Skip to content

armgit / Piik

Public Imported from GitHub
0

Free, open-source screen sharing for private live streams with friends. Watch together in a browser or self-host Piik. 免费开源的私密屏幕共享,支持游戏直播、一起看电影与自部署。

docs(media): record codec hardware decision (#69)

AAA555b728b3203
diff --git a/docs/README.md b/docs/README.md
index fc518b41..b4beaa5b 100644
--- a/docs/README.md
+++ b/docs/README.md
@@ -12,6 +12,7 @@
- [Built-in peer ICE TURN candidate](./research/built-in-peer-ice-turn.md): implemented default-off direct-first TURN fallback, short-lived participant-session bearer, refresh ownership, idle-allocation cost, and deployment gates.
- [Advanced peer distribution](./research/advanced-peer-distribution.md): multi-tree/SVC, encoded-object relay, native RTP forwarding, FEC/network-coding, MoQ, and measurable go/no-go gates.
- [Native shared-encode sender](./research/native-shared-encode-sender.md): consolidated Draft #16/#18/#22/#23/#25/#28 ladder, unclassified product-gate failure, staged revalidation, and stop line.
+- [Native H.264 hardware decision](./research/native-h264-hardware-decision.md): bounded WebCodecs H.264 hardware no-go, constraint/fmtp compatibility boundary, Pion Annex-B assessment, and the hardware-only Media Foundation next slice.
- [Realtime quality adaptation](./research/realtime-quality-adaptation.md): correlated capture/send/receive diagnosis, production generation-rebuild gate, path-isolated dual representations, and bounded simulcast/Dynacast/SVC checks.
- [Browser screen-audio quality](./research/browser-screen-audio-quality.md): capture/source compatibility, standard readback limits, Opus negotiation boundaries, and the retained no-control decision.
- [Display name and presence boundaries](./research/display-name-presence.md): browser-local storage failure behavior, Unicode normalization, control-character policy, and the Web-only presence capability boundary.
diff --git a/docs/research/native-h264-hardware-decision.md b/docs/research/native-h264-hardware-decision.md
new file mode 100644
index 00000000..b7e96e83
--- /dev/null
+++ b/docs/research/native-h264-hardware-decision.md
@@ -0,0 +1,334 @@
+# Native H.264 Hardware Decision Spike
+
+Date: 2026-08-20
+
+Status: `no-go-webcodecs-h264-hardware-unclassified`
+
+This spike does not amend ADR-0006 or authorize another product, room, or
+production run. It answers one narrow question: can the existing local-browser
+Native candidate switch from VP8 to WebCodecs H.264 and satisfy the project's
+physical hardware-encode contract with a small codec change?
+
+The answer is no. The H.264 bitstream and Pion packetization path are
+mechanically plausible, but the one bounded run did not observe a hardware
+encoder and emitted SPS constraint bytes that differ from the requested
+WebCodecs codec string. Their negotiated-fmtp compatibility remains
+unclassified. A WebCodecs H.264 product switch is frozen. The next candidate is
+a separate Windows hardware-only Media Foundation fixture, not a codec ladder
+or fallback framework.
+
+## Decision Summary
+
+| Question | Result |
+| --- | --- |
+| Does Chrome 151 accept H.264 `avc1.42e01f`, Annex B, 720p30, realtime, 3 Mbps, `prefer-hardware`? | Yes. |
+| Did the run prove an NVIDIA, AMD, or other hardware encoder? | No. Chrome-process Windows `VideoEncode` samples were all zero and no NVIDIA encode session was added. The sampling boundary cannot prove software because a later-created or non-child hardware process was not retained. |
+| Did the requested codec string describe the emitted SPS exactly? | No. The requested `42e01f` produced SPS bytes `42 04 1f`. |
+| Did forced key frames carry in-band recovery data? | Yes in this fixture: all six key chunks contained AUD, SPS, PPS, and IDR NAL units. |
+| Can Pion v1.10.5 packetize the resulting Annex-B access units? | Yes statically: its `H264Payloader` splits 3/4-byte start codes, suppresses AUD/filler, emits SPS/PPS as STAP-A, and fragments slices as FU-A. |
+| Is unmodified browser Viewer decode proven? | No. No PeerConnection, room, Viewer, or production service participated. |
+| Should the small WebCodecs/Pion codec diff be implemented? | No. It would preserve the two unresolved product gates: physical encoder identity and exact negotiated H.264 fmtp/constraint compatibility. |
+
+## Standards And Platform Constraints
+
+- WebCodecs defines `hardwareAcceleration: "prefer-hardware"` as a preference.
