Preserve retained Jimmy update order on initial story entry

This commit is contained in:
UnixNight committed 2026-11-06 17:08:59 -05:00
1 parent 4cd62d3627
commit ea04210f52
24 files changed
+1209 -22

No files matched your search

@@ -0,0 +1,125 @@
# Jimmy body-context alignment and transform publication on return
2026-11-06. Static follow-up to the null-body-context limitation in
[the base/STOP probe](jimmy-base-return.md). Original identity and cached REA
instructions are reused; no new native cases or gameplay changes are claimed.
## Source identity and guard
Canonical Neutron.exe SHA256:
`f5281076e1a2492f2d61f87aae3f4544b6d0f535ebe33f9fbb6f78710fe1aea8`.
Complete EXE instruction Evidence:
`ev_0d64cc58fa8f5573fc8e61e0f2c20e36288099212bd2257d3a3121361fd6c904`.
The read-only `tools/binary/jimmy_body_transform_return.py` verifies that subject,
12 source sites,16 virtual bindings and13 bounded source ranges. Its generated
report contains PE file offsets, hashes and the matching retained REA extracts:
`artifacts/jimmy-body-transform-return/source.json`.
The prior OMT transform Evidence is
`ev_af082632c98b34ccfcedef01466cafa03365e7fbf482496d01a4cda6d2f8f787`.
It is a truncated DLL disassembly prefix, not complete DLL coverage. Existing
transform ranges are present and rechecked against the supplied DLL through
`jimmy_model_bridge.collect`; see [render transform](jimmy-render-transform.md)
for the earlier12 angle and14 hierarchy/final-matrix observations and limitations.
JN-Engine's original class notes supplied field/navigation leads only.
## The live context bridge can write current Euler, behind a separate gate
| Original VA / EXE file offset | Recovered receiver and state effect |
|---|---|
|462520..462649 /62520..62649|Runs before common458930→473240/facing. Requires475CA0, own game210→4719F0(byte13E), and nonnull own game338→472570(pointer234) for the context path.|
|46255D..46258F /6255D..6258F|Reads support polygon's+4 link and linked+80 payload; invokes current game complete130(actor game pointer,linked record,payload,dt). Lab4C3CD0,VR044CE910 and outside4C5A80 all bind130→44E030. The callback occurs **before** the alignment gate.|
|462595..4625A2 /62595..625A2|Own game1E8→471950 reads live byte138. False bypasses normal→Euler publication but still reaches common update/facing. This gate is independent of cached support218 and raw input owner.|
|4625A8..462632 /625A8..62632|For true138, obtains current Euler328→4723F0, copies support normal Vec4 from polygon+8..14, calls476FF0 and writes resulting XYZ through330→472420. This changes **current Euler only**: target274, dirtyA, mode704 and turn614 are not written by this suffix.|
|471080/471090 /71080/71090|Property getterC4 reads byte10. Property setterC8 writes byte10 and invokes live setter1EC.|
|471950/471960 /71950/71960|Live getter1E8/setter1EC read/write byte138. Property10 and live138 are separate fields.|
|47059C..4705A9 /7059C..705A9|Generic Init470540 restores live138 through1EC from propertyC4. It does not infer alignment from landing or input ownership.|
Crucially, the ordinary Player constructor explicitly disables this path:
43729E clears EBX;437332 installs the active Player game table4AF6C4;437426
pushes EBX=0 and437429 calls471090. The active constructor table's1EC resolves
471960, as does the eventual Jimmy table4A373C. Thus both property10 and live138
are set false. The inspected constructor interval contains no intervening EBX
assignment; its callees follow the recovered x86 nonvolatile-register convention.
Subsequent generic Init restores the false property into live138. This is a
positive constructor/Init chain, not an assumption from the old null-context probe.
This does **not** prove no uninspected runtime path ever changes the property.
The previously recovered Carl camera preparation/STOP/DEFAULT boundaries do not
supply such a write. Enabling this surface-alignment path for Jimmy, or using it
as a new cutscene-exit leveling rule, would therefore introduce unsupported state.
The generic normal-angle helper's existing Scooter observations remain separate
from whether Jimmy is eligible to invoke it.
## The supplied static lab terrain takes the ordinary callback path
The authoritative `omt/level1b.omt` SHA256 is
`5a33d11311fc222631eb8f7384d1863192cb1f55f7e67035caa0ae12fd9892ab`.
