docs(s31): the measurement ran, PresentMon is retired, and the question it was asked is still open (#95)
§S31 ran on 2026-08-27: three legs, three game launches, Cyberpunk 2077 at frame generation off / ×2 / ×4, driven by tools/s31-capture.ps1. It landed on the pre-committed row P2. FrameType is present and EVERY ROW OF ALL THREE LEGS READS `Application` -- 1,937 / 6,488 / 10,881 rows, no other value anywhere. Counted off the raw CSVs rather than off the tool's own summary, because a falsifier read out of a summary is a falsifier nobody checked. THE HALF THAT MAKES IT A MEASUREMENT RATHER THAN A SHRUG. P2 needs the title to have been demonstrably generating frames, and the two instruments agree on the present rate to within 0.3% on every leg: 43.16 vs 43.04, 144.31 vs 144.18, 241.84 vs 241.80 present/s. They are counting the same stream. Displayed rate tracks the configured multiplier. At ×4 roughly three presents in four cannot be application frames -- and PresentMon classifies 100% of them as application frames. tools/frametype-oracle.ps1 reported this as a FALSIFIER on all three legs and never as a ratio of 1.0. That is the one behaviour it was written to guarantee, and it is the first time it has been fed real input. WHAT P2 RETIRES: PresentMon 2.x FrameType as the application-frame oracle for NVIDIA frame generation. 03_METRICS rung 2 is narrowed with it -- the honest scope is now "no vendor is known to instrument this provider, and one is now known NOT to". Intel's own is still untested, which is the reading this document was already corrected once for asserting. WHAT IT DOES NOT ANSWER, AND THIS IS THE PART THAT WILL BE READ WRONG. §S31's actual question -- is a drained Streamline batch an application frame? -- is exactly as open as it was. presents/batch still reads 2.00 and 3.99 against a title's own ×2 and ×4, still on the premise that the work is recorded once per application frame, and still unpublishable as fg_factor for that reason. What was retired is the INSTRUMENT. Four oracles have now fallen: fl-baseline-probe by its own falsifier, two readings of Steam's overlay, and this one. The entry moves ❓ -> ◐ on exactly that basis. HANDOFF item 3 goes back to the hook routes, which is where P1/P2 said it would go. The decision it has been waiting on is now UNBLOCKED AND UNMADE, and whoever takes it is choosing without an oracle -- which the entry now says, so the next PR states it rather than implying one. ONE SUB-QUESTION IS UNMEASURED AND IT CHANGES THE REASON, NOT THE ACTION: whether --track_frame_type was in effect at all. If FrameType ships in --v2_metrics output regardless of the flag, these legs recorded a default rather than an absence. Both branches end at "PresentMon did not answer the question as invoked", and an oracle that does not answer is retired rather than re-run until it does. The discriminator is one five-second run without the flag, no game and no launch spent; it needs elevation, and §S31 carries the command. Attempted unelevated and refused with the same exit 6 this item has been blocked on from the start. TWO FINDINGS NOBODY WAS LOOKING FOR. The `off` leg drained ZERO Streamline batches, against presents/batch = 1.000 at off in §8. §S30 established that Ray Reconstruction is evaluated once per application frame, so with RR on there are batches even with FG off. RR was not on here. Our side therefore had TWO readable legs, not three -- turn RR on if the off leg needs to carry a ratio. And the CSV's FIRST column is named `Application`, holding the process name, while the value being counted in column 9 is also the string `Application`. A parser resolving by position, or grepping for the word, counts process names and reports a ratio from them. "Columns are resolved BY NAME from the header, never by position" was written into the oracle before it had ever seen a real CSV, against exactly the collision that was waiting in it. Recorded because it looked like ordinary care when it was written and was load-bearing. §M2's 2.x column set is measured at last -- 24 columns, listed. kFeatureDLSS_G is zero on all three legs, taking that to five capture sets across five titles. Artifacts live in %LOCALAPPDATA%\FrameLedger\s31\{off,x2,x4}\ and are not committed: they are measurement output, and the numbers in spike-notes §11 are the record. ./build.ps1 check with no switches: PASSED, one skipped gate (resx-audit, which does not exist). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
1 parent
1c41dc4e80
commit
046ec256f8
6 files changed
+256
-11
No files matched your search
@@ -19,6 +19,43 @@ GitHub release body, so a missing section will mean an empty release note.
|
||||
|
||||
### Changed
|
||||
|
||||
- **§S31 ran, and PresentMon is RETIRED as the application-frame oracle — row P2.**
|
||||
Three legs, three game launches, Cyberpunk 2077 at off / ×2 / ×4. `FrameType` is present and
|
||||
**every row of all three legs reads `Application`** — 1,937 / 6,488 / 10,881 rows, no other
|
||||
value anywhere, counted off the raw CSVs rather than off the tool's summary.
|
||||
|
||||
**The half that makes it a measurement rather than a shrug:** the two instruments agree on the
|
||||
present rate to within **0.3% on every leg** (43.16 vs 43.04, 144.31 vs 144.18, 241.84 vs
|
||||
241.80 present/s). They are counting the same stream, displayed rate tracks the configured
|
||||
multiplier, and at ×4 roughly three presents in four cannot be application frames — while
|
||||
PresentMon classifies **100%** of them as application frames. `tools/frametype-oracle.ps1`
|
||||
reported it as a **falsifier** on all three legs and never as a ratio of 1.0, which is the one
|
||||
behaviour it was written to guarantee.
