WIP: R2 — schedule inference for historic replay / off-policy (draft, do not merge)#2607
Draft
jucor wants to merge 13 commits into
Draft
WIP: R2 — schedule inference for historic replay / off-policy (draft, do not merge)#2607jucor wants to merge 13 commits into
jucor wants to merge 13 commits into
Conversation
…g) vs brute-force oracle
…2e recovery green Iteration findings (probed on synthetic ground truth): - emission delay: votes during a recompute are served under the PREVIOUS weights; at 30s compute this is a whole segment — modeled as the delta-shift from the R2 design (emission changepoint = state changepoint + delay) - soft renewal idle prior: an up worker with pending votes recomputes at the first free poll; exponential penalty on unexplained idleness (soft = robust to unknown stalls; hard bound would be unsound) - count-constrained inference uses the dense lattice (thinning + spacing + exact count creates spurious infeasibility -> InfeasibleScheduleError) - honest metrics: posterior localization (median 1 vote, p90 2) and coverage@3 (>=0.92) both eras; MAP is a fragile summary in emission-flat stretches (single insertion cascades in sorted pairing)
…lized), smoke, perf guard vw finding: bursty real traffic (median inter-vote gap 0s) means localization must be judged in TIME — the irreducible uncertainty is the compute-jitter window. Achieved 9.8s median / 21s p90 on the vw skeleton (physics-limited).
…s) + plan execution notes
…bursty generators, soundness checks, metrics, CLI) + smoke tests
…ON + RESULTS.md, 60 runs)
jucor
force-pushed
the
jc/r2-schedule-inference
branch
from
d340500 to
683d295
Compare
Delphi Coverage Report
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
WIP / DRAFT — not for merge. Pushed to protect and make visible the R2
schedule-inference research chain (previously living only in the local
r2-inferencejj workspace, unbacked-up).What this is
R2 = replaying historic conversations for off-policy learning when the real
recompute schedule that actually took place is unknown and must be inferred
(unlike R1, which uses synthetic schedules we define). See
delphi/docs/REPLAY_HARNESS_DESIGN.md§1.3 / §10 for the framing.Contents (bottom → top)
docs(delphi): replay — R2 schedule-inference implementation planVoteEvent,ReplayDataset,Schedule)era vs later eras)
checks, metrics, CLI) + committed 60-run grid results
Code lives under
delphi/polismath/replay/.Status
Prototype/research — algorithms validated on synthetic + semi-synthetic data,
not yet wired to the Clojure-legacy Python engine (powerit-pca / sequential-bits
port). Kept as a draft PR for backup + visibility, rebased onto current
edge.