Using the existing bounded OMT parser and source shape/checkpoint exclusions,
the helper inventories2869 polygons,2851 eligible for static registration.
Among eligible polygons,2526 have texture references with IDs0..30 and325 have
no texture. The report retains counts, shape IDs/names, original payload offsets
and payload hashes. No texture ID is guessed from the port's material index.
44E030 requires both linked record and payload. It reads polygon material+0C,
then the material's texture-definition ID+8; missing material/payload bypasses
the callback's state paths. All identified static-lab IDs0..30 fall below its
first special selector range50..55, so they select ordinary surface sound5 at
44E407. This excludes the callback's higher-numbered lava/ground-hazard branches
for **that supplied static geometry**.
When desired surface sound changes,44E426 invokes Jimmy **complete**1E4→42AEC0.
That method stops/restarts an existing floor-loop handle at completeA44 only if
the prior handle is valid. It writes the new handle, without touching either
Euler vector, mode704 or turn614. This is distinct from game-interface1E4, which
is another field accessor. The ordinary path supplies no movement reset and
does not grant/remove raw control.
This is not a complete runtime collision-world inventory. It does not prove the
provenance of a retained338 context, rule out additional runtime geometry or
dynamic interactions, or execute the full original body/derived Jimmy frame.
The old Bubble floor component investigation already covers this same callback
with explicit synthetic dependencies; another duplicated finite negative probe
would not establish the missing original Carl entry/resume state.
## Transform publication does not conditionally hide the bank
The later473770 pass calls each game interface3DC and3E0, **not3C8**.
Jimmy3C8→4235B0 is the separate supervisor/editor update selected by470280 when
app+8C is set. It must not be described as normal render synchronization.
For Jimmy,3DC→4624B0 copies current logical/body XYZ into complete44/48/4C.
3E0→4624F0→473540 reads current Euler, converts each degree axis to masked
14-bit angle units, then3E4→462500 writes complete50/52/54. These routines do
not test raw control, movement mode, animation name or cutscene state, and do
not replace current or destination Euler. Previously verified OMT hierarchy
construction consumes those cached pitch/yaw/roll angles. The known publication
path therefore provides no scene-return branch that conceals a residual bank.
Whole-world parent transforms, full morph playback and final render flags remain
separate from the already-probed identity-parent/unit-scale contract.
## Result, implementation and remaining observation
The optional body-context bridge contains a real current-Euler writer, but the
ordinary Jimmy constructor/Init disables its alignment gate. The identified
static lab terrain path affects surface audio, not banking state. The normal
logical→visual synchronization publishes current pitch/roll rather than dropping
them. These facts close specific omissions in the composed return explanation;
they do not explain the user's upright whole-original return or justify a new
reset. Keep the existing playable port unchanged until actual original entry,
registry placement and resumed input state are measured.
Integration follow-up: [the menu/player lifetime chain](main-menu-player-lifetime.md)
subsequently establishes retained player-before-input order for successful original
TSK story entry. The resulting correction changes that entry classification; it
does not enable this optional surface-alignment gate or alter visual publication.
Verification is static: source hash/site/binding/range checks, revalidated prior
transform contracts and complete supplied-static-lab polygon inventory. No new
original execution, native Windows comparison, runtime edit or commit is made.
```sh
python3 tools/binary/jimmy_body_transform_return.py ../Original \
--rea artifacts/jimmy-movement-audit/rea-original-disassembly.json \
--output artifacts/jimmy-body-transform-return
```
@@ -1,5 +1,13 @@
# Focused decompilation checkpoint: cutscene lifecycle
Follow-up: the source gate for successful TSK story entry is now closed by
[menu/player lifetime](main-menu-player-lifetime.md). BackgroundJimmy is reused
before the first playable game, making input follow his update. Program's initial
story order is corrected; camera preparation/landing algorithms are unchanged.
First-neutral-tick/300-idle native Vulkan and eight managed trajectories pass.
The investigation below records the earlier open state; whole-original/native
Windows comparison and integrated package validation remain separate.
2026-11-05. User redirects work toward understanding the original before further
implementation. Preserve runtime baseline a2e89c6 and packaged runtime dae13ac.
