docs(refusal-probe): retire the DS v1 regeneration — unqueued, do not run

brokkr-smithy-dev withdrew the request on the operator's call (msg 01M088G7NQ5G42ES30YPJV4Y3V). Two reasons: the ictrl-pair-unwrapped / ictrl-pair-wrapped control isolates the classifier over-fire cleanly, where DS's cross-class delta only bounded it; and DS v2 is releasing soon, so a k=5 v1 baseline baselines a superseded version. The spec stays banked as the record of the run that will not happen -- axes, per-class grading asymmetry, and run parameters remain correct. Checklist struck through. The staged probe.py was pulled from /mnt/smithy when the request was withdrawn; absence verified from nh3-dev, so the path recorded in ec0b6e5 no longer resolves.
2026-08-17 09:23:24 -07:00
parent ec0b6e5e71
commit 9c1405b1f9
1 changed files with 44 additions and 28 deletions
@@ -1,9 +1,31 @@
 # Dark-Scarlett regeneration spec — the 8 dropped axes

-**Status: QUEUED, gated on the operator.** Running this means taking Fable-Fusion
-down on the `char-rp-reasoning` seat (GPU1 is zero-sum), and that seat is under
-active evaluation. No deadline — the DS weights are not scheduled for reuse
-(103 GB at `ana-ml2:/tank/aimodels/darkscarlett-nvfp4-work/`, /tank at 50%).
+**Status: RETIRED — NOT RUN, and not to be run. Unqueued 2026-08-17.**
+brokkr-smithy-dev withdrew the request on the operator's call
+(msg `01M088G7NQ5G42ES30YPJV4Y3V`); do **not** spend a GPU1 window on it when
+Fable-Fusion clears the `char-rp-reasoning` seat. Two reasons, the second fatal:
+
+1. **Low value-of-information.** The only thing the operational/meta half would
+   have informed is the classifier over-fire — a property of the *scorer*, not of
+   DS. Our `ictrl-pair-unwrapped` / `ictrl-pair-wrapped` pair isolates that
+   cleanly, where DS's cross-class delta only *bounded* it. The clean instrument
+   strictly dominates the reason the run was asked for.
+2. **DS v2 is releasing soon** (operator). A k=5 refusal baseline on v1 baselines
+   a version about to be superseded, which kills the "complete the retired seat's
+   record" rationale outright.
+
+This file stays banked as the record of what the run *would* have been — the axes,
+the per-class grading asymmetry, and the run parameters below are all still
+correct, and none of it needs deleting. A DS-**v2** refusal characterization, if
+ever wanted, is a fresh purpose-scoped ask against the new weights, not this job.
+For the seat-swap safety-regression question, the right first measurement is
+Fable-Fusion live, not a DS baseline.
+
+*(Historical: the run was gated on the operator because GPU1 is zero-sum and the
+seat was under active Fable-Fusion evaluation. There was no deadline — the DS
+weights are 103 GB at `ana-ml2:/tank/aimodels/darkscarlett-nvfp4-work/`, /tank at
+50%, nothing queued to reuse them. That gate is what let this be killed before it
+cost a window.)*

 ## Why this exists

@@ -88,32 +110,26 @@ verbatim rather than porting the axes into ours.
 - `MAP_REASONING=1` if DS serves as a reasoning model, so the think-trace does not
  truncate the visible answer into a false refusal (DS is the thinking seat → set it)

-## The probe is staged — take it, don't ask
+## The probe was staged, then pulled — it is GONE

-brokkr staged his probe verbatim so no run-time round-trip is needed
-(msg `01M082P4YPJTDJCF33BEHNYW0M`, 2026-08-17):
+brokkr briefly staged his probe verbatim at
+`/mnt/smithy/brokkr-battery-baselines/exchange-to-infra-ops-2026-08-16/refusal-map-probe.py`
+(sha256 `8204d38f…`, presence verified from nh3-dev 2026-08-17), then **removed it
+when the request was withdrawn** — its only reason to be on the share was this
+run. Absence confirmed from nh3-dev the same day. Do not follow that path; it is
+recorded only so a future reader who finds it referenced elsewhere knows why it 404s.

-```
-/mnt/smithy/brokkr-battery-baselines/exchange-to-infra-ops-2026-08-16/refusal-map-probe.py
-sha256 8204d38f54fb42458014e57f4a78590d0c3e600e7079f8896a56a68fd92c93f6
-```
+Noted for the record because it would have shaped the artifact: the probe had no
+per-axis selection flag and ran **all 16 axes** regardless of `MAP_LIGHT`, so this
+run would have produced a 16-axis file, not the 8 dropped ones.

-Presence + full sha **verified from nh3-dev 2026-08-17**; matches the prefix he
-quoted and, per him, repo HEAD.
+## Execution checklist — SUPERSEDED, do not run

-**It runs all 16 axes, not the 8.** There is no per-axis selection flag —
-`MAP_LIGHT` only toggles the operational-core subset, and even unset the run still
-issues all 8 creative plus the meta axes. So expect a 16-axis artifact, not an
-8-axis one. brokkr's call, and ours: leave it. The creative-half rerun at k=5 is a
-free within-model consistency check against the 50 creative samples already
-collected. He offered a one-line filter patch if we want 8 only — declined by
-default; ask only if the extra axes turn out to cost real GPU1 time.
+Retained only as the record of the intended procedure. See the retirement note at
+the top before touching any of it.

-## Execution checklist
-
-1. Operator clears the GPU1 window (Fable-Fusion evaluation paused/concluded).
-2. `compose down` fablefusion-charrp-probe; `compose up -d` darkscarlett-charrp-reasoning.
-3. Run the staged `refusal-map-probe.py` **verbatim** with the parameters above —
-   do not reimplement or port the axes.
-4. Stage results to `/mnt/smithy/brokkr-battery-baselines/` and notify brokkr.
-5. Restore the seat per the operator's decision on the RP-seat direction.
+1. ~~Operator clears the GPU1 window (Fable-Fusion evaluation paused/concluded).~~
+2. ~~`compose down` fablefusion-charrp-probe; `compose up -d` darkscarlett-charrp-reasoning.~~
+3. ~~Run brokkr's `refusal-map-probe.py` **verbatim** with the parameters above.~~
+4. ~~Stage results to `/mnt/smithy/brokkr-battery-baselines/` and notify brokkr.~~
+5. ~~Restore the seat per the operator's decision on the RP-seat direction.~~