|
||||
|
||||
**What P2 retires and what it leaves untouched.** Retired: PresentMon 2.x `FrameType` as the
|
||||
oracle for NVIDIA frame generation, and `03_METRICS` rung 2 is narrowed — the honest scope is
|
||||
now *no vendor is known to instrument this provider, and one is known NOT to*. **Not answered:**
|
||||
§S31's actual question. *Is a drained Streamline batch an application frame?* is exactly as open
|
||||
as it was; `presents/batch` still reads 2.00 and 3.99 on the same unverified premise and is
|
||||
still unpublishable as `fg_factor`. **Four oracles have now fallen.** `HANDOFF` item 3 goes back
|
||||
to the hook routes — unblocked, unmade, and now without an oracle behind whatever is chosen.
|
||||
|
||||
**One sub-question is unmeasured and it changes the REASON, not the action:** whether
|
||||
`--track_frame_type` was in effect at all. If the column ships by default, the legs recorded a
|
||||
default rather than an absence. Both branches end at *PresentMon did not answer as invoked*,
|
||||
which is why the row landed without waiting. The discriminator is one five-second run without
|
||||
the flag; it needs elevation, and it is recorded in §S31 with the command.
|
||||
|
||||
**Two findings nobody was looking for.** The `off` leg drained **zero** Streamline batches,
|
||||
against `presents/batch = 1.000` at off in §8 — Ray Reconstruction was not on, and §S30
|
||||
established RR is what produces batches when FG is off, so our side had two readable legs and
|
||||
not three. And the CSV's **first column is named `Application`** while the value being counted
|
||||
in column 9 is also `Application`: a parser resolving by position, or grepping for the word,
|
||||
counts process names. That rule was written into the oracle before it had ever seen a real CSV.
|
||||
|
||||
§M2's column set is measured at last — 24 columns — and `kFeatureDLSS_G` is zero on all three
|
||||
legs, taking that to five capture sets across five titles.
|
||||
|
||||
- **Five real-title captures across four titles, and the RT tri-state is complete** —
|
||||
`spike-notes` §6 and §8 carry the numbers; §8's per-title table was empty and now has five
|
||||
rows. Every run: 40 s, one swapchain, one segment, **0 gaps, 0 dropped**, `faults=0`, all six
|
||||
|
||||
@@ -89,6 +89,30 @@ This is the metric the rewrite exists for. Resolution ladder, highest confidence
|
||||
> the rung outright. `tools/frametype-oracle.ps1` is what produces the input, and
|
||||
> also the reason `spike-notes` §11 now records that the console binary will not run
|
||||
> unelevated on the dev box at all.
|
||||
>
|
||||
> ### 🔴 MEASURED 2026-08-27, AND THE RUNG IS NARROWED: NVIDIA DLSS-G IS NOT COVERED
|
||||
>
|
||||
> §S31 ran — three legs, three game launches, Cyberpunk 2077 at off / ×2 / ×4 — and landed
|
||||
> on row **P2**. `FrameType` is present and **every row of all three legs reads
|
||||
> `Application`**: 1,937 / 6,488 / 10,881 rows, no other value anywhere. Meanwhile the two
|
||||
> instruments agree on the present rate to within 0.3% per leg, so PresentMon SAW the
|
||||
> generated presents — at ×4 roughly three in four cannot be application frames — and
|
||||
> classified all of them as application frames.
|
||||
>
|
||||
> **So for NVIDIA frame generation this rung yields nothing on this driver and this
|
||||
> build.** It is not "unreliable" here; it produces an ABSENCE, and rung 0 turns an
|
||||
> absence into `N/A`. A consumer must never reach rung 4 from it.
|
||||
>
|
||||
> **What is still unmeasured is WHICH vendors it does cover.** Intel's own remains
|
||||
> untested despite being the obvious reading — the reading this document was already
|
||||
> corrected once for asserting — and AMD is untestable on the only machine that exists
|
||||
> (§R5/§R6). The honest scope of rung 2 today is therefore: **no vendor is known to
|
||||
> instrument this provider, and one is now known NOT to.**
|
||||
>
|
||||
> One sub-question does not change the above and is recorded in §S31 rather than here:
|
||||
> whether `--track_frame_type` was in effect at all. If the column ships by default the
|
||||
> legs measured a default rather than an absence. Either way the rung produced no
|
||||
> classification, which is why §S31 landed on P2 without waiting for it.
|
||||
|
||||
> **Rung 0, added 2026-08-06, and it has to come before rung 4 or rung 4 is a lie.** If
|
||||
> `FL_MEASURED_FG` is clear the answer is **`N/A`**, not `none`: no hook capable of answering
|
||||
|
||||
+1
-1
@@ -253,7 +253,7 @@ Findings written to `docs/spike-notes.md`. Nothing in P1 starts until the exit c
|
||||
> measurements that were supposed to feed it.