No src/, regression test, build script, original or packaged binary changes are
@@ -1,5 +1,14 @@
# Carl skip lifecycle: recovered boundaries and the remaining observation
Follow-up: [menu/player lifetime](main-menu-player-lifetime.md) resolves the
successful TSK story entry ordering omitted by this earlier checkpoint. The port
now includes first story entry in retained Jimmy-before-input order. Its baseline
native regression fails on neutral tick0 and passes first input plus300 idle ticks
with the correction. No additional movement/angle reset is introduced.50 derived-
caller source cases and the body/visual-sync audit close other local omissions;
the historical findings below remain scoped to their earlier evidence. Full
original/native Windows gameplay comparison is still unverified.
2026-11-06. Execution of the accepted next-work plan, with the user's requirement
to preserve original behavior exactly. Runtime a2e89c6 and packages dae13ac remain
unchanged. This checkpoint adds evidence and observation tools, not a tilt fix.
+51
View File
@@ -0,0 +1,51 @@
# Derived Jimmy frame around the controller
2026-11-06. This closes a caller-level omission in the earlier component
composition; the lifecycle correction is separately established by
[menu/player lifetime](main-menu-player-lifetime.md).
Canonical EXE SHA256f5281076e1a2492f2d61f87aae3f4544b6d0f535ebe33f9fbb6f78710fe1aea8;
retained REA instruction Evidence
`ev_0d64cc58fa8f5573fc8e61e0f2c20e36288099212bd2257d3a3121361fd6c904`.
The guard verifies14 source sites,21 virtual bindings,15 ranges and the actual
controller call. Text file offsets below equal preferred VA−400000; the helper
PE-maps them independently.
| Source boundary | Recovered ownership |
|---|---|
|424600..424D10 /24600..24D10|Actual Jimmy game3C4 caller. Null raw control closes/hides helper UI and stops sound handles. It does not itself clear movement mode or stored turn.|
|424721..424751|If the menu exists and shared4F8181 is set, poll menu4A8→403A10. State2 calls menu460; both state2 and other states return before the controller. This lock is distinct from raw ownership.|
|424767 complete208→424D10|Objective/UI helper before the controller. The separate objective presentation does not set modal input gate4F8198; that explicit-modal opener is406F90. See objective-timing.md.|
|424BFC complete1A4→425170|Upward world query/velocity handling. Its falling callback425120 maintains mode1; it is not a grounded resume reset.|
|424C08 complete1A8→4252E0|Downward/shadow-distance query and own helper state.|
|424C2B complete194→42AAC0|Airborne sound cleanup.|
|424C34→437940|Actual ordinary player wrapper. The source ownership/support/input-order conclusions remain here.|
|424C3F complete1C8→424E80|Invisible upkeep, with its own mode gate and ordinary non-Invisible return.|
|424C88 complete1B4→4255C0|Impact/material helper; world-hit dependencies remain separate.|
|424C94 complete1B0→425490|History/position correction. Calls game314→472360 Vec4 **root** setter; this is not an Euler setter.|
`jimmy_derived_frame_probe.c` executes the complete derived caller and ordinary
helpers on private copied source.437940 is a recording adapter, upward/downward
world queries return no hit, and menu/resource/visibility transports are explicit
adapters. Thus these cases test what the caller changes outside its controller,
not another execution of the controller/body/whole game.
All50 cases pass:48 combinations of raw ownerNULL/Jimmy/other, menu lock0/1,
menu state1/2, supported0/1 and rising0/1, plus two180-tick unlocked intervals.
The caller dispatches the controller only while the menu lock is clear; raw
ownership is passed onward rather than becoming another caller-side gate.
Mode1, turn11.666668 and distinct current/destination Euler stay unchanged outside
the recorded controller. Queries and menu-close callbacks have checked counts.
Artifacts: `artifacts/jimmy-derived-frame/native-observations.json` binds generated
source, executable, source contracts, original identity, capture metadata and log
hashes; capture wineExit0 and `JIMMY_DERIVED_FRAME_COMPLETE` are required. Primary
integration reran the parser/guards and verified all50 cases. No full body/world
hits, OMT playback, native Windows or actual original Carl frame is claimed.
