docs: the ledger went stale again within hours of the work that closed it (#28)

Three PRs landed today and none of them finished correcting the records the
next session plans from. An audit of docs-vs-repo found six blocking staleness
items, all of which would have caused over-scoping — which is exactly what an
out-of-date ledger caused at the start of this session.

## The six

- **§S19(d) read as open.** #26's generated floor closed its runtime half, by a
  mechanism the entry never considered: the fragments are seeded before the file
  is read, so no `ParseResult` cause is needed and `layer.cpp` is not driven
  inert.
- **§S1 and §S2 still asserted §S18 blocks launch mode and all Vulkan Tier 1**,
  in the present tense, one commit after #25 closed it. Both now say what
  actually blocks Vulkan Tier 1: `vkQueuePresentKHR`, which is P1.
- **`19_SAFETY` still described the pre-#26 three-family floor** as a deliberate
  choice, with the reasoning that #26 disproved.
- **`CHANGELOG` §Known issues** listed §S20 as "nothing installs the rules file"
  and claimed the runtime parser has no fragment floor. Both were fixed today.
- **`15_ROADMAP`** omitted §S20 entirely and said `Program.cs` is `return 0`.
- **`spike-notes` exit criteria** still listed the FPS-impact gate and M3/M4 as
  unmet, and asked for a decision two other documents recorded on 2026-08-02.

## spike-notes gains §12 — the measurements had nowhere to live

`15_ROADMAP` names this file as where P0 writes its answers, and the whole of
today's work was recorded only in commit messages. Now in the file: the §S21
override before/after on two real binaries (`Allow` -> `BlockedModule`), the ANSI
path against the real ACP, what the first floor actually covered (4 of 22 values),
`MoveFileEx` vs `ReplaceFileW` against a live reader, the `rulesVersion` history
that falsified the seeder's update trigger, and the four seeder branches measured
on the machine's real rules location.

Two of them are records of being wrong, kept because the file's own rules ask for
dead ends: an ACP round-trip measured with `[Text.Encoding]::Default` (always
UTF-8 in .NET Core, so it reported no loss), and a canary harness that produced
three confident greens while selecting no tests at all.

## Two things the audit found that were nobody's follow-up

- `19_SAFETY` said `FlStaticPreScan` exposes the pre-scan to the UI. Since #25
  narrowed the guard DLL to the Agent, `FrameLedger.App` has neither a call site
  nor a DLL to load. Corrected in place rather than deleted.
- §R10, new: four artifacts the documents name and the repository does not have —
  `tools/resx-audit`, the managed struct-mirror test, `release.yml`, and three
  gates `build.ps1` runs that `12_BUILD` omits. The struct-mirror row is the one
  that matters: two docs state `dotnet test` includes it, so it reads as a gate
  that runs.

Also re-counted the fragment list's copies while correcting the claim that the
floor reconciled them. It did not, and there are five, not three — including a
`$comment` in the schema carrying the four-fragment version, the exact staleness
§S19(e) was raised about.

Docs only; `build.ps1 check` passes with the standing resx-audit skip.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
poli0981andClaude Opus 5 authored and GitHub committed 2026-08-04 17:32:43 +07:00
1 parent 1ed07c08c6
commit 71f4e33e01
5 files changed
+260 -57

No files matched your search

+26 -9
View File
@@ -42,15 +42,32 @@ GitHub release body, so a missing section means an empty release note.
for.
- Also recorded: `signerField` and `action` are required by the schema and read
by no code (`action` is a `const` with one legal value, so reading it could
change nothing); the fragment list exists in three unreconciled copies; and
the runtime parser has no fragment floor — though the **schema** does, so CI
already refuses an emptied list, which an earlier note here got wrong.
- **Nothing installs the rules file the guard reads** (§S20). The guard reads
`detection-rules.json` under Local AppData; no code in the repository seeds,
copies or downloads it, and the guard never reads `rulesVersion`, so a binary
fix and a data fix have no handshake. It is a **two-consumer** problem — the
same missing file also fails `DetectionRulesFile`, and with it the static
detector P0 item 3 shipped.
change nothing).