|
||||
5. **Vendor SDK reality check. ✅ DONE.** Measured across 34 distinct modules in 162 files from installed titles; `tools/vendor-exports.ps1` regenerates and `docs/vendor-exports.json` is committed. `17_HOOK_ENGINE` §Upscaling is corrected and its caveat discharged. **It needed no feature hooks** — it reads files with `dumpbin` — so the status header above was wrong to group it with 4/6/7. The finding: `sl.common.dll` exports the NGX parameter **accessors**, while the NGX core exports only the **factories**, so NGX-direct titles need a different hook class (`17_HOOK_ENGINE` §The NGX parameter surface splits into two hook classes).
|
||||
6. **RT detection.** Harness + a real DXR title: `DispatchRays` counting *and* `BuildRaytracingAccelerationStructure`; verify the AS-build path catches an inline-RayQuery title that dispatch counting misses. **✅ DONE 2026-08-20.** Both detours are on `ID3D12GraphicsCommandList4`, and the discriminating claim is **proved rather than asserted**: `ctest fl_guard`'s two harness arms differ by one recorded call, and `--hold-presenting-rayquery` reports `AsBuildObserved = 60` against `DispatchObserved = **0**` — a writer with only the dispatch hook sees nothing there and its silence is indistinguishable from a real negative (`spike-notes` §6). Four real titles agree with their own settings menus, `Yes` twice and `No` twice. **Carry the trap forward:** a command list's first `Reset()` swaps in a per-object vtable in which the driver has taken `DispatchRays` over, so the first version of this hook installed, published its family bit and never fired — `HANDOFF` §item 4.
|
||||
7. **Frame Generation ground truth.** ~~Compare rung 1 (FG feature evaluations per present) against Tier-2 ETW `FrameType` on a DLSS-G title.~~ **BLOCKED ON A PRODUCER, not on tooling, measured 2026-08-15:** `slEvaluateFeature(kFeatureDLSS_G)` is never called by Cyberpunk 2077 (0 across ~14,000 Streamline batches at four FG settings), so rung 1 emits nothing for ETW to be compared against. What DID track the setting is `presents / batch` — 1.000 / 2.000 / 4.000 against off / ×2 / ×4 — but a batch is not an application frame and that premise has no independent oracle. `docs/HANDOFF.md` item 3 carries the candidate producers and the decision nobody has taken; `spike-notes` §9 carries the numbers. ~~rung 2 (`GetFrameStatistics` present delta)~~ was removed as structurally impossible, not merely unreliable (`03_METRICS` §Frame Generation). Driver-level FG (AFMF) is undetectable at Tier 1 in v1; whether PresentMon 2.x `FrameType` sees it at Tier 2 is `20_OPEN_QUESTIONS` §M1 — and is untestable on this dev machine, which has no AMD GPU.
|
||||
7. **Frame Generation ground truth.** ~~Compare rung 1 (FG feature evaluations per present) against Tier-2 ETW `FrameType` on a DLSS-G title.~~ **BLOCKED ON A PRODUCER, not on tooling, measured 2026-08-15:** `slEvaluateFeature(kFeatureDLSS_G)` is never called by Cyberpunk 2077 (0 across ~14,000 Streamline batches at four FG settings), so rung 1 emits nothing for ETW to be compared against. What DID track the setting is `presents / batch` — 1.000 / 2.000 / 4.000 against off / ×2 / ×4 — but a batch is not an application frame and that premise has no independent oracle. `docs/HANDOFF.md` item 3 carries the candidate producers and the decision nobody has taken; `spike-notes` §9 carries the numbers. **AND THE ETW HALF IS NOW RETIRED, 2026-08-27.** The comparison this item asks for was attempted: three legs, three game launches, and PresentMon classified **every** frame `Application` at ×2 and ×4 while agreeing with our own hook on the present rate to within 0.3% — §S31 row **P2**. So rung 1 emits nothing and rung 2 cannot see NVIDIA frame generation either; this item has no instrument on either side, and `03_METRICS` rung 2 is narrowed accordingly. What that leaves is a hook-route decision with no oracle behind it (`HANDOFF` item 3). ~~rung 2 (`GetFrameStatistics` present delta)~~ was removed as structurally impossible, not merely unreliable (`03_METRICS` §Frame Generation). Driver-level FG (AFMF) is undetectable at Tier 1 in v1; whether PresentMon 2.x `FrameType` sees it at Tier 2 is `20_OPEN_QUESTIONS` §M1 — and is untestable on this dev machine, which has no AMD GPU.
|
||||
8. **Telemetry layering.** Fill the `18_GPU_VENDOR_APIS` capability matrix on real hardware:
|
||||
- L1 baseline (DXGI + PDH counters) working vendor-neutrally; decide whether the `D3DKMT` perf-data probe is stable enough on Win 10 **and** Win 11 to keep. **🅓 Deferred to Win 11 only, 2026-08-05** — one machine, and it is Win 11. Win 10 22H2 stays a supported floor and is explicitly unmeasured; the deferral holds only for as long as the probe stays non-load-bearing, which `18_GPU_VENDOR_APIS` §L1 now states as the condition rather than as advice.
|
||||
- L2: which fields LibreHardwareMonitor actually returns per vendor, and **whether GPU sensors work unelevated without PawnIO** — this decides whether the default unelevated Agent has temperatures.
|
||||
|
||||
@@ -90,7 +90,7 @@ or it becomes the next stale status claim this file exists to record.
|
||||
|
||||
| **S29** | 🔴 **open, new 2026-08-05 — five ✅ and two ◐, counted off the glyphs below on 2026-08-20** | (a) ◐ **CORRECTED**: the honesty assertion *is* in the merge gate, natively (`fl_guard`, 20.58 s on CI); only the **managed** drain is ungated, and §S19(b) is **not** a prerequisite of the feature hooks — the original claim was wrong and had been used to re-order the work. **Sharpened 2026-08-06:** fixing §S19(b) alone would still gate nothing, because `build.ps1`'s `-SkipIntegration` applies `--filter 'Category!=Integration'` and excludes the class *before* the guard is ever asked. Two independent mechanisms produce one absence, and `ci.yml` must drop the switch too; (b) ✅ `fl_vtable_indices` now pins the Overlay's indices through a shared header; (c) ✅ **closed 2026-08-06 by deleting `ShouldUnhookAsync`** — zero production callers, strictly weaker than `ScanOnceAsync`, and inverted in polarity; the gate's public instance surface is now pinned to `{StartAsync}`; (d) ◐ `vklayer-blastradius` case 3 is now an assertion, but the script still runs only by hand; (e) ✅ **closed 2026-08-06 host-side**, with a held process handle and a classifier that takes no elapsed-time parameter, so a frozen `writeIndex` can never end a session; (f) ✅ **closed 2026-08-20**: the normative contradiction between CLAUDE.md rule 7 and `03_METRICS` about inline RayQuery is settled the way the entry proposed — AS-build activity proves ray tracing is happening, *naming the technique* is what needs a DXIL scan — and rule 7 is amended in the PR that wrote the hooks. Its second half closes with it: `FL_MEASURED_RT` now has a producer (both command-list detours, installed off the game's own device), so all three conjuncts of the `No` branch are live and **`Yes` and `No` are both reachable for the first time**, proved by injection in both directions; (g) ✅ the present-only writer claimed `FL_MEASURED_OUTPUT_RES` unconditionally, including on records with no size |
|
||||
| **S30** | ✅ **closed 2026-08-15** | Answered on the title that raised it: with Ray Reconstruction on, Cyberpunk 2077 evaluates `kFeatureDLSS_RR` on every application frame and `kFeatureDLSS` **not once**, so RR was doing the upscaling and the decode had no arm for it. Evidence rather than inference: `renderW/H` are published only on a frame that drained an evaluation and read 1485×835 = 0.58 × 2560×1440, so the scaling-input tag arrives ON the RR evaluation. Two further defects fell out — the pre-committed decision table had two holes, and the fix contaminated the census that found it (it derived the id from the decoded byte) |
|
||||
| **S31** | ❓ **open, new 2026-08-20** | Is a drained Streamline batch an application frame? The PresentMon 2.x `FrameType` test, with its decision table pre-committed BEFORE the run and `tools/frametype-oracle.ps1` producing the input as two dimensionless ratios. **Two obstacles measured before it could be run**: the console binary will not start a trace session unelevated on the dev box (exit 6; the account is in neither Administrators nor Performance Log Users, and the running shared service does not help), and `--track_frame_type` is a **beta** option whose own help says it needs application or driver instrumentation of the Intel-PresentMon provider — so it may be as unavailable as `fl-baseline-probe` proved to be. Two of the table's six rows retire it outright. Blocks HANDOFF item 3's producer decision |
|
||||
| **S31** | ◐ **row P2, 2026-08-27 — the ORACLE is retired, the QUESTION is not** | Is a drained Streamline batch an application frame? The PresentMon 2.x `FrameType` test, with its decision table pre-committed BEFORE the run and `tools/frametype-oracle.ps1` producing the input as two dimensionless ratios. **Two obstacles measured before it could be run**: the console binary will not start a trace session unelevated on the dev box (exit 6; the account is in neither Administrators nor Performance Log Users, and the running shared service does not help), and `--track_frame_type` is a **beta** option whose own help says it needs application or driver instrumentation of the Intel-PresentMon provider — so it may be as unavailable as `fl-baseline-probe` proved to be. Two of the table's six rows retire it outright. Blocks HANDOFF item 3's producer decision |
|
||||
| **H11** | 🅓 **deferred 2026-08-20, rationale written** | XeFG and FSR3-FG identity. `libxess_fg.dll` and `ffx_fsr3_x64.dll` export nameable entry points, but the newer `amd_fidelityfx_framegeneration_dx12.dll` (3.1.5, three installed titles) exports only five generic `ffx*` names — identity lives in a struct field, and hand-declaring a vendor ABI from observation is the #71 defect class with another vendor's name on it. The licence checklist has not been run on either SDK either. Cost is coverage, not correctness: such a title reports `FL_FG_UNKNOWN`, never `NONE` |
|
||||
|
||||
~~**Six items are ❓ and one is 🚫.**~~ **Recounted 2026-08-05: TWELVE items block
|
||||
@@ -99,13 +99,13 @@ The criterion is *"resolved, **or** explicitly deferred with a written rationale
|
||||
so ⏳ (open, sequenced) and ◐ (partly closed) are just as open as ❓, and each needs
|
||||
either work or a rationale written down:
|
||||
|
||||
- ~~**six ❓**~~ ~~three ❓~~ **four ❓** — S6, S19(c), S20 feed half, ~~**S30**~~ **S31**. *(S19(a),
|
||||
- ~~**six ❓**~~ ~~three ❓~~ ~~four ❓~~ **three ❓** — S6, S19(c), S20 feed half. ~~**S30**~~ ~~**S31**~~ *(S19(a),
|
||||
S19(d) residual and S23-5 resolved 2026-08-05, each by a mechanism rather than by a
|
||||
correction. **S30 closed 2026-08-15 and S31 opened 2026-08-20**, and this bullet went on
|
||||
naming S30 until 2026-08-27 — so the total below stayed right by coincidence rather than
|
||||
by count.)*
|
||||
- **three ⏳** — S2 part three, S4 signing, S23-3
|
||||
- **two ◐** — S14, S23-6
|
||||
- ~~two ◐~~ **three ◐** — S14, S23-6, **S31** (moved out of ❓ on 2026-08-27: its test ran and landed on row P2, so the oracle is settled and the question is not)
|
||||
- **one 🔴** — S29, added by the audit that produced this recount; **five of its seven
|
||||
findings are now closed**, and the residue is (a) the ungated managed drain and (d) the
|
||||
hand-run blast-radius script. *(This bullet listed **three** residues beside a count of
|
||||
@@ -127,6 +127,12 @@ the recount becomes the next thing that needs recounting.
