burrito.space / bez
This repository has no description
Star
Fork 0
Watch

Commits on track/rig-throughput

Keep the roster: restore handoff.md to master track/rig-throughput
dietrich ayala committed
Rig capacity: rig runs on DGX Sparks; measured gateway batch rates for Qwen and GLM
dietrich ayala committed
Roadmap: GLM batch through the gateway, 0 of 8 compiled on mid-analysis fragments
dietrich ayala committed
Model client: log each candidate as it arrives
dietrich ayala committed
Throughput bench: per-model raw results directory; bearer auth for the gateway
dietrich ayala committed
Roadmap: Qwen batch through the gateway, 6 of 8 compiled, 119/227 admitted
dietrich ayala committed
Model client: read null gateway content as empty; show a clipped body on parse failure
dietrich ayala committed
Handoff: GLM reasoning leak diagnosed; gateway batches next
dietrich ayala committed
Record GLM-5.3 Flash throughput on the rig, stopped after the cap was confirmed
dietrich ayala committed
Gateway bearer auth for HttpModelClient; say there is no Percentage variant
dietrich ayala committed
Measure the rig's vLLM server and run a first rig-generated batch
dietrich ayala committed
Roster: rig gateway is the sanctioned model route
dietrich ayala committed
Roster: feature-cost landed; per-feature model cost measured
dietrich ayala committed
Feature-cost: model-driven runs on both seams; measured per-feature model cost
dietrich ayala committed
Feature-cost: give reflection candidates the url crate and say so in the prompt
dietrich ayala committed
Feature-cost: restore the enumerated-reflection getter branch
dietrich ayala committed
Feature-cost: real spec text for the reflection and mutation seams
dietrich ayala committed
Feature-cost: model-run binary for the reflection and mutation seams
dietrich ayala committed
Feature-cost: render script, probes and the voted expected result in failing cases
dietrich ayala committed
Feature-cost: prompts for the reflection and mutation seams
dietrich ayala committed
Feature-cost: follow master compile(target_dir) and Candidate stop_reason
dietrich ayala committed
Feature-cost: candidate plumbing and fixture loop for both seams
dietrich ayala committed
Feature-cost: case corpus, ground-truth cache and scorer for both seams
dietrich ayala committed
Feature-cost: WPT subset, reflection/mutation oracle protocol, and dom seams
dietrich ayala committed
Roster: open rig-throughput; GPU-host login is the next thing the admin grants
dietrich ayala committed
Roster: shared rig machine is an admin VM with no GPU or model server yet
dietrich ayala committed
Roster: candidate-fixes landed; width equation admitted; feature-cost rebases next
dietrich ayala committed
roadmap.md: record the second Claude batch on the width equation
dietrich ayala committed
Fix the three candidate-fixes problems from the first Claude batch
dietrich ayala committed
Roster: open candidate-fixes; lands before feature-cost
dietrich ayala committed
Roster: claude-model-run landed; candidate fixes and width rerun are next
dietrich ayala committed
roadmap.md: record the first Claude batch on the width equation
dietrich ayala committed
generation: add a Claude candidate source beside the OpenAI-compatible one
dietrich ayala committed
Roster: open feature-cost; its model run follows claude-model-run
dietrich ayala committed
Roster: claude-model-run paused before any work; resume instructions
dietrich ayala committed
Roster: open claude-model-run
dietrich ayala committed
Roster: generation-loop landed; model endpoint that completes requests is the blocker
dietrich ayala committed
Drop duplicate .env ignore rule
dietrich ayala committed
roadmap.md: record step 4's width-equation measurements
dietrich ayala committed
model: cap completion length so one candidate cannot stall the whole batch
dietrich ayala committed
real_model_run: sample one failing case per category instead of sending every one
dietrich ayala committed
Isolate the block-width seam and stand up the generation loop's model-independent half
dietrich ayala committed
Roster: prune to goal-required next steps; point optional follow-ups at their docs
dietrich ayala committed
Roster: corpus-shapes landed
dietrich ayala committed
Model the border-box height clamp and close two corpus gaps
dietrich ayala committed
Roster: rig access via machine sharing; open corpus-shapes
dietrich ayala committed
Roster: webkit-webm landed
dietrich ayala committed
Update WebGL figures for WebKit with WebM decoding fixed
dietrich ayala committed
Let WebKit suite runs play WebM with a header-only Vorbis track
dietrich ayala committed
Ignore .env: model endpoint and credentials stay local
dietrich ayala committed
Roster: laptop model reachable for loop runs; rig access still blocked
dietrich ayala committed
Roster: bootstrap-random-corpus landed; remaining corpus gaps
dietrich ayala committed
Run engine bootstrap checks over the seeded random corpus, add dom unit tests
dietrich ayala committed
Roster: provenance landed; open generation-loop
dietrich ayala committed
Add the provenance record for generated engine functions
dietrich ayala committed
Roster: economics-conformance landed; WebGPU independence follow-up
dietrich ayala committed
webgpu-cts.md: Chromium WebGPU is Dawn, so cross-runtime agreement is a second backend, not an independent implementation
dietrich ayala committed
Carry measured conformance coverage into the economics estimate
dietrich ayala committed
Roster: margin-collapse-recipes landed; border-box clamp follow-up; generation model access blocked
dietrich ayala committed
Update fragment_check informational count to 9/25 in platform-areas and content-scoped-engine
dietrich ayala committed