- **The runtime fragment floor is now in place** and this entry no longer claims
otherwise. The compiled-in floor is generated from `rules/detection-rules.json`
at build time, so a rules file with no `heuristic` block can no longer make the
fuzzy tier stop existing.
- **The unreconciled copies are not fixed by that**, and there are more than the
three this entry counted. The generated floor is *derived* and cannot drift,
but `guard_test.cpp`, `rules_budget_test.cpp`, the prose in `19_SAFETY`
§Heuristic tier and a `$comment` in `detection-rules.schema.json` each still
restate the list by hand — and the schema comment restates the **four-fragment**
version, the exact staleness §S19(e) was raised about. No gate cross-checks any
of them.
- **A rules edit reaches no installed machine until a release** (§S20, feed half).
The Agent now installs `detection-rules.json` to the location the guard reads,
so a machine that has never had one no longer refuses every title — that half is
done and measured. What does not exist is `05_DETECTION` §Trust and staleness'
HTTPS fetch, so **FR-7.3 is unmet**: anti-cheat entries cannot arrive on their
own schedule. The guard also still never reads `rulesVersion` or `schemaVersion`,
so a binary fix and a data fix have no handshake — deliberately, since teaching
it to refuse an unknown version would be a second machine-wide refusal lever
pulled by data.
- **`trustedSigners` is the one allow-widening field a foreign rules file still
controls.** Deliberately not floored — flooring an allowlist has the wrong
polarity — and inert today, because `IsTrustedSigner` has no production call
site while §S19(b) is deferred. It becomes live the moment the signer half is
wired, which makes gating it a prerequisite of that work.
### Added
- Design documents (`CLAUDE.md`, `docs/01`–`20`, `legal/`) and the repository
+20 -9
View File
@@ -21,11 +21,21 @@ Findings written to `docs/spike-notes.md`. Nothing in P1 starts until the exit c
> absent hardware (8, and the AMD/Intel half of the capability matrix), or is P1
> by construction (the layer's presentation hooks). Item 0's residual is check 3.
>
> **§S18 and §S21 closed 2026-08-04.** The guard no longer refuses itself, so
> launch mode is no longer blocked by it — and §S21, found while planning this
> phase, closed a **local override of the hard gate**: the rules path came from an
> inherited `LOCALAPPDATA` and the completeness check never read the values, so a
> crafted twelve-line file made the guard allow everything.
> **Three safety items moved 2026-08-04.**
>
> - **§S18 ✅** — the guard no longer refuses itself, so launch mode is no longer
> blocked by it. Measured on three real titles.
> - **§S21 ✅** — a **local override of the hard gate**: the rules path came from
> an inherited `LOCALAPPDATA` and the completeness check never read the values,
> so a crafted twelve-line file made the guard allow everything. Its first fix
> shipped too narrow — the compiled-in floor held 4 of the seed's 22 values —
> and was rebuilt the same day as a table **generated** from the shipped
> blocklist. That also closed §S19(d)'s runtime half.
> - **§S20 ◐, seed half done** — the Agent now installs the rules file. Until
> then the guard answered `RulesUnreadable` on any machine that had not
> hand-installed one, which is what the first real injection hit. **The feed
> half is open**, so FR-7.3's independent anti-cheat schedule is still unmet and
> a rules edit reaches no installed machine until a release.
>
> **Be precise about what unblocking §S18 does and does not buy**, because the
> §S18 entry oversold it and this line used to repeat that. It removes a
@@ -68,10 +78,11 @@ Findings written to `docs/spike-notes.md`. Nothing in P1 starts until the exit c
>
> **Launch mode is still not working, and that is a different sentence.** The
> guard now passes, which is all §S18 was about. Launch-mode *injection* also
> needs §S1 and §S13(c), both open owner decisions, and there is no Agent to
> drive it — `Program.cs` is `return 0`. This row therefore stays "attach mode"
> for the *injection* claim; what is upgraded is the guard's verdict, which is
> the only thing that was measured.