|
||||
> in a pass that touched no code — the §Traps entry *"a document can go stale by NOT being
|
||||
> touched"* wearing the ledger's own name.
|
||||
|
||||
> **STILL TEN on 2026-08-27, for the third consecutive movement.** S31 moved from ❓ to ◐,
|
||||
> so ❓ goes four to three and ◐ goes two to three. **The number has now stayed at ten across
|
||||
> three different changes**, and that is worth saying out loud precisely because a reader
|
||||
> checking only the total would conclude nothing had happened since 2026-08-15. Three things
|
||||
> have. The total is the least informative line in this section and it is the one people read.
|
||||
|
||||
> **A rising count here is the ledger working, not failing.** Every previous movement was downward and came from closing something. This one is upward because the first capture from a real game produced a wrong answer that no fixture had been able to produce — which is the whole reason exit criterion 1 asks for a real title rather than a harness.
|
||||
|
||||
> **Still nine on 2026-08-06.** Item 1 closed §S29(c) and §S29(e) and corrected
|
||||
@@ -791,7 +797,88 @@ a wrong answer into a confident wrong answer.
|
||||
|
||||
</details>
|
||||
|
||||
### S31 ❓ · Is a drained Streamline batch an application frame? — the PresentMon `FrameType` test, pre-committed
|
||||
### S31 ◐ · Is a drained Streamline batch an application frame? — the ORACLE is retired, the QUESTION is not
|
||||
|
||||
> ### 🔴 RUN 2026-08-27 — ROW **P2**. PresentMon is RETIRED as the application-frame oracle.
|
||||
>
|
||||
> Three legs, three game launches, Cyberpunk 2077, frame generation off / ×2 / ×4.
|
||||
> `spike-notes` §11 carries the numbers. What belongs here is the row and what it does
|
||||
> and does not settle.
|
||||
>
|
||||
> **`FrameType` is present and every single row of all three legs reads `Application`** —
|
||||
> 1,937 / 6,488 / 10,881 rows, zero rows of any other value, counted off the raw CSVs
|
||||
> rather than off the tool's summary.
|
||||
>
|
||||
> **And the title was demonstrably generating frames, which is the half P2 needs.** The
|
||||
> two instruments agree on the present rate to within 0.3% on every leg:
|
||||
>
|
||||
> | leg | our hook | PresentMon | our `presents/batch` |
|
||||
> |---|---|---|---|
|
||||
> | off | 43.16 /s | 43.04 /s | N/A — no batch drained |
|
||||
> | ×2 | 144.31 /s | 144.18 /s | **2.00** |
|
||||
> | ×4 | 241.84 /s | 241.80 /s | **3.99** |
|
||||
>
|
||||
> So **both instruments are counting the same present stream**, displayed rate tracks the
|
||||
> configured multiplier, and at ×4 roughly three presents in four cannot be application
|
||||
> frames — while PresentMon classifies **100%** of them `Application`. That is P2 word for
|
||||
> word: *"PresentMon sees the presents but classifies none as generated, while the title is
|
||||
> demonstrably generating them."*
|
||||
>
|
||||
> `tools/frametype-oracle.ps1` reported it as a **falsifier** on all three legs and never as
|
||||
> a ratio of 1.0, which is the one behaviour that entry was written to guarantee.
