sds
bayes for days
Star
Fork
Watch

Commits on claude/tactics-precision
docs(tactics): the firing phase measures the basis, not the tactics
committed
docs(tactics): the corrected damage model moves no tactic pair
committed
feat(tactics): measure how precise an agreement figure is
committed
perf(candidates): answer a volley without building its key
committed
docs(harness): the control for the seat split, 19-18
committed
feat(candidates)!: backing up is the default, and its rules are measured
committed
feat(candidates): a Mek may back up
committed
docs(candidates): three quarters of pairs never reach the weapon walk
committed
docs(candidates): what the line of sight and the profile actually cost
committed
docs(los): correct the end-to-end figure to what the mechanism predicts
committed
docs(los): what a warm line of sight hit costs
committed
perf(los): make a cache hit one lock and one copy
committed
perf(candidates): time the benchmarks on CPU, not on the wall
committed
perf(candidates): benchmark the sweep, which is 93% of the runtime
committed
docs(tactics): what would make advance orderable, and what blocks it
committed
feat(tactics): four generic designs, one per weight class
committed
feat(tactics): a lance vignette per tactic, scored not drawn
committed
refactor(tactics)!: advance is closing to melee range, not moving forward
committed
refactor(tactics)!: screen, reposition and clinch are not built
committed
refactor(tactics)!: implemented or unimplemented, and nothing else
committed
docs(tactics): engage is the hub, and a Task stores nothing
committed
docs(tactics): what the matrix said, and a figure for every tactic
committed
feat(tactics): the vocabulary, and the tool that measures it
committed
docs: fix claudeisms
committed
docs(candidates): where the sweep actually spends its time
committed
perf(candidates): resolve each weapon once a pair, not once a twist
committed
perf(candidates): sample the volley timings instead of taking every one
committed
fix(harness): the by-round table averaged totals, not turns
committed
perf(harness): measure both bots core-for-core
committed
fix(harness): cap a seat's heap
committed
fix(harness): notice a seat that died instead of waiting for the stall
committed
docs(harness): why the turn clock needs pinning to mean anything
committed
feat(harness): pin each seat to a physical core of its own
committed
refactor(harness)!: run each bot seat in its own JVM
committed
docs(plan): decompose strategies into tactics, scoring and objectives
committed
feat(training): --drop-displaced, and what it moves
committed
docs(plan): the questions the asymmetric run left open
committed
feat(scenarios): draw the two sides independently
committed
fix(harness): name the runner image after its Dockerfile
committed
fix(harness): swap sides per scenario, not per game
committed
feat(harness): read decision logs compressed
committed
refactor(features): one palette, read by both figure generators
committed
docs(training): LABELS.md, every label over one worked match
committed
fix(harness): stride the suite when a run is shorter than it
committed
fix(training)!: read the engagement gap from one side
committed
feat(training): sds labelrank, does a label point at winning
committed
docs(plan): who the bot is for, and three more epics
committed
docs(plan): what the audit and the first fit turned up, in their epics
committed
docs(plan): nine epics for the game the bot does not play yet
committed
docs(plan): what done looks like, and two epics for what is missing
committed
docs: sync the feature branch workflow with atgc 0.15
committed
feat(training): a fitted vector records the basis it was fitted against
committed
docs(training): the graft that played cleanly and lost 37 of 38
committed
feat(training): warn as a fit is written, not only as one is loaded
committed
feat(training): warn when a weight vector is lopsided
committed
fix(features): retire target_skill so old vectors still load
committed
feat(observability): what each feature actually reported
committed
docs(los): what the match-lifetime cache actually grew to
committed
test(observability): a reading that went backwards is a cleared cache
committed
feat(observability): report the los cache's growth per turn
committed