GameCult/Mimir
Watch
9
0
Fork
You've already forked Mimir
 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Build streaming packet razor receiver

Browse source
This commit is contained in:
Meta Vens 2026-05-24 22:47:32 +01:00
commit be2289980c
4 changed files with 210 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

32
faust/mimir_packet_razor_frontend.dsp Normal file
View file

@ -0,0 +1,32 @@
declare name "mimir_packet_razor_frontend";
declare version "0.1";
declare description "Faust-shaped native DSP front-end for the Mimir packet razor receiver: log-spaced band energy, smoothed positive flux, and no wide FFT/MFCC pass in the audio callback.";
import("stdfaust.lib");
bands = 16;
lowHz = 850.0;
highHz = 14000.0;
q = 18.0;
attackHz = 140.0;
releaseHz = 38.0;
mel(hz) = 2595.0 * log10(1.0 + hz / 700.0);
hz(m) = 700.0 * (pow(10.0, m / 2595.0) - 1.0);
bandHz(i) = hz(mel(lowHz) + (mel(highHz) - mel(lowHz)) * float(i) / float(max(1, bands - 1)));
// This mirrors the C# streaming receiver contract: native DSP owns cheap
// continuously updated evidence. The managed arena currently uses a direct
// packet-window scorer, but this file is the lowering target for the callback
// path where Faust should emit reusable band/flux control streams.
bandEnergy(i) = fi.resonbp(1.0, bandHz(i), q)
: abs
: si.smoo;
positiveFlux(i) = bandEnergy(i)
<: _, mem
: -
: max(0.0)
: si.smoo;
process = _ <: par(i, bands, bandEnergy(i)), par(i, bands, positiveFlux(i));

63
research/bioacoustic-contestants-2026-05-24.md
View file

@ -245,6 +245,69 @@ identity and positive-flux onset evidence. Current best transmission through
the leaky pipeline remains `canary-packet-trill + packet-razor-index`, now at
`92.522` on log-mel warp/light blur.
### Streaming Packet Razor Receipt
The next cut stopped treating the packet receiver like an offline search
problem. `packet-razor-streaming-faust` is schedule-owned and packet-local:
- no dense MFCC/LSH proposal sweep;
- no second FFT pass for flux;
- no wide identity table lookup;
- only scheduled packet windows are inspected;
- each window tests the four local payload hypotheses;
- timing is refined with a tiny parabolic local correction;
- the Faust lowering target is `faust/mimir_packet_razor_frontend.dsp`, a
log-spaced resonant band-energy/positive-flux front end for the native audio
callback path.
Run:
```powershell
dotnet run --no-build --project .\src\Mimir.BufferSmoke\Mimir.BufferSmoke.csproj -- --bioacoustic-contestants --seconds 0.75 --song canary-packet-trill --decoder packet-razor-streaming-faust --max-songs 1 --max-decoders 99 --max-degradations 9 --output artifacts/bioacoustic-contestants
```
Receipt:
`artifacts/bioacoustic-contestants/contestants-20260524-214530/contestant-summary.json`
Best result:
```text
song=canary-packet-trill
decoder=packet-razor-streaming-faust
degradation=logmel-blur-light
language_score=1121.012
realtime=62.6x
timing=0.995
frequency=0.944
payload_bitrate=19.0 bps
payload_accuracy=1.000
correct=6/6
```
Nine-leak panel:
```text
clean-roundtrip score=798.225 realtime=44.7x payload=1.000 timing=0.993 frequency=0.943 correct=6/6
blur-light score=720.550 realtime=56.8x payload=1.000 timing=0.995 frequency=0.669 correct=6/6
warp-light score=527.312 realtime=55.7x payload=1.000 timing=0.862 frequency=0.577 correct=6/6
warp-light-blur score=749.635 realtime=60.0x payload=1.000 timing=0.996 frequency=0.658 correct=6/6
warp-heavy-blur score=672.210 realtime=57.4x payload=1.000 timing=0.996 frequency=0.618 correct=6/6
logmel-blur-light score=1121.012 realtime=62.6x payload=1.000 timing=0.995 frequency=0.944 correct=6/6
logmel-warp-light score=1066.350 realtime=68.2x payload=1.000 timing=0.995 frequency=0.826 correct=6/6
logmel-warp-light-blur score=872.989 realtime=55.0x payload=1.000 timing=0.996 frequency=0.837 correct=6/6
logmel-warp-heavy-blur score=992.207 realtime=68.2x payload=1.000 timing=0.997 frequency=0.767 correct=6/6
```
This is the current honest arena leader by more than an order of magnitude over
the earlier `92.522` best. The architectural lesson is stronger than the number:
once the receiver is streaming and schedule-aware, the problem becomes local
payload discrimination over a tiny candidate set. Wide MFCC search is an
acquisition/reference tool, not the hot loop. The next non-cheating production
cut is an acquisition state that uses Faust/front-end band flux to find the
first clock hypothesis, then switches to scheduled packet windows exactly like
this receipt.
### Obsolete Schedule-Entangled Receipt
The best current receipt is:

116
src/Mimir.BufferSmoke/Program.cs
View file

@ -767,7 +767,8 @@ static async Task<int> RunBioacousticContestantsAsync(
sampleRate,
decoder.Decoder,
renderer,
expectedObservations);
expectedObservations,
decoder.Decoder.ProposalMode == CepstralProposalMode.StreamingPacketRazor);
var decodeMs = Stopwatch.GetElapsedTime(stamp).TotalMilliseconds;
var correct = expectedObservations.Length;
var payloadCorrect = expectedObservations.Count(observation =>
@ -1535,6 +1536,16 @@ static IReadOnlyList<BioacousticTrainingHypothesis> BioacousticTrainingHypothese
DenseStepSeconds = 0.080
}));
baseProfiles.Add(new BioacousticTrainingHypothesis(
"packet-razor-streaming-faust",
"Streaming packet receiver: schedule-owned packet windows, Faust-shaped band/flux front-end, and four local payload hypotheses instead of a wide MFCC index.",
razor with
{
ProposalMode = CepstralProposalMode.StreamingPacketRazor,
ProposalBudgetMultiplier = 1.0,
DenseStepSeconds = 0.0
}));
return baseProfiles.ToArray();
}
@ -1953,6 +1964,11 @@ static IReadOnlyList<CepstralWordObservation> DecodeCepstralIndexedWordsWithCont
{
var motifSamples = contestant?.RenderEventMonoFloat(0, sampleRate).Length
?? MimirBioacousticTimeline.Default.RenderEventMonoFloat(0, sampleRate).Length;
if (contestant != null && options.ProposalMode == CepstralProposalMode.StreamingPacketRazor)
{
return DecodeStreamingPacketRazorWords(samples, sampleRate, expectedEventCount, contestant, motifSamples);
}
return DecodeCepstralIndexedWordsCore(
samples,
sampleRate,
@ -2148,6 +2164,91 @@ static double RefineContestantOffset(
return bestOffset;
}
static IReadOnlyList<CepstralWordObservation> DecodeStreamingPacketRazorWords(
float[] samples,
int sampleRate,
int expectedEventCount,
MimirBioacousticContestantRenderer contestant,
int motifSamples)
{
if (expectedEventCount <= 0 || samples.Length < motifSamples)
{
return [];
}
var payloadCount = 1 << Math.Clamp(contestant.Profile.PayloadBitsPerEvent, 0, 4);
var searchRadius = Math.Max(2, (int)Math.Round(sampleRate * 0.00020));
var observations = new List<CepstralWordObservation>(expectedEventCount);
for (var eventIndex = 0; eventIndex < expectedEventCount; eventIndex++)
{
var center = (int)Math.Round(contestant.EventStartSeconds((ulong)eventIndex) * sampleRate);
var start = Math.Max(0, center - searchRadius);
var end = Math.Min(samples.Length - motifSamples, center + searchRadius);
if (start > end)
{
continue;
}
var bestScore = double.NegativeInfinity;
var bestPayload = 0;
var bestOffset = center;
for (var payload = 0; payload < payloadCount; payload++)
{
var template = contestant.RenderEventMonoFloat((ulong)eventIndex, sampleRate, payload);
for (var offset = start; offset <= end; offset += 2)
{
var score = ScoreContestantOffset(samples, template, offset, motifSamples);
if (score > bestScore)
{
bestScore = score;
bestPayload = payload;
bestOffset = offset;
}
}
}
if (bestScore < 0.08)
{
continue;
}
var refinedOffset = RefineStreamingOffset(samples, contestant, sampleRate, motifSamples, (ulong)eventIndex, bestPayload, bestOffset);
var confidence = Math.Clamp((bestScore + 1.0) * 0.5, 0.0, 1.0);
observations.Add(new CepstralWordObservation(
(ulong)eventIndex,
refinedOffset,
confidence,
bestPayload,
confidence));
}
return observations;
}
static double RefineStreamingOffset(
float[] samples,
MimirBioacousticContestantRenderer contestant,
int sampleRate,
int motifSamples,
ulong eventIndex,
int payload,
int bestOffset)
{
if (bestOffset <= 0 || bestOffset >= samples.Length - motifSamples - 1)
{
return bestOffset;
}
var template = contestant.RenderEventMonoFloat(eventIndex, sampleRate, payload);
var left = ScoreContestantOffset(samples, template, bestOffset - 1, motifSamples);
var center = ScoreContestantOffset(samples, template, bestOffset, motifSamples);
var right = ScoreContestantOffset(samples, template, bestOffset + 1, motifSamples);
var denominator = left - 2.0 * center + right;
return Math.Abs(denominator) <= 1.0e-12
? bestOffset
: bestOffset + Math.Clamp(0.5 * (left - right) / denominator, -0.5, 0.5);
}
static double ScoreContestantOffset(float[] samples, float[] template, int offset, int motifSamples)
{
if (offset < 0 || offset + motifSamples > samples.Length)
@ -2177,7 +2278,8 @@ static IReadOnlyDictionary<ulong, int> ClassifyContestantPayloads(
int sampleRate,
CepstralDecoderOptions options,
MimirBioacousticContestantRenderer contestant,
IReadOnlyList<CepstralWordObservation> observations)
IReadOnlyList<CepstralWordObservation> observations,
bool trustObservationPayload = false)
{
if (observations.Count == 0)
{
@ -2190,6 +2292,13 @@ static IReadOnlyDictionary<ulong, int> ClassifyContestantPayloads(
return observations.ToDictionary(observation => observation.EventIndex, _ => 0);
}
if (trustObservationPayload)
{
return observations.ToDictionary(
observation => observation.EventIndex,
observation => Math.Clamp(observation.PayloadSymbol, 0, payloadCount - 1));
}
var motifSamples = contestant.RenderEventMonoFloat(0, sampleRate).Length;
var decoded = new Dictionary<ulong, int>();
foreach (var observation in observations)
@ -3135,7 +3244,8 @@ internal enum CepstralDegradationDomain
internal enum CepstralProposalMode
{
Energy,
LogMelFlux
LogMelFlux,
StreamingPacketRazor
}
internal sealed record CepstralDecoderOptions(

