The planted quote-style rate excludes the registered one: δ is carried at 1.614¶
# research(2026-10). The runs of the 2026-10-15 plan, read once each by planted_exchange.py
(planted-exchange-openstack-quotes1.0.json, -quotes0.5.json):
| dose | runs | planted held-out examples | own minus sibling, summed s | summed exact match | rate (nats per token per exact refinement) | 95% |
|---|---|---|---|---|---|---|
| f = 1.0 | 10 | 486 | 151.74 | 94 | 1.614 | [1.479, 1.756] |
| f = 0.5 | 5 | 125 | 25.94 | 11 | 2.358 | [1.791, 4.620] |
No f = 1.0 resample moved exact match by nothing; one of the 10,000 at f = 0.5 did, below the 1%
past which no interval is read. The f = 1.0 interval excludes the registered rate 1.834, so the
rule applies as written: δ is carried at the planted rate, 0.01 × 1.614 = 0.0161 nats per token
in place of 0.0183, a change made after the development pilots were read and disclosed as one in
the Stage 1 design history. likelihood_sesoi.py now applies the rule itself and records both
rates (likelihood-sesoi.json: rate_source planted, positive_control_rate 1.8337, which
reproduces the registered value).
The dose check does not reproduce the rate: the f = 0.5 interval lies wholly above 1.756, and it contains the registered rate 1.834 that the f = 1.0 interval excludes. It rests on 11 exact refinements against the f = 1.0 reading's 94, and a ratio of sums whose denominator is that small is skewed upward, which fits its long upper tail, but the check was registered to show the rate does not depend on the dose and it does not show that. It is reported as a failed check, not explained away: it says the rate at small differences between adapters, which is what δ measures, may be higher than the f = 1.0 reading. A higher rate means a larger δ in s, so the f = 1.0 rate is the smaller margin of the two. It is the stricter one for a bounded reading (a reading bounded below it is bounded below the other) and for the check that detectable effects lie below δ; for a meaningful reading, whose lower bound must clear δ, the smaller δ is the easier bar.
Every bound on δ is recomputed: the two development pilots on the registered metric
(partition-pilot-{openstack,wikimedia}-likelihood.json, rebuilt by their recorded inputs; at the
old SESOI they reproduce the committed files to 1e-12, Wikimedia's drift being the fp32 rescoring
of 2026-10-15), whose verdicts and within_sesoi readings do not change and whose sizing still
clamps K at 10; K coverage, which reads no SESOI and keeps K = 10 and K = 20; and the two full
simulations at those K, rerun at the new SESOI, whose results follow in this entry's PR. The
simulations K coverage reads besides those two (Wikimedia's at K = 10 and the null grid,
-k12-null to -k40-null) keep the 0.0183 they were run at; the rule reads only their null
false-positive rates, which do not depend on δ, so they are not rerun.
The two full simulations at the new δ (partition-sensitivity-openstack-rp1.0-logprob_per_token.json,
TIGRIS 293410, K = 10; partition-sensitivity-wikimedia-rp1.0-logprob_per_token-k20.json, TIGRIS
293411, K = 20; both pinned at 27eb4f3, the commit above before its rebase onto main, whose code
differs from the rebased commit only in files neither simulation imports; 1,000 trials and 4,000
null trials a point) reproduce
every detectable effect, null false-positive rate and bounded-null share of their 2026-10-10 runs
exactly, since none of them reads δ. The one reading δ moves is the share of null studies whose
interval lies inside ±δ, in nats per token:
| organization | spread point | sigma_run | detectable, 97.5% | detectable, 95% | null inside ±δ, 97.5% | null inside ±δ, 95% |
|---|---|---|---|---|---|---|
| OpenStack | pilot lower 90 | 0.0000 | 0.0102 to 0.0112 | 0.0094 to 0.0104 | 1 | 1 |
| OpenStack | pilot estimate | 0.0000 | 0.0102 to 0.0112 | 0.0094 to 0.0104 | 1 | 1 |
| OpenStack | sizing bound 90 | 0.0016 | 0.0103 to 0.0113 | 0.0094 to 0.0105 | 1 | 1 |
| OpenStack | pilot upper 99 | 0.0050 | 0.0118 to 0.0131 | 0.0110 to 0.0121 | 0.999 | 0.9995 |
| Wikimedia | pilot lower 90 | 0.0000 | 0.0037 to 0.0041 | 0.0034 to 0.0038 | 1 | 1 |
| Wikimedia | pilot estimate | 0.0037 | 0.0050 to 0.0054 | 0.0046 to 0.0050 | 1 | 1 |
| Wikimedia | sizing bound 90 | 0.0062 | 0.0067 to 0.0073 | 0.0062 to 0.0067 | 1 | 1 |
| Wikimedia | pilot upper 99 | 0.0085 | 0.0083 to 0.0091 | 0.0077 to 0.0084 | 1 | 1 |
Every detectable effect lies below δ = 0.0161 at every spread point, one to three cells and both levels, so δ still serves as the equivalence margin. The share inside ±δ falls below 1 at one point only, OpenStack's 99% upper spread (0.999 and 0.9995, from 1.0), which leaves the 2026-10-14 entry's statement that it is 1.0 everywhere true of Wikimedia and of OpenStack's other three points.