+ The user agent may ignore it, and `isConfigSupported()` is only a best-effort
+ capability answer. The answer can also change with resource availability.
+- The AVC WebCodecs registration defines `avc.format: "annexb"`. Annex-B output
+ carries SPS/PPS in the bitstream instead of an AVC decoder configuration
+ description, which is the appropriate shape for live RTP packetization.
+- WebRTC browsers must implement VP8 and H.264 Constrained Baseline. H.264
+ endpoints must support RFC 6184 and `packetization-mode=1`, include
+ `profile-level-id`, and send parameter sets in-band.
+- NVIDIA lists the local GeForce RTX 4070 SUPER as an Ada, eighth-generation
+ NVENC device with H.264 4:2:0 support. That proves the machine has suitable
+ hardware; it does not prove Chrome selected it.
+- Media Foundation can enumerate hardware MFTs explicitly with
+ `MFT_ENUM_FLAG_HARDWARE`. A D3D11-aware transform can receive an
+ `IMFDXGIDeviceManager`, allowing capture and encode surfaces to remain on the
+ selected device. Hardware enumeration and D3D awareness are read-back gates,
+ not hints.
+
+## Bounded Fixture
+
+### Isolation
+
+- Browser: Google Chrome `151.0.7922.138`.
+- GPU cohort: NVIDIA GeForce RTX 4070 SUPER, NVIDIA driver `610.88`; an AMD
+ Radeon integrated adapter and virtual display adapters were also present.
+- Source: local `127.0.0.1` page with a synthetic 1280x720 Canvas. No Screener
+ server, signaling endpoint, room, account, Viewer, Oopz process, or production
+ endpoint was started or contacted.
+- Chrome used a unique temporary profile, D3D11 ANGLE, background throttling
+ disabled, and an off-screen window. The two temporary profiles contained no
+ reparse points. After the run, both exact temporary trees were audited again
+ without following links, deleted, and confirmed absent.
+- Each codec was attempted at most once. H.264 performed one encode run. VP8
+ failed the exact `prefer-hardware` capability check and was not silently
+ retried with `no-preference`.
+
+### Encoder Fixture
+
+```text
+width/height: 1280 x 720
+framerate: 30
+duration: 360 inputs / about 12 seconds
+bitrate: 3,000,000 bits/s
+latencyMode: realtime
+bitrateMode: variable
+hardwareAcceleration: prefer-hardware
+H.264 codec/format: avc1.42e01f / annexb
+key-frame request: frames 0, 60, 120, 180, 240, 300
+source: moving synthetic Canvas, one VideoFrame per 33.333 ms
+```
+
+The output callback recorded each chunk's timestamp, size, callback latency,
+and Annex-B NAL types. It parsed the SPS profile, constraint, and level bytes.
+It did not save encoded frames.
+
+The temporary fixture and sampling scripts were not retained. This is a single
+bounded observation, not a repository-reproducible benchmark. The 15.984-second
+CPU sampling window also extends beyond the 11.978-second encode loop, so its
+CPU value describes the whole sampled Chrome process tree rather than active
+encode cost.
+
+### Physical Evidence Method
+
+The hardware decision did not use `isConfigSupported()` as proof.
+
+1. The launched Chrome process tree was resolved after the target attached and
+ before encoder configuration. It included the browser, GPU, utility, and
+ renderer processes.
+2. Windows sampled
+ `\\GPU Engine(*)\\Utilization Percentage` once per second for 15 samples and
+ retained only `engtype_videoencode` instances owned by that process tree.
+3. The same samples retained cumulative CPU seconds for the process tree.
+4. `nvidia-smi` sampled global encoder session count, FPS, latency, and encoder
+ utilization. An unrelated `nvcontainer.exe` session already existed, so
+ global utilization was discarded as attribution evidence; only a new
+ session could have corroborated Chrome NVENC use.