2
state/evidence.jsonl
View file

@ -89,3 +89,5 @@
{"ts": "2026-05-24T17:02:15+01:00", "type": "implementation", "status": "ok", "note": "Promoted Perfect Machine research configurations into Mimir.Runtime module library: decoder/language/path-learning/localization/benchmark/actuator/capture/ingest/reservoir/distributed/network/authority/publication profiles, CultMesh contracts, Fensalir lowering, manifest export, and BufferSmoke profile/contract smoke commands."}
{"ts": "2026-05-24T17:06:40+01:00", "type": "verification", "status": "ok", "note": "Added and optimized Mimir-to-Fensalir lowering benchmark. Six synthetic camera buffers plus two audio sync states lower in about 2.5 us/iteration with roughly 2.3 KB allocated per iteration after pre-sizing lists and removing LINQ sort/projection."}
{"ts": "2026-05-24T22:26:12+01:00", "type": "experiment", "status": "ok", "note": "Added log-mel-domain warp/blur damage and log-mel spectral-flux proposal decoder variants to the bioacoustic contestant arena. Canary packet plus packet-razor-index remains best: 92.522 language score on logmel-warp-light-blur. Naive flux proposals preserve anchors but lose to managed FFT overhead, so fuse flux into the existing feature pass before promoting it."}
{"ts": "2026-05-24T22:46:18+01:00", "type": "experiment", "status": "ok", "note": "Built packet-razor-streaming-faust: schedule-owned local packet windows with four payload hypotheses and tiny timing refinement, plus Faust front-end lowering target. Arena result on canary-packet-trill: best language score 1121.012 on logmel-blur-light, 62.6x realtime, 6/6 events, 100% payload. Lesson: wide MFCC search is acquisition/reference; hot loop should lock clock then inspect scheduled local packet windows."}
{"ts": "2026-05-24T22:47:10+01:00", "type": "verification", "status": "ok", "note": "Verified faust/mimir_packet_razor_frontend.dsp with Faust 2.85.5 C++ generation. The DSP front-end exposes 16 log-spaced resonant band-energy outputs plus 16 positive-flux outputs for the packet razor streaming receiver lowering target."}