```sh
python3 tools/binary/jimmy_derived_frame.py ../Original \
--output artifacts/jimmy-derived-frame/build \
--rea artifacts/jimmy-movement-audit/rea-original-disassembly.json \
--capture artifacts/jimmy-derived-frame/probe
```
@@ -0,0 +1,93 @@
# First story entry inherits the menu's Jimmy registry position
2026-11-06. This closes the missing source lifetime distinction behind the Carl
neutral-return failure. It does not establish a universal rotation reset or a
native Windows/full-original cutscene capture.
## Source and supplied data
Canonical Neutron.exe SHA256
`f5281076e1a2492f2d61f87aae3f4544b6d0f535ebe33f9fbb6f78710fe1aea8`;
retained REA instruction Evidence
`ev_0d64cc58fa8f5573fc8e61e0f2c20e36288099212bd2257d3a3121361fd6c904`.
`tools/binary/main_menu_player_lifetime.py` verifies8 calls,15 sites,15 bindings,
12 instruction spans and both actual GAMs. Coordinates below are preferred VAs;
the helper PE-maps file offsets independently (these text offsets equal VA−400000).
| Boundary | Source fact |
|---|---|
|46A641→46B160;46A660→46A310→44D0E0|Construct application and MAIN. This bounded prefix alone does not create Jimmy.|
|46BA12/46BA18|Application constructor seeds elapsed+804=0 and splash stage+808=0.|
|46BEC0..46C067|Application update increments elapsed. Strict elapsed>3s or an accepted Escape edge advances stages0→1→2→3; the3s float is48E5A0.|
|46BF2F..46C00A|Stage1 advances to2, clears elapsed and dispatches app114(`level1d.gam`) at46C004. This is the previously omitted positive background-load edge.|
|app114→46DFB0;46BEDF;app104→46DFF0|Queues filename+74D with signed callback countdown+7FA=2; underflow sets pending+74C; the application then invokes ordinary selective teardown/load.|
|Level1D GAM|LV1D,54 records, first3JIM/JIM1 at raw8. SHA256276133d750eaa82724b35c12a91fce41a9e074224e3ef12c3624564ae8ed4d9e. Loading the supplied background creates Jimmy before a user story load.|
|46E570..46E67D|After GAM records, the whole surviving registry receives Init28 then PostLoad40C, including retained MAIN; Init is not limited to newly parsed records.|
|MAIN/Level1D game28→44D4E0→4587F0→470540|Common Init restores stored+11 into live+139 through4705B3/4705BC. Both games' stored value is1. MAIN's constructor-default live0 therefore does not imply indefinite retention after background initialization. Their40C is44DF30, a noop call plus own game520=0 store, with no later lifetime reset.|
|404B4F→460E70→460EA0/460EF0→4603F0|The NewGame action queues NewGame.tsk; the reader consumes that request. The request itself does not initialize Jimmy.|
|460A50..460A6A;460BD1|The reader requests a destination GAM only through an existing non-null controlled receiver accepting **C3DJIMMY** at4ECB20. It dispatches Jimmy complete168→43B5A0 with the staged filename/start/root data. The no-Jimmy path does not supply that request.|
|43B5A0→app100/46DF70|Uses the ordinary player scene-transition request. The successful TSK-to-GAM path does not construct a new player before teardown.|
|Jimmy4226B5..BF;42275A|Stored/live teardown predicate0 via470F40(0), reusable game09=1. The initialized old games qualify for removal; Jimmy survives.|
|46E38B/46E45A→46E6A0,46E6E0|Matching reusable3JIM is found in the existing registry. Reuse does not invoke its constructor or re-append its node. The new game header is allocated/appended after that existing node.|
Lab Level1B has129 records and first3JIM/JIM1 at raw8, SHA256
316fff40db565c2add5b07cbccdef640ddb67c64d6625fea1429ce2a0542a234.
The prior [game-lifetime proof](../checkpoint/level-game-lifetime.md) supplies
the initialized-game predicate, destructor unlink and retained-character order.
The earlier fresh game→Jimmy proof remains correct **when no reusable Jimmy
exists**. It was incorrectly applied to the port's first TSK-backed story scene.
## Causal consequence and implementation boundary
A successful original TSK character/destination load requires an existing Jimmy;
the ordinary loader retains him and appends the new game afterward. Thus the
normal New Game story path has **Jimmy update before level-game input**, even
though it is the first playable lab. The background edge explains where that
player comes from; successful-request and retention guards establish the ordering
without treating a black screenshot as proof of the complete New Game sequence.