+
+This method has one explicit limitation: the process tree was not re-enumerated
+after encoding began. A hardware service created later or outside Chrome's
+parent/child tree would not have been attributed. The result is therefore
+"hardware not proven", not a claim that a particular software encoder was
+positively identified.
+
+## Raw Bounded Evidence
+
+### H.264
+
+| Measure | Retained value |
+| --- | --- |
+| `isConfigSupported` | `true`; returned config retained `prefer-hardware`, `realtime`, `variable`, and `annexb` |
+| Inputs / outputs | `360 / 360` |
+| Encode-loop wall time | `11,977.6 ms` |
+| Output bytes / observed VBR rate | `2,275,233 / 1,521,047 bit/s` |
+| Encoder queue peak | `8` |
+| Output callback latency | p50 `10.0 ms`; p95 `11.3 ms`; max `241.8 ms` |
+| Scheduling delay | p50 `1.07 ms`; p95 `14.17 ms`; max `15.8 ms` |
+| Pending outputs / fatal error | `0 / none` |
+| Decoder config | codec `avc1.42e01f`, 1280x720, description `0` bytes |
+| Chrome-tree `VideoEncode` | 15 samples: `[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]` |
+| NVIDIA global sessions | 15 samples: `[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]` |
+| Chrome-tree CPU | `3.21875 CPU-s / 15.984 wall-s`; `1.259%` of 16 logical processors, or about `20.1%` of one logical processor |
+
+CPU includes Chrome, Canvas drawing, frame construction, and encode work; it is
+not an isolated encoder CPU measurement.
+
+All six requested key chunks had NAL types `[9, 7, 8, 5]`, meaning AUD, SPS,
+PPS, and IDR. Their byte sizes were
+`[6074, 8675, 6672, 9567, 8483, 10011]`. Each began with the same retained
+prefix:
+
+```text
+000000010910000000016742041fe8a02802ddff
+```
+
+The parsed SPS bytes were `profile_idc=0x42`, constraints `0x04`, and
+`level_idc=0x1f` (`42041f`). That is not the requested codec string's
+`42e01f`. WebCodecs requires an encoder to produce a stream at least as
+constrained as the requested string, but this spike does not attempt to prove
+the H.264 semantic relationship between those two constraint bytes. The
+product gate is simpler: do not advertise or request an fmtp value until the
+emitted stream's constraint bytes are proven compatible, then prove an
+unmodified Viewer negotiation and decode.
+
+### VP8
+
+The exact same 720p30 configuration with codec `vp8` and
+`hardwareAcceleration: "prefer-hardware"` returned
+`isConfigSupported().supported=false`. No VP8 frames were encoded and no
+latency or CPU comparison is claimed. The current ADR-0006 candidate deletes
+`hardwareAcceleration` and continues when that check fails; the existing
+shared-encode research already classifies that behavior as an implementation
+preference, not physical hardware evidence.
+
+## Static Product-Path Evaluation
+
+The small bitstream path is plausible:
+
+```text
+one VideoEncoder
+ -> existing 17-byte Frame envelope
+ -> one Go Fanout queue and RTP packetizer
+ -> one TrackLocalStaticRTP
+ -> at most two PeerConnections / transport-local bindings
+```
+
+- `native/sender/internal/app/ui/app.js` owns one `VideoEncoder` object and
+ marks the next submitted frame as a key frame when any downstream request is
+ pending.
+- `native/sender/internal/media/fanout.go` owns one queue, one timeline, and one
+ packetizer. Switching `VP8Payloader` to `H264Payloader` would not create a
+ per-Viewer encoder.
+- `native/sender/internal/remote/peer.go` adds the same track to every bounded
+ PeerConnection. PLI or FIR on any leg emits the existing shared key-frame
+ request. A connected late Viewer also requests a key frame.
+- Pion RTP v1.10.5 accepts Annex-B access units. Its H.264 payloader stores SPS
+ and PPS, emits them as STAP-A before the following NALU, fragments large NALUs
+ as FU-A, and discards AUD/filler. This requires negotiated
+ `packetization-mode=1`.