> needs §S1 and §S13(c), both open owner decisions, and there is no Agent
> capture path to drive it — `Program.cs` gained a rules seeder the same day
> (§S20) but still builds no host, no watcher and no injector control. This row
> therefore stays "attach mode" for the *injection* claim; what is upgraded is
> the guard's verdict, which is the only thing that was measured.
>
> Three refusals came before the success, and each was a real defect: a
> machine-wide `EasyAntiCheat_EOS` service that refused every process on the
+39 -11
View File
@@ -226,10 +226,18 @@ Implemented in `FrameLedger.Injector` and reached from managed code through a th
> nothing produced either.
>
> **It runs inside the chokepoint**, as the last of the four checks in
> `EvaluateImpl`. `FlStaticPreScan` also exposes it to the UI for FR-2.2's
> pre-launch question, but that export is **advisory**: it gates nothing, and
> a caller who never asks it — or ignores the answer — changes nothing about
> what injection allows.
> `EvaluateImpl`. `FlStaticPreScan` also exports it for FR-2.2's pre-launch
> question, but that export is **advisory**: it gates nothing, and a caller who
> never asks it — or ignores the answer — changes nothing about what injection
> allows.
>
> **The UI cannot reach that export today**, and this sentence used to say it
> could. Since 2026-08-04 `FrameLedger.Guard.dll` ships beside the **Agent
> only** (`FrameLedger.Guard.targets`, §S18 blocker 3), so `FrameLedger.App`
> has no `FlStaticPreScan` call site and no DLL to load. FR-2.2's advisory
> answer reaches the UI through the Agent over IPC (`07_IPC`) when that exists.
> Recorded rather than quietly reworded: an advisory export with no consumer is
> a claim about the product that is not true yet.
>
> **Which directory.** The **install root**, resolved by walking up from the
> executable to a known platform boundary (`steamapps\common\<X>` and
@@ -295,9 +303,9 @@ install directory, nothing left behind.
**The fix is §S8's mechanism applied to data instead of to symbols.** A token
that escapes can be ignored; a symbol that does not exist cannot be called; a
family that data cannot remove cannot be bypassed. `fl::guard::FloorFamilies`
carries those three families **inside the binary**, `ParseRules` seeds them
before it reads a byte of the file, and nothing merges, rewrites or removes
them. The file may add families and values. It can take nothing away.
carries the blocklist **inside the binary**, `ParseRules` seeds it before it
reads a byte of the file, and nothing merges, rewrites or removes it. The file
may add families and values. It can take nothing away.
Three consequences worth stating rather than discovering:
@@ -309,10 +317,30 @@ Three consequences worth stating rather than discovering:
the merged set would make it a gate that cannot fail, since the floor
satisfies it by construction — so an empty or incomplete rules file still
refuses, and still says so.
- **The floor is deliberately three families, not the whole blocklist.** A
larger one would be a second blocklist drifting from
`rules/detection-rules.json`; ctest `fl_rules_budget` asserts every floor
value is present in the shipped seed under the same family and group.
- **The floor is GENERATED from `rules/detection-rules.json` at build time**
(`tools/gen-ac-floor.ps1`), so it is the whole shipped blocklist plus the
heuristic's name fragments.
> This paragraph used to read "the floor is deliberately three families, not
> the whole blocklist", kept small because a larger hand-written table would be
> a second blocklist drifting from the data. Measured, that bought **4 of the
> seed's 22 values, 2 of its 5 groups and none of its 5 fragments** — so it
> closed *"a crafted file makes the guard allow everything"* and left open
> *"a crafted file removes most of the blocklist"*. Generating it removes the
> objection that kept it small: a table derived from the data cannot drift from
> it. `trustedSigners` is deliberately excluded, because flooring an
> ALLOW-widening list has the wrong polarity.
>
> A file family identical to a floor entry is deduplicated, so the file's own
> budget is half `kMaxFamilies` — `tools/rules-validate.ps1` checks that worst
> case, and ctest `fl_rules_budget` asserts the generated floor reproduces the
> shipped seed exactly.
