Live on vllm-voices (fv-ml1 GPU0 :8027) beside voices-base, lv-yarros and lv-bronte.
Healthy 190 s after recreate, four models served, GPU0 96,092 -> 96,090 MiB. The adapter
was verified byte-identical to checkpoint-850 by sha256 across both transfer hops, and the
seat was verified by generating, not by reading its config: base emits 170 words of <think>
planning and never writes the passage, lv-hemingway writes the scene.
Gate design was pre-registered before any generation existed (0bb4938). Three arms, 60
held-out beats, 4 seeds, 240 generations per arm.
A. VOICE PASS 6.4x +0.413 delta_cb, pairwise floor 0.064 -- and it clears the OLD
all-arms floor (0.113) too, so this verdict does not lean on the
rule change. Closes 73.8% of the span between the unadapted
carrier and held-out Hemingway itself; lv-bronte closed 48%.
B. NOT COPIED see below
C. NO DAMAGE PASS ran-on +0.08, on-beat -0.14, both inside a 0.217 floor
AXIS B: THE NEGATIVE CONTROL WAS THE WRONG ONE, AND FIXING IT MADE THE RESULT WORSE, NOT
BETTER. memorization_check.py uses the base-unadapted arm as its control. Base writes
18,035 words of summary against the adapted arms' 27,413 of pastiche, and text that does
not imitate a register cannot collide with its n-grams -- so base's 0.00 measures "different
register", not "did not memorise". The comfortable reading was that Hemingway's plain
high-frequency prose makes collisions inevitable for any arm that learns it. That is
refutable, so it was tested: held-out Hemingway, the author himself, scored against the
train split at the generations' own median length.
HELD-OUT HEMINGWAY (never trained) 370 chunks 0.01 hit-rate mean-longest 0.1 max 10
base-unadapted 240 gens 0.00 0.0 0
ckpt850 (shipped) 240 gens 0.07 0.6 9
positive control (train vs train) 160
The hypothesis is false: the adapter reproduces train n-grams ~7x more often than the
author reproduces himself. That is real and is on the record. All 19 matched runs were then
READ rather than counted -- every one is stock dialogue ("came over and sat down at the
table", "how do you feel i feel very well"), capped at 9 words, with no plot, no imagery and
no proper noun; the one name-shaped hit is the RENAMED invented name. Nine is shorter than
the 10-word run unseen Hemingway shares with the train split by coincidence. Elevated rate,
zero protectable content. Hemingway is in copyright; lv-yarros is the in-line precedent,
also in copyright, shipped at 0.10 against a 0.07 control. Unload is 0.003 s.
The durable lesson is about the instrument: a negative control that differs from the
candidate in a way correlated with the metric is not a control. memorization_selfsim.py and
memorization_dump_matches.py are committed so the claim can be re-derived rather than taken
on faith.
SHIPPED ckpt850, NOT the loss minimum at step 1750. The two are indistinguishable on voice
-- 0.072 apart against a 0.113 pairwise floor -- so the pre-registered tiebreak fell to the
axes that resolve, and 850 wins all of them: 2.3x tighter seed spread (0.050 vs 0.113),
lower memorisation, less ran-on, half an epoch less overfit. ckpt1750's spread is one seed
(0.491, 0.449, 0.468, then 0.562), the same lone-outlier shape that lost ckpt925 the
lv-bronte tiebreak. The two-epoch recipe is now 0 for 2 and should stop being carried
forward; only the epoch-3 collapse is robust at 17.4x jitter.
servers/fv-ml1/ssh-target was a bare IP, so deploy-stack.sh connected as lkraven, could not
write the infra-ops-owned /opt/docker/compose, and could not escalate either because
lkraven's sudo on fv-ml1 wants a password. Now infra-ops@10.251.50.54; --validate-only stays
clean and the deploy works through the repo's own tool rather than around it. Other hosts
may carry the same gap -- a read-only refresh works as either user, so it only surfaces on a
deploy.
The gate design is written before any generation exists, because lv-bronte's
verdict turned on a choice that was only visible after the numbers printed.
THE FLOOR RULE IS NOW PAIRWISE. lv-bronte computed the noise floor as the largest
within-arm seed spread across ALL arms present. Its ckpt475 shipped at +0.193
against a 0.251 floor set entirely by ckpt925 -- a third arm nobody was shipping,
on one outlier seed. Scored against the arm it was actually compared to, the floor
is 0.092 and the same gap clears at 2.1x. A candidate's verdict must not depend on
which other arms happened to be generated. voice_distance.py now prints both floors
and flags any disagreement, so the lv-bronte record stays comparable.
audit_pairs_sourcenames.py closes the blind spot leak_gate.py has by construction:
it reads the corpus and the renamed copies, never the generated beats, so it cannot
see a beat-writing model restoring the author's real character names. Run over the
Hemingway pairs, which predate build_sft_pairs.py --source-entities:
val 0 of 200 -- the eval fixture is clean, the gate is unconfounded
train 70 of 7,094 (0.96%) -- Santiago x16, Catherine x7, Rinaldi x3, Brett,
Harry, Jake, Pablo, Nick, Maria ...
responses 0 of 7,294 -- the lv-bronte beat-only signature exactly
A matched surface is only counted when the rename actually removed it, verified
against the renamed copies, so a beat naming a held real-world place is not a leak.
Controls run every time: 941/941 surfaces found in the unrenamed source, nonce
absent from both trees, and 6 planted canonical names detected 6/6.
voice_distance.py --author is now REQUIRED. It was hardcoded "Yarros" and printed
"reference: held-out Yarros" over Brontë's numbers into a committed artifact. A
default would have moved the silent-wrong-label failure rather than removed it. The
stale "one seed-pair per arm / corroborates Base < Instruct" footer is replaced with
what the run actually carries.
Gate design: three arms (base-unadapted, ckpt1750, ckpt850), 60 beats, 4 seeds.
ckpt850 is present because the loss curve cannot separate it from ckpt1750 -- +0.0040
against a 0.0044 median neighbour jitter, with three checkpoints inside one jitter of
the minimum. adapter/ is excluded: +0.0762 is 17.4x the jitter and is resolved without
a gate.
vh
Vuong Hoang