+
+The fixture's repeated `[SPS, PPS, IDR]` key chunks make the shared feedback
+shape credible: one merged PLI/FIR can generate one access unit for every
+bound Viewer. This is static evidence plus an encoder-only fixture, not a
+Viewer proof.
+
+The unresolved constraint-byte/fmtp compatibility boundary blocks the
+mechanical switch. Pion v4.2.18's
+default H.264 mode-1 capabilities include `42001f`, `42e01f`, `4d001f`, and
+`64001f`; they do not include the observed `42041f`. Binding a track by MIME
+type alone can fall back to a partial codec match and advertise a different
+format. The gate must instead:
+
+1. parse emitted SPS before publishing;
+2. register and offer only an exactly compatible mode-1 capability;
+3. retain the selected answer `profile-level-id` without raw SDP;
+4. prove Chrome, Edge, Firefox, and Safari Viewer decode/render;
+5. prove every join/PLI/FIR recovery unit carries usable in-band SPS/PPS and
+ IDR; and
+6. retain one encoder invocation while two independent RTP bindings advance.
+
+The obvious WebCodecs/Pion edit is roughly 40-80 core lines plus focused tests,
+but it cannot supply either a portable hardware identity or this interop proof.
+It therefore does not satisfy this spike's 200-300-core-line implementation
+exception despite being textually small.
+
+## WebCodecs Versus Native Hardware APIs
+
+| Route | Advantage | Blocking cost / risk | Decision |
+| --- | --- | --- | --- |
+| WebCodecs H.264 Annex B + Pion | Smallest diff; observed SPS/PPS/IDR cadence; existing shared fanout remains intact | Hardware request is a hint; no encoder identity/readback; emitted constraint bytes differ and fmtp compatibility is unclassified; no Viewer proof | Freeze |
+| Media Foundation hardware H.264 MFT | Windows hardware-only enumeration, exact transform activation, D3D11 device manager, ICodecAPI support/readback, vendor-neutral across installed hardware MFTs | COM/asynchronous MFT lifecycle and GPU-surface ownership are materially larger than 300 core lines | Next bounded fixture |
+| Direct NVENC | Explicit NVIDIA encoder session and detailed low-latency controls | NVIDIA-only, SDK/API lifecycle and redistribution review, separate AMD/Intel future decisions | Hold unless the Media Foundation hardware contract fails |
+
+Media Foundation is the Occam candidate because it can fail closed on an
+official hardware category while avoiding a vendor dispatch layer. It is one
+Windows H.264 path, not an abstraction for VP8/VP9/AV1/HEVC and not a software
+fallback.
+
+## Next Native Slice
+
+Build an offline Windows executable fixture before touching capture, rooms, or
+the existing sender:
+
+1. Create one D3D11 device on the explicitly selected adapter with
+ `D3D11_CREATE_DEVICE_VIDEO_SUPPORT`; create and reset one DXGI device
+ manager.
+2. Enumerate only `MFT_CATEGORY_VIDEO_ENCODER` transforms matching H.264 output
+ with `MFT_ENUM_FLAG_HARDWARE | MFT_ENUM_FLAG_SORTANDFILTER`. Record bounded
+ activation identity and fail when the count is zero.
+3. Activate exactly one transform. Require `MF_SA_D3D11_AWARE=true`, send
+ `MFT_MESSAGE_SET_D3D_MANAGER`, and never retry a synchronous/software MFT.
+4. Use fixed NV12 1280x720@30 input and H.264 3 Mbps output. Require and read
+ back Constrained Baseline or another profile that exactly matches a Pion
+ mode-1 capability, level 3.1 or lower, realtime/low-latency mode, CBR mean
+ bitrate, bounded VBV, no frame reordering, and a 60-frame maximum GOP.
+ Unsupported, read-only, or weakened properties fail the fixture.
+5. Feed 360 synthetic GPU-resident frames. Retain input-to-output latency,
+ output cadence, queue depth, CPU, and adapter-specific Windows
+ `VideoEncode`. Add a vendor session counter only where the selected adapter
+ exposes one. Capability enumeration alone does not pass.
+6. Force an IDR at frames 0/60/120/180/240/300 and through the same
+ key-frame-request entry point. Parse every recovery access unit for Annex-B
+ SPS/PPS/IDR and exact profile/level bytes.
