The hypothesis under test was the lockstep discipline: that each run of an AI client consumes a
fixed number of draws regardless of the path it takes, so a reimplementation could keep the
generator aligned with the right COUNT and order of draws while getting the decisions wrong.
It is false, and it fails at four scopes. Measured with a new bracket on
StrategyClient::OnResumePlaying over the per-client generator at +0x134, six runs across VM140 and
VM145, five fresh processes; every unpinned run's autosaves are byte-identical to the published
oracle, so the instrument is behaviour-neutral (rules 19 and 26 both discharged).
across clients turn2->turn3: 3 / 0 / 0 words for AI players 32 / 496 / 512; human 0
across turns client 32: 3 words on turn 2, 7 on turn 1
across processes client 512 makes ONE cl_RandRange call on turn 1 -- the research-target
tie-break at 0x006a8495, phase 18 -- and it cost 1 word in one process and
3 in another, because RNG_NextInt is an unbounded rejection loop
per site RNG_Chance costs ZERO words at p<=0 and p>=1
Twenty-one live draw sites in an AI turn, in twelve functions (plus two provably dead ones); two
fired on the reference turn, three on turn 1. Only three are unconditional, and all three only
given that their enclosing function was called. Six of client 32's seven turn-1 words come from
the ship-design composer 0x006ad700, which is also where the only loop-carried draw lives.
Also: cl_RandFloat 0x00579c70, a third cl_* RNG facade, found twice independently. It reaches
RNG_NextFloat by a TAIL JUMP, so no rel32 sweep for the entry points can see it -- which is why
ai-turn-logic.md 5's 'zero NextFloat calls from the AI module' reads as true and is not. All 29
call sites of the three facades are inside the AI band: the cl_* RNG facade is AI-only surface.
Positives for the engine: the AI draws from nothing but its own client's generator (foreign_words
0 on every bracket), the human client draws nothing at all, and the per-turn cost is single digits.
Rung B is unaffected. Rung C needs the decisions.
gen_addresses.py distinguishes "addr" (VA -> RVA) from "offset" (a member
offset). Three lane-K entries used "addr" for offsets and generated negative
RVAs, which broke the engine build. Two are re-expressed as offsets; the third,
StrategyServer_off_RNGPtr_S_frame, is withdrawn - addresses.json already has
StrategyServer_off_RNG = 0x16c for the S frame alongside off_RNGPtr = 0x168 for
S+4, which is precisely lane T's rule. The second call site found this lane
confirms it rather than correcting it.
The second turn driver, 0x007d92a0, read byte for byte. Corrects turn-driver.md
section 5: of the four subsystems it said live here, only bankruptcy does.
- 36-phase map with strides enumerated (EncounterResults 0x178, Encounter 0x74,
member 0x44) and the arity check that logs but does not return.
- Phase 7 is encounters.clear(), not a filter: the erase pair is the same
four-argument shape vector<Encounter>::operator= uses, and both arms converge
three instructions later.
- Bankruptcy: ProcessBankruptcy at phase 15, UpdateBankruptcyLimits at phase 31.
Three corrections to formula-gaps Q1 - the divisor is the double
-0.15000000596046448 not -0.15, the per-system income term is clamped at 0
before summing, and the 3.3 factor lives in .bss and is DB-loaded.
- Turn results are FILLED here (phases 6, 11, 18 write S+0x2f4[PlyrIdx]) but
rotated by ApplyEncounterResults and dispatched by SynchronizePlayer as event
0x25 afterwards. sizeof(SETurnResults) = 0x11c, enumerated five ways.
- BuildTurnEvents is misnamed: it is the setup/load/rejoin resync push, gated on
a pending descriptor, and references no EVENT_ string at all.
- TurnEvents_Write and TurnEvents_Read are swapped in Ghidra (layouts.json is
right). sizeof(TurnEvents) = 0x18, enumerated four ways.
- The autosave: StrategyHost::Autosave 0x00895210, its four localized paths, the
rotation that fires only on the post-turn call, the connection detach around
the write, and why the payload carries nothing time-, name- or machine-derived.
- Two RNG sources in the tail that nothing models: one NextFloat per expired node
line, plus draws inside the combat resolver. Both run before the autosave.
