Skip to content

Measure p: read-only shadow mode (unblocks the whole roadmap) #471

Description

@phall1

The largest unforced error in the plan. Every quantitative conclusion in the roadmap is a function of p, the per-atom merge-time failure rate, and it has never been measured. There is currently no instrument in the design that could measure it without first taking exclusive write access to someone's trunk.

p decides: the drainability wall (p* = 1/(1+lambda*E[S]) — 24.2% sustained, 11.3% peak at our envelope), required speculation width, the entire cost table, Bar B's thresholds, cohort size n*, and whether the attribution design is even the maximum-likelihood answer.

Decompose it into four components, because only one of them justifies a merge queue at all. Every atom already passed its own PR CI before enqueue, so a merge-time failure is exactly one of:

  • (a) stale base — solved by requiring up-to-date branches, no queue needed
  • (b) spurious red q — a test-quality problem invariant 6 puts irreducibly into the throughput budget
  • (c) genuine interaction r — the thing a merge queue uniquely prevents. This is the number that decides whether the product has a reason to exist.
  • (d) replay conflict c — a separate absorbing state costing zero CI, strongly correlated rather than independent, compounding with prefix depth as 1-(1-c)^k. Plausibly binds speculation depth before p does at monorepo scale.

Eight independent analyses agreed p was the pivot and every one modelled it as a single number.

The strongest argument FOR the premise, which nobody in the corpus made: interaction failures scale as C(n,2)*r in the number of changes merging per CI window, so the case for a queue grows quadratically with agent concurrency while the case for merely faster CI grows only linearly.

Build: read-only shadow mode. contents:read, pull_requests:read, checks:read. For every merged PR, reconstruct the candidate the queue would have built using the existing merge-tree path. Yields c at zero CI spend, plus stack prevalence and depth distribution. On any repo already running GitHub MQ, harvest merge_group check conclusions as a free verdict oracle for p.

Cost: roughly what the M0a kernel cost. ~$0 CI. No trunk write access — that is the point. Unblocks W3's baseline, M2's width decision, and de-risks Bar A.

See docs/INVESTIGATION.md §6 Q1, §7 R3, §8 E1.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions