sots-re/findings/subsystems/roll-parity.md

# Roll parity — does an AI client's turn cost a fixed number of RNG words?

- **Type:** subsystem (live measurement + static classification)
- **Status:** **verified — the answer is NO**, at every scope the question can be asked at
- **Confidence:** high on the negative (measured on two boards, four AI clients, five fresh
  processes, with the instrument's autosaves byte-identical to the published oracle);
  medium on the completeness of the site inventory (§6 says what is not covered)
- **Owner / date:** lane PAR · 2026-09-08 · VM140 (reference) + VM145
- **Instrument:** `sots-engine` `wip/par`, new shim module `src/shim/hooks/ai_rng.cpp`,
  configs `shim.cfg.par{ctl,br,pin,orders}`, builds `par-…` … `par4-6ba2c79-20260908T2322Z`
- **Answers:** the coordinator's redirect — *"each run of an AI client consumes a fixed number of
  RNG draws regardless of the path it takes"*, the lockstep-discipline hypothesis
- **Corrects:** `ai-turn-logic.md` §5 (zero NextFloat calls from the AI module), §3.1 (the phase
  list is not flat), §4.3 (the `cl_*` façade has three RNG members, not two, and ten
  `cl_RandRange` call sites, not eight); `ai-task-system.md` §10.1's six direct `NextInt` sites

---

## 0. The verdict, at four scopes

**The hypothesis is false, and it is false in the strongest available way: it fails at the
primitive, at the site, at the client and at the turn.** There is no scope at which an AI client's
word count is fixed.

| scope | claim | measured |
|---|---|---|
| **per turn, across clients** | all AI clients of one turn cost the same | **3 / 0 / 0** on `turn2→turn3`; **7 / 0 / 1** on `turn1→turn2` |
| **per client, across turns** | one client costs the same each turn | client 32: **3** on turn 2, **7** on turn 1 |
| **per client, across processes (path fixed)** | one client's count is a property of the board | client 512, turn 1, **one and the same `cl_RandRange` call** (the research tie-break, `0x006a8495`): **1 word** in one process, **3 words** in another |
| **per site** | a site that is reached always costs the same | `RNG_Chance` costs **0 words** at `p ≤ 0` and `p ≥ 1`; `RNG_NextInt` is an **unbounded rejection loop** |

The two shadow empires (496, 512) own nothing at all — no colonies, no fleets, no systems — and
they still do not agree with each other: on turn 1, 496 spends **0** and 512 spends **1 call
costing 1 or 3 words**. So the count is not even a function of "what the empire owns".

**What this means for `game/ai`, in one line: keeping the generator aligned requires the AI's
*decisions* to be right, not merely its draw count. C-exact does not get cheaper.** §7.

---

## 1. The instrument

`StrategyClient::OnResumePlaying 0x00777480` is an exact bracket around one client's whole turn:
`if (this->+0x12c) agent->vt[1](0x26, ev)` at +0x9d runs the entire AI Process Turn synchronously,
and `this` names the client, its player net id (`+0x148`), its agent (`+0x12c`, non-null only on an
AI client) and its private generator (`+0x134`).

One MinHook detour on that function, plus an **observer registered on lane H's existing seven
draw-site detours** rather than a second set of hooks on the same targets (MinHook allows one hook
per target, and those detours are installed unconditionally whenever `hooks != off`). Two
independent measurements of the same number, per bracket:

* `observed` — the sum over the entry points, with per-return-address attribution;
* `left_delta` — `left` (RNG+0x9c8) read straight off the object at entry and exit.

They are printed together with their difference on **every** bracket, including when it is zero.
Also printed unconditionally: `agent=`, so "this client has no AI" is distinguishable from "the AI
ran and drew nothing" (method rule 20); and `foreign_words`, the count of draws that happened inside
the bracket on some *other* generator.

Two further hooks, on `cl_Chance 0x00578cf0` and `cl_RandRange 0x005798e0`, record
`__builtin_return_address(0)` so a draw is attributed to the **AI call site** and not to the one
address inside the façade where every façade draw would otherwise pile up. `cl_RandFloat 0x00579c70`
is deliberately **not** hooked — §2.

`airng.pin_seed=<hex>` re-seeds the client generator at bracket entry, so two runs start from an
identical generator state. It changes the game's behaviour on purpose; a pinned run's autosave is
never compared against the oracle.

### 1.1 Rule 19 and rule 26, discharged

The instrumented, **unpinned** runs on the canonical pair produce
`(Autosave EndTurn).sav = bb4fd9ac89f41e3b` and `(Autosave).sav = 978041acd168b56e` —
**byte-identical to the published oracle**, which was produced with `hooks=off` in a different
process on a different day. That is a stronger control than a same-key `airng=off` pair, and it was
reproduced in **two fresh process launches** (rule 26). The detour is behaviour-neutral.

### 1.2 The instrument caught its own defect, which is the point of having two measurements

The first bracket containing a `cl_Chance` printed `observed=8` against `left_delta=7`.
`RNG_Chance`'s own body calls `RNG_NextFloat`, and **both are detoured**, so one drawn word was
reported twice — once against the game's call site and once against `0x008e6e09`, the instruction
after Chance's internal call. This is exactly the nested-entry-point double count `draw_sites.h`
warned about for `IntRangeBell`/`GaussianRange` and had never seen. Fixed by rejecting the inner
row's *words* while keeping its *call count* visible, so the nesting stays legible instead of
silently disappearing. `left_delta` was never affected — which is why there are two measurements.

---

## 2. `cl_RandFloat 0x00579c70` — a third façade, found twice, and why nobody saw it

`ai-turn-logic.md` §5 states, from a rel32 scan of the AI band for the seven RNG entry points:
*"Zero calls from the module to `NextFloat`, `Chance`, `NextUInt`, `Twist`, `Seed` or
`GaussianRange`."* The very first live bracket recorded a **`NextFloat` at return address
`0x006ad878`** — inside the AI band, inside the ship-design composer `0x006ad700`.

The mechanism is a **tail jump**:

```
00579c70  mov eax,ds:0xae4808               ; g_CurrentClientIndex
00579c75  fldz
00579c77  mov eax,[eax*4+0xae47e4]          ; g_StrategyClients[idx]
00579c7e  test eax,eax
00579c80  je  0x579c92                      ; no current client -> return 0.0f, NO DRAW
00579c82  mov ecx,[eax+0x134]               ; the per-client generator
00579c88  fstp st(0)
00579c8a  lea ecx,[ecx+0x4]                 ; &mt
00579c8d  jmp 0x47d830                      ; JMP, not CALL -> RNG_NextFloat
```

A scan that follows only `E8` edges to the entry points cannot see it, and a scan that follows only
`E8` edges *from* the AI band sees a call to `0x00579c70`, an ordinary-looking façade. It has
**exactly one caller in the image**, `0x006ad873`.

Because it tail-jumps, the draw-site detour's return address is the **AI call site itself**, which
is why this is the one façade whose consumer shows up directly in a draw-site table — and why
hooking it would have *destroyed* that attribution rather than adding to it. It is left alone.

Independently re-derived by a boundary-accurate image-wide scan that follows `E9` as well as `E8`.

### 2.1 The `cl_*` RNG façade is AI-only surface

Image-wide, at real instruction boundaries:

| façade | call sites | in the AI band 0x00680000–0x006e0000 |
|---|---:|---:|
| `cl_Chance 0x00578cf0` | 18 | **18** |
| `cl_RandRange 0x005798e0` | 10 | **10** |
| `cl_RandFloat 0x00579c70` | 1 | **1** |

**Every call site of all three is in the AI module.** Nothing else in the image draws through them.
(`ai-turn-logic.md` §4.3 says eight `cl_RandRange` sites; there are ten.)

---

## 3. The measurements

Every run: load the save, one End Turn, `hooks=trace` with every template hook off. Autosave hashes
in each row. `words` is `left_delta`, the measurement taken on the generator object.

### 3.1 `turn2-state → turn3-state` — the canonical deterministic pair, VM140

Two fresh process launches, unpinned (each process seeds its AI clients differently — lane L1).

| bracket | player | AI? | run p1 | run p2 |
|---|---|---|---:|---:|
| 1 | 16 `re` (human) | no | 0 | 0 |
| 2 | 32 Fane Lao | yes | **3** | **3** |
| 3 | 496 Singularity | yes | **0** | **0** |
| 4 | 512 Singularity | yes | **0** | **0** |
| 5 | 16 (again, new turn) | no | 0 | 0 |

Both runs: `(Autosave EndTurn).sav bb4fd9ac…`, `(Autosave).sav 978041ac…` — the oracle.

Client 32's three words, both runs, same breakdown:

| site | entry | calls | words |
|---|---|---:|---:|
| `0x006ad878` — `cl_RandFloat` in `AIComposeShipBlueprint 0x006ad700` | NextFloat | 1 | 1 |
| `0x00579915` — `cl_RandRange`'s `NextInt` | NextInt | 2 | 2 |

`residual = 0` and `foreign_words = 0` on every bracket in both runs.

### 3.2 `turn1-state → turn2` — a different board for the same players, VM145

Two fresh process launches, unpinned.

| bracket | player | AI? | run p1 | run p2 |
|---|---|---|---:|---:|
| 1 | 16 (human) | no | 0 | 0 |
| 2 | 32 | yes | **7** | **7** |
| 3 | 496 | yes | **0** | **0** |
| 4 | 512 | yes | **1** | **3** |
| 5 | 16 | no | 0 | 0 |

Client 32's seven words, identical breakdown in both runs: `cl_RandFloat` ×2 = 2, `cl_RandRange`
×4 = 4, `cl_Chance` ×1 = 1.

**Client 512 is the result.** In both runs it made **exactly one `cl_RandRange` call**. That one
call cost **1 word** in the first process and **3 words** in the second. Same board, same save,
same code path, same number of calls — a different number of MT words, because `RNG_NextInt` is a
rejection loop and the two processes seeded that client differently.

### 3.3 Third process, with façade call-site attribution — which sites those words are

A third fresh process on `turn1-state`, with the two hooked façades recording their **AI call
sites** rather than the one address inside the façade:

| player | words | AI call sites (`airngcall`) |
|---|---:|---|
| 32 | 7 | `0x006ad94c` `cl_RandRange` ×2, `0x00691ea0` `cl_RandRange` ×2, `0x006adfca` `cl_Chance` ×1, plus `0x006ad878` `cl_RandFloat` ×2 |
| 496 | 0 | — |
| **512** | **1** | **`0x006a849a` `cl_RandRange` ×1** |

Two things fall out.

**Six of client 32's seven words come from the ship-design composer `0x006ad700`** — sites 13, 14
and 19 of §5. The seventh is a `cl_RandRange` in `0x00691e90`. So on this board the AI's entire
random consumption is *ship design* plus one other pick; nothing in the task system drew at all.

**Client 512's one word is the research-target tie-break.** `0x006a849a` is the return of site 10,
`0x006a8495` in `0x006a8390`, which is reached from **Process Turn phase 18** — the research-target
selection — and which is guarded by `cmp eax,[edi+4] / je return`: *the shortlist is empty or has
one member.* 496 takes that branch and draws nothing; 512 does not and draws once.

That is the mechanism behind lane L5's *"only one of three empires' research pick varies between
runs"* and lane L1's *"all three seeds move but only one empire's pick does"*, and it is now read
off the running game rather than inferred from outcomes: **two of the three empires never reach the
seed-sensitive branch at all.**

### 3.4 Pinning the seed: the same three words, and the oracle bytes unchanged

`airng.pin_seed=5A17C0DE` re-seeds each AI client's generator at bracket entry (`base + seq-1`, so
the three clients do not share a stream). On the canonical pair:

| player | idx in→out | words | sites |
|---|---|---:|---|
| 32 | 0 → 3 | **3** | `cl_RandFloat` ×1, `cl_RandRange` ×2 (`0x006ad94c`, `0x00691ea0`) |
| 496 | 0 → 0 | 0 | — |
| 512 | 0 → 0 | 0 | — |

**And the autosaves are still `bb4fd9ac…` / `978041ac…` — the oracle.** The AI's entire random
stream was replaced with a different one and the turn's output did not move by a byte. So on this
board every one of those three words is a draw *whose result never reaches the save*: two of them
pick among candidates that are equivalent as far as the recorded state is concerned, and one is the
composer's `cl_RandFloat` acceptance test. This is the first direct measurement of that phenomenon —
previous lanes could only observe that a *different* seed sometimes produced a different save.

It is also the sharpest available warning against the wrong conclusion: **"the count did not change
when the seed changed" is not evidence for roll parity.** The count is stable here because the
*path* was stable. §3.2 holds the path fixed and moves the seed, and the count moves.

### 3.5 An observation this lane could not resolve

All three `turn1-state` runs produced the same post-turn autosave, `d59bb9f2fd0eb535`, even though
client 512's stream demonstrably differed between them (1 word, 3 words, 1 word). Lane L5 reports
**three different files** from three runs of this pair on VM146, and `campaign/board.md` records
`turn1-state → turn2` as non-deterministic on that basis.

Both cannot be a complete description. Three identical outcomes from a k = 6 tie set (lane L4) is a
1-in-36 coincidence, so "it is deterministic after all" is not supported either. The honest
statement is that **the draw differs and the outcome did not**, on this guest, three times — and
that whatever maps the drawn word onto the chosen tech is not a plain uniform index into the
shortlist. Flagged, not resolved; it wants the tech id logged next to the draw, which is one line in
a probe on `0x006a8390`.

---

## 4. Why the count moves — the mechanism, read from the instruction stream

### 4.1 The primitives are not fixed-cost

`RNG_NextInt 0x004271c0`, whole body, `ret 4`:

```
4271c9  edi = *bound
4271cb  edi |= edi>>1 |= edi>>2 |= edi>>4 |= edi>>8 |= edi>>16    ; mask, computed ONCE
4271f0  LOOP: lazy twist if left==0; y = *next++; left--; temper
42723e        and eax,edi
427240        cmp eax,DWORD PTR [ebx]        ; re-reads *bound every pass
427242        ja  0x4271f0                   ; reject iff (y & mask) > bound   -- UNSIGNED
```

One back-edge, **no iteration counter and no bail-out**. `E[words] = (mask+1)/(bound+1)`, in
`[1, 2)`, and **exactly 1 only when `bound+1` is a power of two**. Almost every bound in the AI turn
is a container size minus one, so extra words are the ordinary case, not the exception. Measured:
one call costing 3 words (§3.2).

`RNG_Chance 0x008e6dd0` returns **without a draw** at `p <= 0` and at `p >= 1`; otherwise exactly
one word, no rejection. Two details for a reimplementation: the unit is stored to **float32** before
the comparison, and the accept test is **strict `p > r`** — the opposite convention from the inlined
site in `EncounterDetect_AssignContacts`, where equality accepts. Two conventions in one image.

`RNG_NextFloat 0x0047d830` is exactly one word, always.

**This alone refutes the hypothesis.** A design that drew unconditionally to stay in lockstep would
not use a rejection loop, because a rejection loop's cost is not knowable in advance.

### 4.2 The draws are conditional, and mostly do not fire

Twenty-three draw call sites are reachable inside an AI turn, in twelve AI functions; **two of them
are provably dead code** (`0x006adf5d`, `0x006adf6c`: the block is guarded by `cmp esi,ebx / jle`
with `esi` and `ebx` both zeroed on every inbound edge), leaving **21 live sites**. On the reference
turn, **two** fired. On turn 1, **three**.

Sites and their guards (full table in §5). The shapes that matter:

* **species gates** — Process Turn phase 2 returns immediately unless `ClientPlayer+0x5c == 1`
  (Hiver); phase 20's task creation is a seven-way species switch in which **species 4 creates
  nothing at all**; the helper `0x00694820` costs 1 word for species 0/4/5/6 and 1-or-2 for
  species 1/2/3;
* **empty-container gates** — nearly every `cl_RandRange` sits behind
  `cmp begin,end / je skip`. This is the dominant shape;
* **success gates** — the surrender roll at phase 12 gates phases "13" and "14", which are **not
  sequential phases at all** but the body of its success arm (a correction to
  `ai-turn-logic.md` §3.1). Six further draw sites live in there and only run on a turn where an AI
  actually rolls to surrender;
* **loop-carried draws** — `0x006ae575` is one `cl_Chance(0.2f)` **per qualifying empty section
  slot** of a ship design, inside a composer that is itself called 1–3 times per design request
  inside two outer loops. This is unbounded and it is where most of the variance lives.

### 4.3 The single most legible instance: 496 versus 512 on turn 1

`ai-order-emission.md` §2 establishes that all three AI players set a research target on turn 1.
496 and 512 own nothing — same `Sav = 0`, no colonies, no fleets, no systems. Yet:

* **496 draws zero words.** Its research pick is forced; the shortlist has one member, the
  `cl_RandRange` that would break a tie is behind `cmp eax,[edi+4] / je return` and never runs.
* **512 draws exactly one `cl_RandRange`.** Its shortlist has more than one member, so the tie is
  broken by a draw.

That is the mechanism behind lane L5's *"one of three empires' research pick varies between runs"*
and lane L1's *"all three seeds move but only one empire's pick does"*, and it is now a read of the
code rather than an inference from outcomes: **two of the three empires never reach the
seed-sensitive branch at all** — precisely the situation method rule 26's corollary warns about.

---

## 5. The draw-site table

Every RNG draw reachable inside `OnResumePlaying` for an AI client. `U` = unconditional once its
enclosing function is entered; `C` = guarded. Phase numbers follow `ai-turn-logic.md` §3.1.

| # | draw VA | enclosing fn | entry | reached from | U/C | guard |
|---:|---|---|---|---|---|---|
| 1 | `0x006c592a` | `0x006c57e0` | `cl_Chance(p)` | phase 2 | **C** | `Species == 1` (Hiver), an agent latch `+0x104`, and an empty-candidate test; `p` is live, so `Chance`'s early-outs apply |
| 2 | `0x006aeeed` | `0x006aee70` | `cl_RandRange` | phases 2, 14 via `0x006b23b0` | **C** | message-list vector empty |
| 3 | `0x006cc57e` | `0x006cb570` | `cl_Chance(clamp01(r))` | phase 7 | **C** | turn delta ≤ 15 returns; `p` clamped to `[0,1]`, so both zero-word early-outs are reachable |
| 4 | `0x006cfc24` | `0x006cf8a0` | `cl_Chance(min(1/T,0.5))` | **phase 12** | **C** | survival outlook false, or `T > 30`, or `p <= 0` |
| 5 | `0x00694830` | `0x00694820` | `cl_Chance(0.75f)` | phase 14 | **U** | first instructions of the function |
| 6–8 | `0x00694865/87/a6` | `0x00694820` | `cl_Chance(0.8f)` | phase 14 | **C** | species jump table: 1 / 3 / 2 respectively; species 0/4/5/6 draw nothing here |
| 9 | `0x006c3447` | `0x006c33e0` | `cl_RandRange` | phase 14 | **C** | `switch(arg)` — only arm 2 reaches it, and `arg` is the return of #5–8 |
| 10 | `0x006a8495` | `0x006a8390` | `cl_RandRange` | **phase 18** (research target) | **U** past 2 gates | `this == 0`, or the shortlist vector is **empty** |
| 11 | `0x006817ae` | `0x00681750` | `cl_RandRange(0,5)` | phase 20 | **C** | the **fallback** arm: only when no category scored |
| 12 | `0x0069086a` | `0x00690860` | `cl_Chance(0.5f)` | phase 20 | **U** | none — 4 instructions, `return chance ? 18 : 19` |
| 13 | `0x006ad873` | `0x006ad700` | **`cl_RandFloat`** | phases 20, 25 | **C** | an "ok" latch cleared by three tech lookups, and request-flag bit `0x40` |
| 14 | `0x006ad947` | `0x006ad700` | `cl_RandRange` | " | **C** | hull slot already filled, or zero candidates |
| 15 | `0x006adf30` | `0x006ad700` | `cl_Chance(0.5f)` | " | **C** | matching-section count ≤ 0 |
| 16 | `0x006adf3f` | `0x006ad700` | `cl_RandRange` | " | **C** | nested under #15's success |
| 17 | `0x006adf5d` | `0x006ad700` | `cl_Chance(0.8f)` | " | **DEAD** | `cmp esi,ebx / jle` with both zero on every inbound edge |
| 18 | `0x006adf6c` | `0x006ad700` | `cl_RandRange` | " | **DEAD** | same block |
| 19 | `0x006adfc5` | `0x006ad700` | `cl_Chance(0.8f)` | " | **U** in its block | both arms above converge before it |
| 20 | `0x006ae413` | `0x006ad700` | `cl_Chance(0.3f)` | " | **C** | flag bit `0x20000`, or a counter ≤ 0 |
| 21 | `0x006ae575` | `0x006ad700` | `cl_Chance(0.2f)` | " | **C, LOOP** | once per qualifying empty section slot |
| 22 | `0x006cce84` | `0x006ccdb0` | `RNG_NextInt` | phases 20, 25 (design names) | **C** | no name-gen object, a **cached** name hit, or an empty name pool |
| 23 | `0x0068bf03` | `0x0068bdc0` | `cl_RandRange(2,5)` | phase 20, **only** via `AITEscortGate::Execute` vtable slot 5 | **C** | a local bool; what sets it was not read |

**Three of twenty-one live sites are unconditional, and all three are unconditional only *given that
their function was called*** — which is itself gated. Not one draw in the AI turn is unconditional
with respect to the turn.

### 5.1 Sites on the same generator that are **outside** the turn

Listed so a later lane does not double-count them:

* **Prepare Turn only** (packet code 2, raised by `RaiseAIPrepareTurn 0x00815f20`, whose two callers
  `RunAI` and `CreateGame` are both one-shot): `0x006cfa06`, `0x0069620e`, and **three** draws in
  `0x0069dbb0` — `0x0069dbfd`, `0x0069dc29` and **`0x0069dcff`, which lane AI1's six-site list does
  not have**. `0x0069dbb0`'s guard `if (agent->+0x36c == 0)` is a true one-shot because the third
  draw **re-arms** the field to `turn + 20 + r`, which is never 0. These land at construction and
  never again — verified live: none of those return addresses ever appears in a bracket.
* **Net-message paths, not the turn:** `0x0069df9d` (agent vtable slot 3) and `0x006ca11c` (slot 4)
  are dispatched only from `StrategyNetworkClient::OnMessage 0x00784640`. *(This also corrects
  `ai-turn-logic.md` §2, which lists `0x0069de50` as the `AIObject` packet sink: it reads `[pkt+4]`
  as a 0/1/2 selector, while the four real `AIObject` sinks read `[pkt+0]` as the code — and none of
  those four draws anything.)*