- S+0x8 advances twice per turn, not once.
Repo-wide correction: the research-event roll costs one or two RNG words, not
one. Fixed in unlock-cascade.md, addresses.json and lane-u.json; the captured
compare artefacts under verify/results are left alone as run records.
Reads: TechTree::PrereqsMet 0x0057d8e0 (AND of ORs; zero groups TRUE, an empty
group FALSE), the prerequisite layout at TechDef+0x88/+0x98, the tail collector
0x00587cc3, and the head of OnTechResearched (RecordObservedTech unconditional;
the research-event roll gated on ResT == def && ResearchRollPending).
Live: 35 compared calls over three workloads, 0 divergences, tracecmp exit 0,
End-Turn oracle hashes unchanged. The EVENT_TECHS_UNLOCKED residual lane P
predicted and lane V measured is closed by running the cascade, not by posting
on completion.
12 new addresses in ghidra/addresses.d/lane-u.json (header 615 -> 627).
Every serializable class carries an enumeration of its own fields -- its
Write(Stream&), walking the members in order with a 4-char tag. This decodes
that idiom mechanically for the whole binary in 0.35 s.
Validation first (tools/serializers.py validate), against answers the campaign
already had before the tool existed:
A 305/307 field offsets+kinds exact across 17 classes, 0 WRONG, vs
struct-recovery.md 1-4 and observedtech-append.md
B sizeof from the container-stride divides: ObservedTech 0x2c, MoraleEvent
0x50, PlayerReport 0x30, DiplomacyStats 0x24 -- all matching
C 22 of save_reader.py's shapes, tag order identical (Sys 78 tags,
Player 104, CreateParams 25, Ship 22): 22 agree, 0 disagree
D Read/Write cross-check on every class: 437/437 field offsets agree
At scale: 386 classes with a Write, 1,682 member fields.
verified 87 (542 fields) | clean 77 (328) | unnamed 176 (471)
partial 31 (341) | empty 15
58 classes with a sizeof corroborated by a second line of evidence
(45 container stride, 13 enumeration meeting the embedding bound); the rest
report a lower bound and say so.
Four things each worth 10-170 classes: the RTTI class hierarchy descriptor as
the only honest "is this an IStreamable" test (a 3-slot vftable also matches
TacAISquadRule_* and the row parsers); mod=0 memory operands, which x86disp.py
cannot index and which hide every field at offset 0; the member->id pointer
idiom behind every handle field; and sub-writers, both base-class and private
(StrategyServer's six id lists live in FUN_00794cd0).
Failure classes are enumerated in the finding -- 176 anonymous-tag classes are
a hard limit on names but not on layout, and the other 64 are bounded
mechanical fixes. Two fields lost to a value assembled across a branch were
left unrecovered rather than patched with an unverifiable heuristic.
Write-back: 288 structures + 328 labels into Ghidra (0 failures), +201
addresses.json entries, header regenerated with tools/gen_addresses.py.
Note: ghidra/addresses.json also carries lane V's already-written live
confirmation text on ObservedTech_sizeof and ServerPlayer_off_ObservedTechs --
their edit, swept in only because we share the file.
Settles the 0x18-vs-0x1c contradiction lane X raised. 0x1c is right, everywhere,
and there is exactly one std::string instantiation in this binary:
_Bx@0, _Mysize@0x10, _Myres@0x14, _Alval@0x18.
ObservedTech+0x24 is that string's trailing empty-allocator word, not the
unaccounted data field it was read as. Three complete enumerations of the element
each skip it: ObservedTech::Write 0x00817cf0, the ctor 0x008562a0, and the copy
ctor inlined at 0x0079a184. Generalised with a new scanner, tools/strfootprint.py,
which recovers every (base, disp, tag) handed to the Mars::Stream string helpers:
65 std::string members off a non-stack base across every serializer in the exe,
ZERO with a sibling member inside the 0x1c span, and 51 of the 52 measurable
inter-member gaps exactly 0x1c. Corroborated by the vector<string> walk stride
(add esi,0x1c @0x00699c29), PostEvent's by-value strings at [ebp+8]/[ebp+0x24]
with RET 0x4c, and MoraleEvent 0x50 = name@0x34 + 0x1c.