- **The fuzzy tier is floored too**, which is what stops a rules file with no
`heuristic` block from making it silently cease to exist. That was
`20_OPEN_QUESTIONS` §S19(d), and a floor closes it without the new
`ParseResult` cause that entry proposed — which would have made
`kRulesIncomplete`'s signal a lie and driven `layer.cpp` to machine-wide inert
passthrough.
### Blocklist seed
+62 -18
View File
@@ -11,8 +11,13 @@ Rules for this document:
rephrased. The answer goes in the owning doc, and the entry here is deleted
with a pointer in the commit message.
- Every S-series item is a **safety** item and blocks the first real injection.
- Nothing here is a known bug in shipped code — there is no code yet. That is
the point: these are cheap now and expensive in P1.
- ~~Nothing here is a known bug in shipped code — there is no code yet.~~ **No
longer true, and the change matters for how this file reads.** There is a built
guard, a Vulkan layer, a rules seeder and 11 registered ctests, and §S21 records
a fail-open found in *shipped* guard code. Entries here are now a mix of open
questions and recorded defects.
- **A ✅ entry is closed.** They are kept rather than deleted, with their
reasoning, so counting entries overstates the open work — read the markers.
---
@@ -42,7 +47,10 @@ Each item below is a place where the documents themselves leak a gap.
> the session goes `Detected → Guarded → Capturing` without ever entering
> `Injecting`. Vulkan Tier 1 is nevertheless launch-mode-only
> (`17_HOOK_ENGINE:161` — only the launching process can set
> `FRAMELEDGER_ENABLE_VK_LAYER=1`), so **§S18 is its sole blocker**.
> `FRAMELEDGER_ENABLE_VK_LAYER=1`), and §S18 was its sole blocker.
>
> **§S18 closed 2026-08-04, so that sentence is history.** What now blocks
> Vulkan Tier 1 is `vkQueuePresentKHR`, which is P1 and not started.
`04_CAPTURE` §Launch mode prefers `CreateProcess(CREATE_SUSPENDED)` → guard →
inject → `ResumeThread`. But a suspended process has **loaded no modules yet**,
@@ -90,10 +98,13 @@ the DLL — and it compares the variable's *value*, so a stray `=0` does not
enable us. The cost is accepted: **Vulkan Tier 1 is now launch-mode-only.**
`tools/vklayer-blastradius.ps1` runs the check and unregisters in a `finally`.
> **That cost is larger than it read.** Launch mode is blocked by §S18, and §S1
> does *not* cover the Vulkan path (no injection, no suspended target), so §S18
> alone blocks every Vulkan Tier-1 session. Recorded 2026-08-03; it is why §S18
> is not merely "launch mode, which is blocked anyway".
> **That cost read larger than it was.** Recorded 2026-08-03: launch mode was
> blocked by §S18, and §S1 does *not* cover the Vulkan path (no injection, no
> suspended target), so §S18 alone blocked every Vulkan Tier-1 session — which is
> why it was not merely "launch mode, which is blocked anyway".
>
> **§S18 closed 2026-08-04.** The remaining blocker is the layer's own
> `vkQueuePresentKHR`, P1 and not started.
> **The measurement also killed the design that looked obvious.** Declining
> `vkNegotiateLoaderLayerInterfaceVersion` for a process that did not opt in
@@ -756,9 +767,14 @@ configurable and never was; the field that would express it has no consumer.
> is the defect class this file exists to record. Wiring it means first deciding
> that `warn` or `allow` may exist, and CLAUDE.md rule 2 says they may not.
**(d) The fragment list has three unreconciled copies, and the RUNTIME parser has
no floor.** `rules/detection-rules.json`, `guard_test.cpp` (the inline
`GoodRulesJson()` fixture) and `rules_budget_test.cpp` each carry their own.
**(d) ◐ The fragment list has unreconciled copies — more than three. The runtime
hole is CLOSED.** `rules/detection-rules.json`, `guard_test.cpp` (the inline
`GoodRulesJson()` fixture) and `rules_budget_test.cpp` each carry their own, and
re-counting on 2026-08-04 found two the entry had missed: the prose in
`19_SAFETY` §Heuristic tier, and a `$comment` in `detection-rules.schema.json`
that restates the **four-fragment** version — the exact staleness §S19(e) was
raised about, sitting in the schema that gates the data. The generated floor is
not a copy: it is derived at build time and cannot drift.