|
||||
>
|
||||
> #### What is NOT yet excluded, and the one command that excludes it
|
||||
>
|
||||
> **Whether `--track_frame_type` was in effect is unmeasured.** If `FrameType` appears in
|
||||
> `--v2_metrics` output *regardless* of the flag, these legs show PresentMon's default rather
|
||||
> than an absence of vendor events. The discriminator costs five seconds, no game and no
|
||||
> capture — run PresentMon **without** the flag and look at the header:
|
||||
>
|
||||
> ```
|
||||
> PresentMon-2.5.1-x64.exe --process_name explorer.exe --v2_metrics \
|
||||
> --output_file %TEMP%\pm-noflag.csv --timed 5 --terminate_after_timed --stop_existing_session
|
||||
> ```
|
||||
>
|
||||
> No `FrameType` column there ⇒ the column exists only because of the flag ⇒ the vendor
|
||||
> emitted nothing. Column present ⇒ these three legs measured a default and this entry needs
|
||||
> re-reading. **It still needs elevation, so it is the owner's** — attempted unelevated
|
||||
> 2026-08-27 and refused with the same exit 6 this item has been blocked on from the start.
|
||||
>
|
||||
> **THE ACTION DOES NOT DEPEND ON IT, WHICH IS WHY THIS ROW LANDS ANYWAY.** Both branches end
|
||||
> at the same place: either the vendor emits nothing, or the invocation produced no frame-type
|
||||
> data at all. In both, PresentMon did not answer the question **as invoked**, and an oracle
|
||||
> that does not answer is retired rather than re-run until it does — the discipline
|
||||
> `fl-baseline-probe` was retired under. What the discriminator changes is the *reason*, and
|
||||
> therefore whether a later attempt with a different invocation is worth anyone's time.
|
||||
>
|
||||
> #### What P2 retires, and what it leaves exactly where it was
|
||||
>
|
||||
> **RETIRED:** PresentMon 2.x `FrameType` as the application-frame oracle for NVIDIA frame
|
||||
> generation. `03_METRICS` §Frame Generation rung 2 is narrowed accordingly.
|
||||
>
|
||||
> **NOT ANSWERED:** *is a drained Streamline batch an application frame?* This entry's actual
|
||||
> question is untouched. `presents / batch` still reads 2.00 and 3.99 against a title's own
|
||||
> ×2 and ×4, still on the unverified premise that the work is recorded once per application
|
||||
> frame, and still unpublishable as `fg_factor` for that reason. **Four oracles have now
|
||||
> fallen** — `fl-baseline-probe` by its own falsifier, two readings of Steam's overlay (§S30),
|
||||
> and this one. `HANDOFF` item 3's producer decision goes back to the hook routes, which is
|
||||
> where P1/P2 said it would go.
|
||||
>
|
||||
> #### Two things the run produced that nobody was looking for
|
||||
>
|
||||
> **The `off` leg drained ZERO Streamline batches**, where `spike-notes` §8 recorded
|
||||
> `presents/batch = 1.000` at off. The difference is Ray Reconstruction: §S30 established that
|
||||
> RR is evaluated once per application frame, so with RR on there are batches even at off. RR
|
||||
> was not on here. **Consequence for anyone rebuilding this comparison: our side had only TWO
|
||||
> readable legs, not three** — enable RR if you need the off leg to carry a ratio.
|
||||
>
|
||||
> **The CSV's FIRST column is named `Application`** — it holds the process name — while the
|
||||
> value being counted, in column 9, is also the string `Application`. A parser resolving by
|
||||
> position, or grepping for the word, counts process names and reports a ratio. The
|
||||
> *"columns are resolved BY NAME from the header, never by position"* rule was written into
|
||||
> `tools/frametype-oracle.ps1` before it had ever seen a real CSV, against exactly the
|
||||
> collision that turned out to be there. Recorded because the rule looked like ordinary care
|
||||
> when it was written and was load-bearing.
|
||||
|
||||
**The question HANDOFF item 3 cannot close without.** `presents / batch` reads
|
||||
1.000 / 2.000 / 4.000 against Cyberpunk 2077's own off / ×2 / ×4, and a *batch* is
|
||||
@@ -3151,7 +3238,7 @@ instead, which needs no thread at all.