At Carl completion, grounded/released Jimmy clears mode704/turn614 in437940
before438A60 can select air input. The stale airborne destination is not republished.
This agrees with the already-executed original-component composition and the
user's upright-original observation. No Euler/turn reset is added to control grant,
DEFAULT, landing, task or animation callbacks.
The port previously selected retained input order only when `carriedSave!=null`.
Its initial story also consumes NewGame.tsk, but incorrectly used fresh order.
Correct only that scene-entry classification: TSK-backed story and retained
transitions use player-before-input; direct scene inspection/fresh simulations
keep fresh order. Remaining scalar/helper/lifetime reconstruction stays bounded.
## Runtime evidence and limits
An isolated generated-copy diagnostic observed MAIN followed by Level1D allocation
and a Jimmy registry node. It required a recorded, one-time Win32 ValidateRect
acknowledgement after a pending full-window WM_PAINT. Thread snapshots suspend/
resume owned threads briefly, so timing is perturbed. The supplied HDnoCD and
640×480 NoCD variants have separately authenticated identities; their only mutual
byte differences are the3 resolution-data bytes. Neither is silently equated with
the canonical EXE. These captures show allocation/lifetime progress, not usable
rendering, completed PostLoad, user New Game or Carl playback.
Local artifacts: `artifacts/original-startup-diagnostic/{before-ack,with-ack,long-hd}/`
and `artifacts/main-menu-player-lifetime/source.json`. The baseline port's strengthened
native Vulkan control-phase regression fails on **first neutral tick0**:
inputAfterPlayer=false, current roll360, destination roll−10.000001, mode0/turn0.
This is the exact source-composed stale writer. After the correction, the native
Vulkan fixture stays upright on first resumed input and300 neutral ticks, then
resumes held movement. It preserves heading and locked body/fire/energy updates.
The existing standing/landing/air algorithms are unchanged; airborne return is
still governed by original air input, not an invented blanket leveling rule.
Verification: all336 Python plus2 HUD checks, all managed projects and36
GPU/render passes succeed. Eight new managed trajectories coverLeft/Right,
natural/skip and repeated Carl scenes. Native Linux SDL/Vulkan1920×1080 control
replay and VR04 powered steering/keyboard/800×600 resize pass, zero Vulkan errors.
The first post-return held-input readback is inspected upright. Final Debug build
has zero warnings/errors and Windows x64 cross-publish succeeds. No native Windows
execution or full-original Carl runtime is claimed.
Logs, baseline failure, source contracts, capture hashes and verified-change.json
are in artifacts/main-menu-player-lifetime/. Clean package validation follows as
a separate checkpoint. Full original/native Windows comparison remains separate
from this source-supported initialization correction.
@@ -0,0 +1,88 @@
# Original startup: pending repaint and background Jimmy allocation
2026-11-06. This is a local Wine diagnostic, not an untouched timing or visual
reference. Originals remain unmounted in the execution namespace and unchanged.
`tools/recon/original_startup_probe.c` reuses the observer's exact-subject/live
instruction guards and runs only at the generated wrapper directory
`Z:\work\game`. It samples registry/actor state, loaded modules and owned thread
contexts. Each successful SuspendThread is paired with ResumeThread; all recorded
resumes succeed. Stack words are retained as raw candidates, not a full validated
unwind. Suspension perturbs timing. No game code or actor memory is written.
## Repaint observation and controlled test
Two initial main-thread snapshots contain the guarded OMT window-procedure return
100F1ADA and its consistent local message value0F (WM_PAINT), with internal event7.
Read-only GetUpdateRect later reports the entire1024×768 window pending. Original
DLL source100F1330 converts paint to event7 and calls100F7050; its validation of
engine-owned rectangles includes100F70BA→ValidateRect. The message pump
100F11D0..100F129C drains queued messages before returning to the periodical loop.
These bounded facts explain the diagnostic target; they do not prove all startup
time is spent in a single call or that native Windows has the same failure.
`--acknowledge-repaint` makes one ordinary Win32 ValidateRect(window,NULL) call
after the second sample. The operation is explicit in output and changes only
the OS window update region. It is an intervention, so no result is presented as
an untouched original frame/timing trace.
Matched24s tests use the same standard-resolution executable, generated prefix,
sampling tool and no synthetic keyboard input:
| Condition | Observed result |
|---|---|
|No acknowledgement|16 samples through24.35s: MAIN only, no Jimmy.|
|One acknowledgement at3.10s|Pending full-window region acknowledged. By9.18s, current game table4C444C (Level1D), Jimmy table4A373C, with MAIN still in the registry.|
The HD variant also progresses to the same allocation boundary after the recorded
acknowledgement. Both resolutions remain black with sampled gameplay-viewport
delta0 and raw ownerNULL. No usable Carl, completed New Game, completed PostLoad,
or playable-render reference is established. The source trace of
[menu/player lifetime](main-menu-player-lifetime.md) supplies the subsequent
successful-TSK retention proof; these captures establish the previously unseen
background allocation, not that complete later route.
## Subject and code identity
HDnoCD SHA2568cd5d21997835a40978b8b83e344c97c92217bf6cd6858f7458b76935d64dedb,
standard NoCD SHA2565e1d3bed2d649d16ec62ead02514119a16590380d4d5a6a73103db900a15c62e.
Their only mutual byte differences are resolution-data offsetsF655E,F655F,F6561:
standard640×480 versus HD1920×1080, both16-bit. The Xvfb desktop is1024×768.
This difference prompted the controlled standard-resolution check; switching
resolution alone is not claimed to repair rendering. Neither variant is assumed
equivalent to canonical Neutron.exe.
Standard-subject REA file-header Evidence:
`ev_62814154b8ddc13d5db3c1150082b91f0f7bf1eaccf1bca208b36671a7a1bd81`.
HDnoCD complete instruction Evidence:
`ev_23555fac476985c913a80805a12b25c26c25347c6e8210ee188471dd815c276d`.
Relevant OMT instructions come from retained DLL Evidence
`ev_af082632c98b34ccfcedef01466cafa03365e7fbf482496d01a4cda6d2f8f787`;
that saved disassembly is a prefix, not complete DLL coverage. The relevant
window/message ranges are present. Original file identity is independently
checked by the project inventory and PE mapping.
Artifacts under `artifacts/original-startup-diagnostic/` retain each probe binary,
capture identity, stdout, screenshots, module/stack/state samples and final
standard probe/dependency source. `standard/` is the acknowledged matched run;
`standard-no-ack/` is the control. All actual live anchors pass. Successful capture
requires process exit0 and `ORIGINAL_STARTUP_DIAGNOSTIC_COMPLETE`.
## Reproduction
Build against the source observer in the same tools directory. Reuse only the
existing generated game copy/prefix; never use an original directory as output.
```sh
i686-w64-mingw32-gcc -std=c11 -O2 -Wall -Wextra -Werror \
tools/recon/original_startup_probe.c -lbcrypt -o artifacts/startup-probe.exe
python3 tools/recon/reference_capture.py --asset-probe \
--probe-executable artifacts/startup-probe.exe \
--probe-model=--standard-resolution --probe-arg=--seconds=24 \
--probe-arg=--acknowledge-repaint \
--probe-success ORIGINAL_STARTUP_DIAGNOSTIC_COMPLETE --seconds 30 \
--output artifacts/jimmy-skip-original-sw
```
Omit the acknowledgement argument for the control. Keep each capture in a
separate generated directory or preserve its logs/identity before another run.
+10
View File
@@ -59,6 +59,16 @@ The supplied HD resolution variant preserves the observed operands. The HDnoCD
profile is independently identified; its observation is not automatically a
canonical-original behavior claim. Other supplied variants are rejected.
2026-11-06 follow-up: the supplied standard-resolution NoCD image is now a fourth
exact profile, SHA256
`5e1d3bed2d649d16ec62ead02514119a16590380d4d5a6a73103db900a15c62e`.
Its eight live global operands match the HDnoCD profile; independent source checks
verify32 anchors across four subjects. The two NoCD images differ only at three
resolution bytes; canonical-equivalence is still not assumed. Source checks are
in artifacts/original-state-observer/profiles-4.json. Standard-subject REA Evidence
and perturbed startup observations are in
[the startup diagnostic](original-startup-diagnostic.md).
Each global address is independently guarded by its actual live instruction
operand before sampling. Runtime inspection is refused if those anchors do not
match within5s. This avoids assuming that the other variant has identical data