The load-time cost is large and asymmetric. Measured with the lifetime census on the canonical pair:
the three AI clients arrive at their first turn already **17, 424 and 427** words into their
streams, and the human client's generator is at 0. `g_GlobalRNG 0x00af6e58` shows **3 words drawn,
one per AI client construction** — a second, independent falsification of
`ai-turn-logic.md` §5.1's "every draw from it returns 0", after lane L1's.

---

## 6. What is **not** established

1. **No inlined draw runs in an AI turn — and that is now measured, not assumed.** The
   `left_delta` vs `observed` residual was **0 on every bracket of every run**. Statically: an
   image-wide scan for the MT tempering immediates at real instruction boundaries reproduces lane
   I's 15 functions exactly, and **none is in the AI band**; the three nearest
   (`0x006ec720`, `0x006f65f0`, `0x006f7890`) are `StrategyNetworkServer` and sit above
   `0x006e0000`. Lane AI1's rule-16 caveat is discharged with the band bounds it asked for.
2. **The site inventory is a lower bound past two indirect hops.** `IAITask::Execute` (slot 5, all
   31 classes) and the `AIObject` packet sink (slot 3, all four) were resolved; a virtual call
   *inside* a task's `Execute` subtree is unmodelled. Weak evidence that it is close to saturated:
   adding all 31 `Execute` roots to the 31 phase roots added exactly **one** new draw function.
3. **Only two boards were measured**, both from the same 2-player-plus-shadows game, and no
   **species 5** or **Hiver** AI turn was measured. The species terms in §5 are read from the
   instruction stream and are rule-6 hypotheses live. A Hiver AI would exercise sites 1 and 6 and is
   the cheapest remaining workload.
4. **The dead-code claim on sites 17–18 is a static read.** No live hit has been observed, but no
   run has exercised the enclosing arm heavily either.
5. **`0x0068bdc0`'s `[ebp-0x1]`** — what sets it was not read.
6. **Whether `agent+0x36c` is serialised** in the `StrategyAIAgent::Streamable` block. If it is, the
   Prepare-Turn cost differs between a new game and a loaded save and the whole stream is offset.
7. **The number of ship-design requests per turn is unquantified.** The chain
   phase 20/25 → `0x006ce190` → `0x006cda40` → `0x006ae620` → `0x006ad700` was confirmed as edges
   and two enclosing loops, but the intermediate bodies were not read. Since the composer holds 9 of
   the 21 live sites including the only loop-carried one, **this is where the variance is**, and it
   is the first thing a follow-up should instrument (entry probe on `0x006ad700` logging `arg[3]`
   and `0x006ae620`'s loop counter).

---

## 7. What this means for `sots-engine`

**The cheap version of `game/ai` does not exist.** The redirect's hope was that a reimplementation
could stay in lockstep by reproducing the *count and order* of draws while getting the decisions
wrong. It cannot:

* every draw in the AI turn is behind a branch on game state, so a wrong decision changes **whether**
  a draw happens, not only its result;
* `RNG_NextInt`'s rejection loop makes the cost of a *single call* depend on the stream, so even a
  correct call sequence cannot be costed in advance — the engine must reproduce the rejection loop
  bit-for-bit (it already does; `mars/rng` is closed);
* `RNG_Chance`'s two zero-word early-outs mean a probability computed slightly wrong can change the
  word count from 1 to 0 and desynchronise everything after it.

