement (primary data, downloaded and computed over this pass)

Fastino (vendor-primary)

Secondary, carried as (U)

Vault paths cited