Blast radius: zero recovered struct tables were wrong. Every string-bearing layout
already used 0x1c spans and 0x1c gaps -- ServerPlayer::pswd @0x2dc..0x2f7, the row
flagged for re-checking, included. Only prose carried the 0x18 number: the
loader-prototypes conventions line, the GlobalConst_ParseString prototype, and the
ObservedTech element table. struct-recovery S0 additionally had _Mysize/_Myres
transposed (size@0x14, res@0x18) while every table in the same file used the
correct offsets; fixed.
ObservedTech's four on-disk fields are now mapped rather than guessed, by reading
the serializer as lane X suggested: +0x04 uint16 otnF, +0x06 uint16 otnL, +0x08
bool odet (ONE BYTE, WriteBool), +0x0c std::string otch (0x1c), +0x28 int owith
= 0x2c exactly. That matches save_reader.py's on-disk order already. Game::
ObservedWeapon (0x00817bc0/0x00817b10) is the same element with tag owep.
Oracles unaffected and re-run: save_reader 36/36 and --strict exit 0 on all three
real saves; state_checksum 38 tests OK, coverage PROVED byte-for-byte on turn1 and
turn3. sots-engine wip/strings 32d3e36 syncs the header and corrects two stale
"unpinned" comments: clean_room_check OK, host ctest 33/33.
Standing rule this produced: never size a struct member from the offsets the code
touches. This build's STL puts the empty allocator LAST in both string (0x1c) and
vector (0x10), and an empty allocator is never loaded or stored, so a touch-scan
undercounts by exactly 4 every time. Size from an enumeration instead.
Ghidra does not index ModRM displacements, so `lea reg,[reg+disp]` -- the MSVC
idiom for taking a member's address -- is invisible to find-constant-uses. That
blind spot parked ServerPlayer+0x274 and covers every non-trivial member of the
~1,600 classes still to map.
tools/x86disp.py: full x86-32 length decoder (prefixes, 1/2/3-byte opcodes,
ModRM, SIB, sign-extended disp8, disp32, every immediate form) swept from
Ghidra's 41,089 function starts so decodes begin on real instruction boundaries.
2,174,504 instructions, 612,166 displacement sites, 100.0% code coverage, 70
desyncs (0.17%), zero unknown opcodes. Excludes no-base disp32 forms
(mod=0/rm=5, sib.base=5) which are absolute globals, not member offsets.
Commands: build/query/cohort/func/dis/stats/brute. Works off a gitignored local
cache in dumps/ rather than hammering CT111.
Validated before use: re-finds lea eax,[ecx+0x29c] in ServerPlayer::GetEventStorage
(0x0080db00) and both known OnTechResearched +0x29c sites, plus a new one in
ProcessTurn. Positive control: the ServerPlayer serializer scores 50/50 known
offsets.
sizeof(Game::ObservedTech) = 0x2c (44), proven three ways: the exact magic
divide 0x2e8ba2e9 sar 3 at 0x0087239f, imul reg,reg,0x2c at 0x0087243a and
0x007b735b, and the search stride add edi,0x2c at 0x007ba257.
Append site: RecordObservedTech+0xdf (0x007ba27f) --
lea ecx,[player+0x274]; call vector_ObservedTech_push_back 0x007b7320
RecordObservedTech (0x007ba1a0) is a direct callee of OnTechResearched and
de-duplicates by tech name before appending. The realloc through 0x007b5820 is
why lane R's guard saw all three vector words move. Element carries a vptr
(RTTI .?AVObservedTech@Game@@) at +0 and a 0x18-byte std::string at +0x0c; the
four on-disk ints map onto +0x04/+0x06/+0x08/+0x24/+0x28 in an order this read
does NOT determine, and is not guessed.
Also corrects harness-audit row 11: ComputeBudget has no store to Budget+0x64
(its only +0x64 accesses are loads off a different base), and ProcessResearch's
int* overbudget arg is a ProcessTurn stack local, not Budget+0x64. Agrees with
lane R's guard seeing 0 changes in 4284 calls.