`ParseRules` guards the entire heuristic read behind `heurTok >= 0` and
`IsCompleteEnoughToGate` never looks at fragments, so a rules file with no
`heuristic` block parses `kOk` and the tier silently stops existing.
@@ -776,9 +792,12 @@ no floor.** `rules/detection-rules.json`, `guard_test.cpp` (the inline
>
> What genuinely remains is narrower and sharper:
>
> - **The runtime hole**, above. It matters only for a file that did not come
> through CI — which is to say, only once §S20 gives one a way to reach a
> machine. That is why this is sequenced with §S20 and not before it.
> - ~~**The runtime hole**, above, sequenced with §S20.~~ **Closed 2026-08-04 by
> the generated floor**, and by a mechanism this entry did not consider: the
> fragments are seeded into `Rules` before the file is read, so the tier cannot
> stop existing because the data never supplied it. That needs no new
> `ParseResult` cause, so `kRulesIncomplete`'s signal stays true and `layer.cpp`
> is not driven inert. See §S21's floor note.
> - **The schema canary does not discriminate on this constraint.** It is
> `{"schemaVersion":"not-a-number"}`, which any schema still pinning
> `schemaVersion` rejects. Delete `minItems` from `nameFragments` and the canary
@@ -1023,11 +1042,14 @@ FILE_SHARE_DELETE`. The guard denied delete sharing and the layer denied nothing
> the two `layer.cpp` comments that documented measured-wrong designs: a reader
> designs a gate around them.
Proven red, all four ways, before being called done: emptying `FloorFamilies`
makes the disarmed-rules test allow; pointing a floor value at something the seed
does not carry fails `fl_rules_budget` by name; running the completeness check
from index 0 lets a file with no BattlEye through; and renaming
`kFloorFamilyCount` makes `rules-validate.ps1` **fail rather than skip**.
Proven red before being called done. Two of the four canaries named here were
rewritten when the floor became generated — `kFloorFamilyCount` no longer exists
and no floor value can be hand-pointed — so the current set is: a generator that
drops four of the five groups, a floor keeping one of five fragments, a
`kMaxFamilies` below twice the seed, and a `kMaxNameFragments` below twice its
fragments. Emptying `FloorFamilies` still makes the disarmed-rules test allow, and
running the completeness check from index 0 still lets a file with no BattlEye
through.
#### Measured end to end, not only in fixtures (2026-08-04)
@@ -1382,3 +1404,25 @@ no drain) stays in P0. Items 5–9 below are still open.
`09_I18N` fails the build until a human signs off on safety translations.
Identify the reviewer during bootstrap and draft the `Safety_*` keys as soon
as the consent wording is stable.
10. **Four artifacts the documents name do not exist, and one of them is claimed
as a gate that runs.** Found 2026-08-04 by an audit of the repository's own
status records, and grouped here rather than fixed one by one because the
decision is the same for all of them: build it, or stop describing it as
present.
| Named where | Artifact | State |
|---|---|---|
| `12_BUILD:139`, `13_CI_CD:11`, `09_I18N:28`, `CLAUDE.md:66` | `tools/resx-audit` | Absent. **Honestly handled** — `build.ps1:309` skips it loudly with a reason, and no `.resx` files exist yet |
| `12_BUILD:136`, `13_CI_CD:11` | the managed **struct-mirror** test | Absent. Nothing under `tests/` references `FlFrameRecord`. Both docs state `dotnet test` includes it, so this one reads as a gate that runs |
| `13_CI_CD:21`, `12_BUILD:121`, `CHANGELOG:9` | `.github/workflows/release.yml` | Absent. `CHANGELOG`'s header instructs an author to write for a consumer that does not exist |
| `12_BUILD` §Local quality gate | three gates omitted from the list | `build.ps1` runs **13** steps; the doc lists 10, missing `coverage-gate`, `versioninfo-check` and `chokepoint-check` |
The struct-mirror row is the one that matters. `CLAUDE.md` §Struct mirroring
makes that test the mechanism protecting the shared-memory ABI, and a doc that
says a gate runs is worse than a doc that says it is missing — the second
prompts someone to write it. **`build.ps1`'s skip-loudly discipline is the
right answer here**: a gate that is not written should be named and skipped,
not omitted.