|
||||
| # | Question |
|
||||
|---|---|
|
||||
| M1 | Can PresentMon 2.x `FrameType` see **driver-level** frame generation (AMD AFMF)? `03_METRICS` now says v1 cannot detect it at Tier 1. If Tier 2 can, that is a genuine and surprising capability inversion worth surfacing in the UI |
|
||||
| M2 | ◐ **Partly measured 2026-08-20** (`spike-notes` §11). It exists — PresentMon **2.5.1**, 956,768 bytes, SHA256 `9BEC…A191` — and it carries **no VERSIONINFO at all**, so pinning means pinning the hash and the filename. It does **NOT** run unelevated on this machine: exit 6, *"requires either administrative privileges or to be run by a user in the Performance Log Users user group"*, and the account is neither. The 2.x column set is therefore still **unmeasured**, which is why `tools/frametype-oracle.ps1` resolves columns by name and refuses loudly. `15_ROADMAP` parks the Service + API2 in v2, so there is no planned fallback if the console is gone |
|
||||
| M2 | ◐ **Partly measured 2026-08-20** (`spike-notes` §11). It exists — PresentMon **2.5.1**, 956,768 bytes, SHA256 `9BEC…A191` — and it carries **no VERSIONINFO at all**, so pinning means pinning the hash and the filename. It does **NOT** run unelevated on this machine: exit 6, *"requires either administrative privileges or to be run by a user in the Performance Log Users user group"*, and the account is neither. The 2.x column set is therefore still **unmeasured**, which is why `tools/frametype-oracle.ps1` resolves columns by name and refuses loudly. **✅ The column set is MEASURED 2026-08-27: 24 columns, listed in `spike-notes` §11, and the parser has now eaten three real CSVs.** It resolved `FrameType` correctly against a first column literally named `Application`, a collision that would have made a position-based or grep-based parser count process names. What stays open in this row is the elevation-free route, not the column set. `15_ROADMAP` parks the Service + API2 in v2, so there is no planned fallback if the console is gone |
|
||||
| M5 | **Do LHM GPU sensors work unelevated, without PawnIO?** This decides whether the default unelevated Agent has temperatures at all, and therefore how ADR-9 reads to users |
|
||||
| M6 | ◐ **Half of it is answered, 2026-08-20, and it is the half nobody expected to fall.** *"or the PresentMon Service"* — **no.** `PresentMonSharedService` is installed and `Running` as LocalSystem on this machine, and the console binary still fails with access denied, because the console starts its **own** trace session rather than talking to the service. So the service being present is not a route to elevation-free Tier 2 for the console. The *"add me to Performance Log Users"* half stays open: adding an account to a local group is a system settings change needing admin, and it is the owner's rather than a session's. Would change the README, Disclaimer and EULA wording back |
|
||||
| M7 | `18_GPU_VENDOR_APIS` §Runtime policy says telemetry is never read from the game process, but `17_HOOK_ENGINE` reads per-process VRAM and Reflex latency there. Reconcile the wording — the rule means "no vendor SDK polling loops in the game", not "no measurement in the game" |
|
||||
|
||||
+26
-5
@@ -437,11 +437,32 @@ arithmetic, the refusals and the fixtures are in and gated. What is missing is n
|
||||
> said three oracles had fallen; a file whose whole subject is staleness contradicted itself
|
||||
> inside one entry, which is recorded rather than quietly repaired.
|
||||
>
|
||||
> **And the owner's list for the §S31 run, in the order it has to happen:** elevate (or join
|
||||
> `Performance Log Users`) → capture Cyberpunk 2077 at off / ×2 / ×4 with
|
||||
> `--v2_metrics --track_frame_type` → run `tools/frametype-oracle.ps1 -Csv … -Report …` →
|
||||
> land on a §S31 row. **Three settings, not two:** §S30 already burned a draft on the ×2 leg,
|
||||
> which cannot discriminate because "displayed ÷ a fixed 2" predicts the same number there.
|
||||
> ~~**And the owner's list for the §S31 run**~~ — **RUN 2026-08-27. `tools/s31-capture.ps1`
|
||||
> drove it, three legs, three game launches, and it landed on row **P2**: `FrameType` present
|
||||
> and **every row of all three legs `Application`**, while the two instruments agreed on the
|
||||
> present rate to within 0.3% — so PresentMon saw the generated presents and classified none
|
||||
> of them as generated.
|
||||
>
|
||||
> **PRESENTMON IS RETIRED as the application-frame oracle for NVIDIA frame generation.** That
|
||||
> is the fourth oracle to fall, after `fl-baseline-probe` and two readings of Steam's overlay.
|
||||
>
|
||||
> **AND THIS ITEM IS BACK WHERE P1/P2 SAID IT WOULD GO: the hook routes.** What was retired is
|
||||
> the *instrument*, not the question. "Is a drained Streamline batch an application frame?" is
|
||||
> exactly as open as it was, `presents / batch` still reads 2.00 and 3.99 against a title's own
|
||||
> ×2 and ×4 on the same unverified premise, and it is still not publishable as `fg_factor`.
|
||||
>
|
||||
> **So the decision this entry has been waiting on is now unblocked and unmade.** The candidate
|
||||
> routes above are unchanged and none is costed: the interposer's swapchain proxy;
|
||||
> `slGetFeatureFunction` + `slDLSSGGetState`; vendoring `sl_dlss_g.h` with its consumer;
|
||||
> NGX-direct `nvngx_dlssg`; or shipping `presents/batch` with the premise stated. **The
|
||||
> measure-before-hooking rule has been discharged — the measurement was taken and it did not
|
||||
> choose a route for us.** Whoever picks this up is choosing without an oracle, and should say
|
||||
> so in the PR rather than implying one.
|
||||
>
|
||||
> Numbers in `spike-notes` §11. Two loose ends recorded there and not here: our own `off` leg
|
||||
> drained ZERO batches (Ray Reconstruction was off, so there is nothing to divide), and whether
|
||||
> `--track_frame_type` was in effect at all is unmeasured — it changes the reason, not the
|
||||
> action.
|
||||
|
||||
|
||||
> **§H5 case 3 is MEASURED as of 2026-08-15, and the answer is half of what the entry below
|
||||
|
||||
@@ -1671,6 +1671,82 @@ into a real title and the process survives it (§7).
|
||||
anything it does not recognise, and its decision table is the only thing standing
|
||||
behind it. Said plainly rather than left for the next reader to assume.
|
||||
|
||||
### 🔴 THE RUN HAPPENED — 2026-08-27, three legs, three game launches, and row **P2**
|
||||
|
||||
Owner-run from an elevated shell via `tools/s31-capture.ps1`. Cyberpunk 2077, frame
|
||||
generation off / ×2 / ×4 set in the title's own menu, 40 s per leg, one game launch
|
||||
each. Analysis is in `20_OPEN_QUESTIONS` §S31; what belongs here is what was measured.