Honest limits are recorded in the note and the board: this is a recall tool, not
an oracle. Class-level precision at 0x274 is ~13% by function, i.e. a ~900x
search-space cut that still needs one call-graph check. Cohort ranking must not
be used as a hard filter -- it would have discarded the correct answer here.
Ghidra writeback: labels + plate comments on RecordObservedTech,
vector_ObservedTech_push_back, ObservedTech_ctor, vector_ObservedTech_assign,
vector_44B_grow, vftable_ObservedTech.
Read TechTree::SetResearched (0x00581e10) end to end while wiring the research event
posts into ours (sots-engine wip/post-events). Two things it settles:
- SetResearched invokes the owner callback as vft+0x10(def, (flags>>2)&1), so bit 2 of
flags IS OnTechResearched's "silent" argument. ProcessResearch passes flags = 2, so
silent is FALSE and the completion event IS posted. Previously this was only
inferable from lane R's observation that a completion call moved EvNxID by two.
- The unlock cascade in full: the node's child edges set child state 0->1 and
child.costRP = min(child.costRP, edge.costRP); then a whole-tree sweep sets state 2
and stamps turnAvailable from ModCount ONLY when it currently reads -1 (so it is
sticky, and a node re-entering state 2 is not re-announced), and recurses into any
zero-cost available node. Both loops test the node itself, not a parent -- there is
no parent clause anywhere in the function, confirming lane E's correction.
Also records what is and is not known about vector<ObservedTech> at ServerPlayer+0x274:
the on-disk element shape is confirmed against turn3-state.sav (its string holds a tech
name), while sizeof(ObservedTech) and the append call site are still unpinned. The
cheapest route to the stride is now a measurement from the new observed_techs region,
not a search -- find-constant-uses does not index lea displacements.
New entries: ServerPlayer_off_ObservedTechs, TechNode_off_TurnAvailable /
TurnResearched / Order / Children, TechEdge_off_CostRP / ChildDef,
TechTree_off_OrderCounter, TechTree_SetResearched_flag_Force / _flag_Silent.
Generated header regenerated with tools/gen_addresses.py (386 -> 396 entries).
Closes B3's oracle gap and B2's known gap. Both milestones write into the
owner's event list; nothing modelled it, so their clean compares bounded the
economy fields only.
Container: EventStorage embedded at ServerPlayer+0x29c (0x1c bytes), verified by
ServerPlayer::GetEventStorage 0x0080db00 whose whole body is 'lea eax,[ecx+0x29c];
ret'. EvNxID at +0x14 == player+0x2b0, exactly the byte run the harness guard
reports. The list is a vector<TurnEvents{int EvTurn; vector<PlayerEvent>}> --
bucketed by turn, not flat, which the save-editor struct note had wrong.
Entry point: int __thiscall EventStorage::PostEvent(this, std::string BY VALUE,
std::string BY VALUE, void* obj, Vector3* pos, int turn, const char* img,
int act), 0x008862b0, RET 0x4c. 161 call sites in 110 functions: this is the
whole simulation's event API, not a research helper. Dedup is per turn bucket on
message/image/location/position/action but NOT summary; act 0 with no subject and
no position is stored as 2; EvNxID starts at 0 and is promoted to 1 on first post;
PruneOldTurns drops buckets older than turn-50 with an off-by-one that always
leaves one stale bucket behind.
Record: 0x74 bytes, EvEID(+4) EvDsc(+8) EvMsg(+0x24) EvLoc(+0x40) EvPos(+0x44)
EvImg(+0x50) EvAct(+0x6c) EvCID(+0x70) -- confirmed field by field against
turn3-state.sav, which contains the EVENT_RESEARCH_OVERBUDGET record itself.
B3's defect is fully explained: 0x00587b97, reached only from the
completion-roll-FAILED branch, under !wasDone && nowDone && owner.
Three note corrections: EvPos defaults to FLT_MAX, not infinity (writing +inf
changes the save bytes); the save array is turn-bucketed, not flat;
EVENT_TECHS_UNLOCKED has no 'parent researched' clause.
56 entries added to ghidra/addresses.json (header regenerated); 11 prototypes,
13 labels, 12 comments and 2 structs written back to the Ghidra project.