Not urgent — the ring is P1 and there is nothing to mirror yet. Recorded so
it is found by reading, not by trusting.
+113 -10
View File
@@ -1,7 +1,7 @@
# P0 spike notes
P0's named deliverable (`15_ROADMAP` §P0). **Empty by design** — this file is
where measured results go, not predictions.
P0's named deliverable (`15_ROADMAP` §P0). Where measured results go, not
predictions. (It began empty by design and is now the longest record here.)
Rules:
@@ -663,7 +663,8 @@ Record the SDK version each name came from.
hooked yet**, so this clears the loader, not the unhook path; P1's fault
policy and unhook still need the same check once `Present` is hooked.
- Reached through the shipped C ABI (`FrameLedger.Guard.dll`), so there was no
path that skipped the gate. There is still no injector CLI (§S9).
path that skipped the gate. There is no injector CLI and there will not be —
§S9 closed that as a design decision, not as a gap.
**The three refusals that came first are the valuable part.** Every one was a
real defect, and none would have surfaced without a real machine and a real
@@ -676,7 +677,7 @@ title:
| Same, from the launching shell | `SuspiciousUnsigned` `FrameLedger.Guard.dll` | **Defect, open.** Our own DLL trips our own `guard` name fragment. See below. |
| Same, from a non-ancestor process | `Allow` → injection refused | Correct, and now legible: **Deadly Heart Gambit is x86**, and the Overlay is x64-only. This fills `14_TESTING`'s manual-matrix row for a 32-bit title. |
### 🔴 Open — the guard refuses itself in launch mode
### ✅ Closed — the guard refused itself in launch mode *(fixed 2026-08-04)*
Isolated cleanly on the same title, same machine, same rules:
@@ -995,6 +996,104 @@ construction. So these rows are blocked on P1, not on finding a game.
- 2.x column set over stdout:
- Performance Log Users sufficient without admin:
## 12 · The gate's own inputs — §S21 and §S20 *(2026-08-04)*
Analysis and the decisions live in `20_OPEN_QUESTIONS` §S21 and §S20. What
belongs here is what was **run**.
### ✅ §S21 · The override, demonstrated end to end on two real binaries
Same input to both — a crafted twelve-line rules file naming the three families
`IsCompleteEnoughToGate` required, with values that match nothing, plus a real
`EasyAntiCheat_x64.dll` loaded in the target:
| Binary | Verdict |
|---|---|
| `bb0da0a`, the tree before the fix | **`Allow`** |
| The fix branch | **`BlockedModule`**, family "Easy Anti-Cheat" |
No admin, no write to our install directory, nothing left on disk. This is the
one measurement to keep from the whole item: the hard gate could be turned off by
an inherited environment variable and a text file.
### ✅ §S21 · The ANSI path, on the real Win32 calls (system ACP 1252)
A profile component the system code page cannot spell exists on disk and
`CreateFileW` opens it, while `CreateFileA` fails `ERROR_INVALID_NAME (123)`.
Under an ASCII profile the same file opens both ways — which is exactly why a dev
box cannot see this, and why it would have shipped.
**A measurement of my own that was wrong first.** The initial round-trip used
`[Text.Encoding]::Default`, which in .NET Core is **always UTF-8** and not the
process ANSI code page, so it reported no loss. Re-run against the real ACP, it
lost the characters. *Measuring the wrong API is indistinguishable from measuring
a working one.*
### ✅ §S21 · What the first floor actually covered
Measured against the shipped seed, the hand-written three-family floor held
**4 of 22 values, 2 of 5 groups, 0 of 5 name fragments** — so it closed "a crafted
file allows everything" and left "a crafted file removes most of the blocklist"
open. Recorded because the write-up read as though it bounded more than it did.
The floor is now generated from `rules/detection-rules.json` at build time and
carries all of it.