+7. Pass the same encoded access unit through Pion `H264Payloader` to two local
+ transport bindings, then prove an unmodified browser Viewer answer and
+ decode/render. Do not enter a Screener room.
+
+NVIDIA's low-latency guidance supports CBR, a very small (approximately one
+frame) VBV, low/ultra-low-latency tuning, and explicit IDR recovery. The first
+Media Foundation fixture should use only controls that its selected hardware
+MFT reports as supported and modifiable and then returns unchanged. Do not add
+an NVENC fallback when a Media Foundation control is absent.
+
+### Rough Size
+
+| Responsibility | Estimated core LOC |
+| --- | ---: |
+| D3D11 adapter/device, DXGI manager, synthetic NV12 surfaces | 180-260 |
+| Hardware-only MFT enumeration, activation, identity/readback | 100-160 |
+| Media types, ICodecAPI configuration, asynchronous input/output pump | 260-380 |
+| Annex-B/profile/key-frame validation and bounded measurements | 120-180 |
+| **Offline fixture total** | **660-980** |
+
+Focused tests and harness glue are likely another 200-350 lines. Windows
+Graphics Capture, BGRA-to-NV12 conversion, audio, packaging, and product bridge
+integration are explicitly outside that estimate. This is not a 200-300-line
+product change, so no implementation was started in this branch.
+
+## Go / No-Go Gates
+
+The Media Foundation slice is `go-native-h264-hardware-fixture` only when all
+of these are retained in one run:
+
+- exact hardware MFT identity and D3D11-aware readback;
+- adapter-attributed `VideoEncode` activity and no software fallback;
+- 360/360 bounded outputs with p95 input-to-output latency recorded;
+- exact SPS profile/constraint/level matching the offered mode-1 fmtp;
+- SPS/PPS/IDR after startup, late-join request, PLI, and FIR;
+- the same encoded access unit packetized once and delivered over two transport
+ bindings; and
+- unmodified Viewer decode/render in the target browser matrix.
+
+Any missing physical, profile, feedback, or Viewer evidence remains no-go. Do
+not infer success from codec support, adapter model, transform enumeration,
+low CPU, or a single browser decode.
+
+## License Boundary
+
+- Pion RTP/WebRTC is MIT and remains the packetization/transport candidate; no
+ Pion source was copied into this repository.
+- Media Foundation and D3D11 are Windows platform APIs. The fixture should use
+ installed hardware MFTs and avoid redistributing vendor codec binaries.
+- Direct NVENC would require a separate NVIDIA Video Codec SDK license and
+ redistribution review before product work.
+- H.264 patent/licensing obligations remain unresolved regardless of whether
+ encoding is reached through WebCodecs, Media Foundation, or NVENC. The
+ project license and distribution model must be decided before distributing a
+ sender executable.
+
+## Primary Sources
+
+Accessed 2026-08-20:
+
+- [W3C WebCodecs](https://www.w3.org/TR/webcodecs/)
+- [W3C AVC WebCodecs registration](https://www.w3.org/TR/webcodecs-avc-codec-registration/)
+- [RFC 7742: WebRTC video codec requirements](https://www.rfc-editor.org/rfc/rfc7742.html)
+- [RFC 6184: H.264 RTP payload format](https://www.rfc-editor.org/rfc/rfc6184.html)
+- [NVIDIA video encode/decode support matrix](https://developer.nvidia.com/video-encode-decode-support-matrix)
+- [NVIDIA NVENC application note](https://docs.nvidia.com/video-technologies/video-codec-sdk/13.0/nvenc-application-note/index.html)
+- [NVIDIA NVENC programming guide](https://docs.nvidia.com/video-technologies/video-codec-sdk/13.0/nvenc-video-encoder-api-prog-guide/index.html)
+- [NVIDIA System Management Interface](https://docs.nvidia.com/deploy/nvidia-smi/index.html)
+- [Microsoft `typeperf`](https://learn.microsoft.com/en-us/windows-server/administration/windows-commands/typeperf)