The positives, which are real:

* **the AI draws from nothing but its own client's generator.** `foreign_words = 0` on every
  bracket of every run — a live confirmation of what `ai-turn-logic.md` §5 could only assert from a
  direct-call scan. The strategic generator and the AI are genuinely independent, so an engine can
  get the server's stream exactly right while the AI's is still approximate.
* **the human client draws nothing.** Every human bracket measured 0 words with `agent = 0`. There
  is no client-side RNG on the human path at all.
* **the per-turn cost is tiny** — single digits — so the AI's generator is not a large unknown; it
  is a small, sharp one.
* **the `cl_*` façade is a clean, complete, AI-only API**: three functions, 29 call sites, all in
  the AI band. An engine implements exactly those three and has covered every strategic-AI draw.

The consequence for the roadmap is the one lane L1 already reached from the seed side, now with a
mechanism: **Rung B (replay a recorded command stream) is unaffected — it does not run an AI.
Rung C (behavioural AI) needs the decisions, and the decisions live in the ship-design composer and
the task system**, neither of which is modelled.

---

## 8. Appendix — the *call*-parity question, and what this lane can say about it

This lane was briefed on call parity ("does the AI reach the simulation through the same interface a
human client does?") and redirected mid-flight to roll parity. The redirect is recorded because the
call-parity deliverable is **not complete**: the ~22 `hypothesis` rows in lane Q's list table were
**not** converted by driving the UI-reachable order methods, and no new order method was found. What
follows is only what this lane's own work establishes.

**Parity holds at the wire and apply layers, and it breaks at the call layer** — the reading the
brief proposed, and nothing measured here contradicts it. Two things are added:

1. **A new, clean piece of AI-only call surface: the `cl_*` RNG façade** (§2.1). Three functions —
   `cl_Chance 0x00578cf0`, `cl_RandRange 0x005798e0`, `cl_RandFloat 0x00579c70` — with 29 call sites
   between them, **all 29 inside the AI module**. Nothing else in the image draws through them. This
   sits alongside `cl_SetResearchTarget 0x00578f60` and `cl_EndTurn 0x00579310` (whose `+0x12c` test
   makes it AI-only by construction) as part of a coherent AI-only façade over "the current client".
   It is not order surface — it emits no `TurnCommands` element — but it *is* interface the engine
   must implement separately, and `cl_RandFloat` was not previously known to exist.

2. **A live confirmation that the human client has no client-side RNG at all.** Every human
   `OnResumePlaying` bracket measured 0 words with `agent = 0`. So the asymmetry at the call layer is
   not merely "the AI calls some methods the UI does not" — the AI has an entire *stateful* facility,
   a private seeded generator, that the human path never touches. `StrategyClient+0x134` exists on
   the human client too, and stays at word 0 for the whole game.

For `sots-engine` the answer to the original question is therefore: **one order path, two client
APIs.** The order methods and everything downstream of `StrategyClient+0x160` are shared and need
implementing once; the AI-only surface the engine must implement separately is the `cl_*` façade
(three RNG entries plus `cl_SetResearchTarget` and `cl_EndTurn`) and the four order methods with no
UI caller that lane AI1 P4 named (`0x00762ca0`, `0x00763670`, `0x00763720`, `0x00763990`) — with
lane AI2's caveat that its own sample was incomplete still standing, because this lane did not
extend it.

---

## 9. Files

- Instrument: `sots-engine` `wip/par` — `src/shim/hooks/ai_rng.{h,cpp}`, the `draw_sites` observer
  hook, `src/shim/shim.cfg.par{ctl,br,pin,orders}`
- Predictions, committed before the build: `sots-engine/docs/PAR-predictions.md`
- Raw brackets and autosave hashes: `verify/results/shim/airng/` — six runs,
  `par-{br-p1,br-p2}` (canonical pair, unpinned, two fresh processes),
  `par-pin-a1` (canonical pair, pinned), `par-{t1-br-p1,t1-br-p2,t1-br-p3}` (`turn1-state`, three
  fresh processes, the third with façade attribution), plus `parse_airng.py`, the reader
- Addresses: `ghidra/addresses.d/lane-par.json`