Two residuals measured rather than asserted:
- The known-folder path **is** still user-relocatable — the `User Shell Folders`
key under `HKCU` is FullControl for the user. That is a persistent change
affecting every application, not a per-launch variable, so the fix is a
narrowing and is written down as one.
- Moving the layer to `SHGetKnownFolderPath` added `ole32` and `SHELL32` to its
imports. Resident in **62%** and **61%** of inspectable processes respectively,
so the layer takes on no meaningful new load-time surface.
### ✅ §S20 · `MoveFileEx` vs `ReplaceFileW` against a live reader
`MoveFileExW(MOVEFILE_REPLACE_EXISTING)` returns **`ERROR_ACCESS_DENIED`** when
the destination is open by a reader holding the guard's exact share mode
(`READ|WRITE|DELETE`). `ReplaceFileW` succeeds. **This repository had prescribed
the failing one** — in a source comment and in the ledger — as guidance for
whoever built the seeder. A prescription nobody ran is a guess in the imperative
mood.
### ✅ §S20 · `rulesVersion` does not track the blocklist
The first seeder design replaced the installed file when the packaged seed was
strictly newer. Measured against this repository's own history of that file:
`31825cf` and `ba5355e` both changed the `anticheat` block **without** touching
`rulesVersion`, and `a4a2c63` — the only commit that bumped it — changed the block
**not at all**. The rule would have delivered none of the two changes it existed
for. The seeder records a hash of what it installed instead.
### ✅ §S20 · The seed half, on the machine's real rules location
Remove the file → the guard answers `RulesUnreadable`. Run the Agent → the guard
reads its rules and reaches check 1. Four Agent branches exercised against the
real location, with the machine's own file backed up and restored by hash:
already-current, installed-from-nothing, idempotent second run, corrupt file
replaced.
The feed half is untouched, so a rules edit still reaches no installed machine
until a release.
### 🔴 The canary harness lied three times before it caught anything
Kept because this file's rules ask for dead ends, and because this one produced
three confident greens in one session:
1. `ctest`/Catch2 with `-c "*name*"` **selects nothing and exits 0**, so every
"the canary passed" was vacuous.
2. The rewritten check matched only Catch2's **failure** summary, so a genuinely
green run parsed as "the filter selected nothing".
3. `build.ps1`'s throw escaped the harness and **left a canary edit on disk** —
the next build then failed on an unrelated `/W4 /WX` unused-parameter warning,
which was nearly diagnosed as the canary working.
Now: tag filters, an assertion that assertions actually ran, and `try`/`finally`
around the build. A canary that cannot go red is the defect it was written to
catch, one level up.
---
## Exit criteria
@@ -1005,12 +1104,16 @@ construction. So these rows are blocked on P1, not on finding a game.
reporting **correct** upscaler, quality preset, render → output
resolution, FG factor and RT state — verified against the game's own
settings menu.
- [ ] Measured game FPS impact ≤ 0.5%.
- [x] **Moved to the end of P1** (§R4, decided 2026-08-02) — ~~measured game FPS
impact under 0.5%~~. As written it imported P2's drain, aggregate and
recorder paths; the harness-level per-present cost stayed in P0 and is
measured in §3 (8.4 ns).
- [ ] Every S-series item in `20_OPEN_QUESTIONS.md` resolved, or explicitly
deferred with a written rationale.
- [ ] M3 and M4 answered **before** any NVAPI or LHM code is written.
- [x] M3 and M4 answered **before** any NVAPI or LHM code is written — both
CLEAR, §0 above.
> Note (`20_OPEN_QUESTIONS` §R4): the first two criteria import P2 work — a real
> session needs the drain, aggregate and recorder paths. Either scope a
> throwaway drain harness into P0 or move the FPS-impact criterion to the end of
> P1. Decide before starting, not halfway through.
> ~~Note (§R4): … Decide before starting, not halfway through.~~ **Decided
> 2026-08-02** — the FPS-impact criterion moved to the end of P1, and the
> harness-level per-present cost stayed in P0 as item 2. This file went on asking
> for a decision that two other documents had already recorded.