+- [Microsoft `MFTEnumEx`](https://learn.microsoft.com/en-us/windows/win32/api/mfapi/nf-mfapi-mftenumex)
+- [Microsoft hardware MFTs](https://learn.microsoft.com/en-us/windows/win32/medfound/hardware-mfts)
+- [Microsoft `MF_SA_D3D11_AWARE`](https://learn.microsoft.com/en-us/windows/win32/medfound/mf-sa-d3d11-aware)
+- [Microsoft `MFCreateDXGIDeviceManager`](https://learn.microsoft.com/en-us/windows/win32/api/mfapi/nf-mfapi-mfcreatedxgidevicemanager)
+- [Microsoft H.264 video encoder](https://learn.microsoft.com/en-us/windows/win32/medfound/h-264-video-encoder)
+- [Microsoft low-latency codec property](https://learn.microsoft.com/en-us/windows/win32/medfound/codecapi-avlowlatencymode)
+- [Microsoft force-key-frame property](https://learn.microsoft.com/en-us/windows/win32/medfound/codecapi-avencvideoforcekeyframe)
+- [Microsoft H.264 profile enumeration](https://learn.microsoft.com/en-us/windows/win32/api/codecapi/ne-codecapi-eavench264vprofile)
+- [Pion RTP v1.10.5 H.264 payloader](https://github.com/pion/rtp/blob/v1.10.5/codecs/h264_packet.go)
+- [Pion WebRTC v4.2.18 media engine](https://github.com/pion/webrtc/blob/v4.2.18/mediaengine.go)
diff --git a/docs/research/native-shared-encode-sender.md b/docs/research/native-shared-encode-sender.md
index 1d17e6c7..f3afef1a 100644
--- a/docs/research/native-shared-encode-sender.md
+++ b/docs/research/native-shared-encode-sender.md
@@ -93,6 +93,45 @@ Audio, per-process WASAPI capture, WGC/DDA fallback, source selection, Electron,
installers, updating, and cross-platform abstractions stay outside the first
spike.
+### 2026-08-20 Windows codec comparison
+
+The target workstation exposes an NVIDIA GeForce RTX 4070 SUPER. NVIDIA's
+published NVENC contract encodes H.264, HEVC, and AV1, but not VP8. The current
+Native canary instead fixes VP8 and asks WebCodecs for `prefer-hardware`; its
+code removes that preference when the capability check rejects it. The prior
+bounded run did reject the preferred configuration and continued with
+`no-preference`. One `VideoEncoder` and its first output callback therefore do
+not establish hardware encoding, and VP8 cannot use this GPU's NVENC engine.
+
+Discord and the read-only Oopz 0.87.425 package inspection both make H.264 the
+smallest Windows hardware candidate. Discord publishes native OS/driver capture
+and encoding with hardware preferred and WebRTC transport. Oopz's package
+contains Agora screen-share integration, WGC/DXGI/D3D11 capture components, an
+H.264 Web viewer configuration, NVENC/QSV/AMF plus software codec paths, and an
+observable hardware-acceleration field. The inspection found one channel screen
+publication and no evidence of an application-owned per-viewer PeerConnection
+loop. That is consistent with a common publication, but static packaging cannot
+prove one physical encoder, the codec/profile selected for a real host, or
+actual GPU use.
+
+The bounded WebCodecs H.264 run, its no-go result, and the hardware-only Media
+Foundation successor are owned by
+[Native H.264 hardware decision spike](./native-h264-hardware-decision.md).
+Do not recreate a codec ladder here. AV1 waits for the same physical-encode
+proof and the desktop/mobile Viewer decode matrix; compression efficiency alone
+cannot advance it.
+
+Reproducibility anchors for the proprietary-package inspection, without copied
+code or user data:
+
+- Oopz `data/app.so`: SHA-256
+ `462E081D03E49008AD8D64A032BD6396F0B97CDADBDF715F7C6810D081A714F8`.
+- Oopz screen-viewer bundle `screenShare/page/assets/index-C6S4n49C.js`:
+ SHA-256
+ `1EAECD2C0DC332E5D2628047EE1BECB874C8130C2C383355EE313961EFC6E1CA`.
+- Oopz `agora_rtc_sdk.dll`: SHA-256
+ `66B8B34A57BA0EE9DCC6C516E2CC20F5F1EB5D35ED7150FDF37201A46BC8D4A0`.