|
||||
|
||||
**The 2.x column set, which §M2 left unmeasured: 24 columns.**
|
||||
|
||||
```
|
||||
Application, ProcessID, SwapChainAddress, PresentRuntime, SyncInterval, PresentFlags,
|
||||
AllowsTearing, PresentMode, FrameType, CPUStartTime, FrameTime, CPUBusy, CPUWait,
|
||||
GPULatency, GPUTime, GPUBusy, GPUWait, DisplayLatency, DisplayedTime, AnimationError,
|
||||
AnimationTime, MsFlipDelay, AllInputToPhotonLatency, ClickToPhotonLatency
|
||||
```
|
||||
|
||||
**`FrameType` is present, and every row of all three legs reads `Application`.**
|
||||
Counted off the raw CSVs rather than off the tool's own summary:
|
||||
|
||||
| leg | rows | distinct `FrameType` values |
|
||||
|---|---|---|
|
||||
| off | 1,937 | `Application` × 1,937 |
|
||||
| ×2 | 6,488 | `Application` × 6,488 |
|
||||
| ×4 | 10,881 | `Application` × 10,881 |
|
||||
|
||||
**And the two instruments agree on the present rate to within 0.3%**, which is what
|
||||
makes the finding a measurement rather than a shrug — PresentMon SAW the generated
|
||||
presents:
|
||||
|
||||
| leg | our hook | PresentMon | our `presents/batch` | our `evaluations` |
|
||||
|---|---|---|---|---|
|
||||
| off | 1,676 presents / 38.82 s = **43.16 /s** | 1,937 / 45 s = **43.04 /s** | N/A — no batch drained | 0 |
|
||||
| ×2 | 5,746 / 38.64 s = **144.31 /s** | 6,488 / 45 s = **144.18 /s** | **2.00** | 0 |
|
||||
| ×4 | 9,652 / 38.77 s = **241.84 /s** | 10,881 / 45 s = **241.80 /s** | **3.99** | 0 |
|
||||
|
||||
Every leg: `status=Ready`, `faults=0`, `layoutVersion=3`, `attach=Ok`, one stream, one
|
||||
segment, **0 gaps and 0 dropped**, `unidentified=0`.
|
||||
|
||||
So displayed rate tracks the configured multiplier, both instruments count the same
|
||||
present stream, and at ×4 roughly three presents in four cannot be application frames
|
||||
— while PresentMon classifies **100%** of them `Application`. That is §S31 row **P2**,
|
||||
and `tools/frametype-oracle.ps1` reported it as a **falsifier** on all three legs
|
||||
rather than as a ratio of 1.0, which is the one behaviour it was written to guarantee.
|
||||
|
||||
**Unmeasured, and it decides the REASON rather than the action:** whether
|
||||
`--track_frame_type` was in effect at all. If `FrameType` ships in `--v2_metrics`
|
||||
output regardless of the flag, these legs recorded a default. The discriminator is one
|
||||
five-second run **without** the flag, checking only whether the header carries the
|
||||
column — no game, no capture, no launch spent. It still needs elevation; attempted
|
||||
unelevated 2026-08-27 and refused with the same exit 6 this section opens with.
|
||||
|
||||
> **Both branches end at the same action**, which is why §S31 landed anyway: either the
|
||||
> vendor emitted nothing, or the invocation produced no frame-type data at all. In
|
||||
> neither case did PresentMon answer the question **as invoked**.
|
||||
|
||||
**Two findings nobody was looking for.**
|
||||
|
||||
- **The `off` leg drained ZERO Streamline batches**, against `presents/batch = 1.000`
|
||||
at off in §8. The difference is Ray Reconstruction: §S30 established RR is evaluated
|
||||
once per application frame, so with RR on there are batches even with FG off. RR was
|
||||
not on here. **Our side therefore had two readable legs, not three** — turn RR on if
|
||||
the off leg needs to carry a ratio.
|
||||
- **The CSV's first column is named `Application`** and holds the process name, while
|
||||
the value being counted, in column 9, is also the string `Application`. A parser
|
||||
resolving by position — or grepping for the word — counts process names and reports a
|
||||
ratio from them. `frametype-oracle.ps1`'s *"resolved BY NAME from the header, never by
|
||||
position"* was written before it had ever seen a real CSV, against exactly the
|
||||
collision that was waiting in it.
|
||||
|
||||
**`kFeatureDLSS_G` is zero on all three legs**, which takes the count to five capture
|
||||
sets across five titles. `HANDOFF` item 3's premise has now failed everywhere it has
|
||||
been looked for.
|
||||
|
||||
**Artifacts:** `%LOCALAPPDATA%\FrameLedger\s31\{off,x2,x4}\` — `presentmon.csv` and
|
||||
`capturehost-report.txt` per leg. Not committed; they are measurement output and the
|
||||
numbers above are the record.
|
||||
|
||||
### 🔴 `--track_frame_type` is a BETA option that needs the VENDOR to cooperate
|
||||
|
||||
This is the finding, and it lands on a pre-committed oracle. `PresentMon 2.5.1
|
||||
|
||||
Reference in new issue
Block a user