+
## Deferred Mobile Sender Boundary
This section records later capability work; it does not broaden ADR-0006's
@@ -414,6 +453,10 @@ unless its own bounded capability and transport gate is accepted.
- [Pion send-side bandwidth estimator](https://github.com/pion/interceptor/blob/main/pkg/gcc/send_side_bwe.go)
- [Pion WebRTC license](https://github.com/pion/webrtc/blob/main/LICENSE)
- [WebCodecs](https://www.w3.org/TR/webcodecs/)
+- [NVIDIA NVENC application note](https://docs.nvidia.com/video-technologies/video-codec-sdk/13.1/nvenc-application-note/index.html)
+- [Agora Windows screen sharing](https://doc.shengwang.cn/doc/rtc/windows/basic-features/screen-share)
+- [Agora Windows encoding preference](https://doc.shengwang.cn/api-ref/rtc/windows/API/enum_encodingpreference)
+- [Oopz help center](https://help.oopz.cn/)
- [W3C WebRTC Statistics](https://www.w3.org/TR/webrtc-stats/)
- [W3C WebRTC SVC](https://www.w3.org/TR/webrtc-svc/)
- [W3C Media Capabilities](https://www.w3.org/TR/media-capabilities/)
diff --git a/docs/research/realtime-quality-adaptation.md b/docs/research/realtime-quality-adaptation.md
index 2aaa53b5..ab3236d6 100644
--- a/docs/research/realtime-quality-adaptation.md
+++ b/docs/research/realtime-quality-adaptation.md
@@ -122,12 +122,24 @@ software fallback fails the gate.
Discord's published Go Live material is a useful architecture comparison, not
a preset to copy. It describes native OS/driver-integrated capture and encoding,
-GPU hardware encoding, WebRTC transport, and product-specific rate-control
-tuning. It also documents a feedback loop that could lock 60 fps output down to
-30 fps. This supports measuring the complete capture/encoder/congestion loop;
-it does not show that Discord servers re-encode each viewer stream or that AV1
-is universally cheaper. No equivalent first-party implementation evidence was
-found for KOOK or Oopz, so they are not used as design facts.
+GPU hardware encoding, WebRTC transport, and one codec that the sender and all
+current viewers can decode. Discord lists VP8 and H.264 as its broad baseline,
+with HEVC and AV1 on selected platforms; its RTC worker routes the sender's
+stream rather than requiring one browser PeerConnection per viewer. This is
+evidence for a common publication, not proof of one physical encoder object.
+Discord also documents a feedback loop that could lock 60 fps output down to
+30 fps, accidental keyframes as often as once per second, and the resulting
+quality gains from a stable encoder lifetime, explicit keyframe cadence, and
+rate-control headroom. These findings support measuring the complete
+capture/encoder/congestion loop; they do not make AV1 universally cheaper.
+
+A read-only inspection of the locally installed Oopz 0.87.425 package is
+consistent with a native, hardware-first common publication: its Viewer bundle
+selects H.264, the package exposes a hardware-acceleration status, and the
+inspection found one screen-track publication. It does not prove the runtime
+codec/GPU choice or one physical encoder. Components, retained hashes, and the
+privacy boundary are owned by
+[Native shared-encode sender](./native-shared-encode-sender.md).
## Evidence Before Adaptation
@@ -607,3 +619,7 @@ is a separate optimization.
- [Jitsi desktop degradation preference](https://github.com/jitsi/lib-jitsi-meet/blob/master/modules/RTC/TraceablePeerConnection.ts)
- [Discord Go Live architecture](https://discord.com/blog/how-it-all-goes-live-an-overview-of-discords-streaming-technology)
- [Discord encoder-quality case study](https://discord.com/blog/from-blocky-to-brilliant-improving-video-quality-on-discord-go-live-on-amd-gpus)
+- [NVIDIA NVENC application note](https://docs.nvidia.com/video-technologies/video-codec-sdk/13.1/nvenc-application-note/index.html)
+- [Agora Windows screen sharing](https://doc.shengwang.cn/doc/rtc/windows/basic-features/screen-share)
+- [Agora Windows encoding preference](https://doc.shengwang.cn/api-ref/rtc/windows/API/enum_encodingpreference)
+- [Oopz help center](https://help.oopz.cn/)