# Pyx: a field guide to merge decisions

Field guide — spoken transcript.

Pyx. A field guide to merge decisions.

This field guide helps engineering leaders inspect a proposed automatic-merge decision. Assay coordinates delivery by agents under human governance; Pyx is our merge-risk assessment work. We'll use an illustrative research-note PR to explain category admission, the current six-signal score, missing evidence, fresh-head checks and lane health. You'll know what record to inspect and what still needs implementation. The evaluator exists; automatic merge writing and the broader learned feedback loop remain incomplete.

Our example changes one dated research note inside an opted-in area. Every changed path must fit that permission, including the old path of a renamed file. The author must be the configured worker App, with a linked work item and an acceptable surface classification. Stream briefs, risk-classed changes and agent instructions are excluded. Only a reviewer-applied admission label counts, and later evaluations rederive the decision.

For this note correction, record the worker author, the reviewer's admission event and revision Head A. The evaluator combines hosting-platform records with repository policy and operator configuration. Review history, current head, check results and label events come from the forge. Label events establish who applied a size or model record. A dispatcher-attested model stamp says what was launched; it does not prove correctness. An author's statement that a change is low risk cannot replace these inputs.

The score adds one point for each fired reason: review rework, CI non-success, a head changed after a request, large size, missing model attestation, or unreadable evidence. For our PR, a zero requires readable evidence with none of these signals. The initial house ejection line is zero. Reasons can overlap: an old request for changes can fire both rework and changed-head signals. Read the reasons with the number; this is a rule-based count, not a defect probability.

For checks, the latest run of each name supersedes older reruns. Completed success, neutral and skipped conclusions are accepted; commit status contexts must say success. Empty, pending or truncated results do not establish green. This describes reported checks, not proof that every relevant test ran. Your required-check coverage and validation against the latest target branch still need their own controls.

When our research-note PR is waiting for CI, unreadable evidence alone returns could-not-check without a permanent ejection. A confirmed demotion reason changes the route. In an activated lane, the command removes admission, records the reason and sends the PR to human review. Once ejected, that PR does not automatically return just because a later evaluation looks clean.

Suppose our note now has readable evidence, a zero score and reviewer approval at Head A. The final gate also requires an open, ready PR, no earlier ejection, acceptable checks, mergeability and an armed lane. It resolves human activation and rereads the head. Our worker has pushed Head B during that evaluation: the command refuses. Head A's approval cannot authorize Head B.

The inspected public implementation provides evaluation, gated admission and ejection, and a merge dry run. It deliberately has no final merge mutation. A non-dry-run call that passes every condition still refuses at the write boundary. The house activation sign-off remains empty too. Verify the exact consumed release and enablement before describing any installation as an automatic-merge service.

The initial policy uses a three-per-day repository cap and a ninety-percent first-pass-yield floor over twenty-eight days, once ten lane merges exist. Below that sample, yield is an early state; absent data is a hold. The current collector omits verification-failure history from this ratio. The full sticky hold, failure response and human re-arming control still need end-to-end verification before operation.

For learning, repmetrics is the collector that links main-branch CI failure episodes to candidate merged changes. A green-to-red transition identifies an episode, not necessarily its cause. Several intervening changes can leave an ambiguous range. Current captures contain real episode records, while the audited-label register and model results remain pending. Preserve ambiguity and audit the evidence before training on it.

The proposed progression is transparent rankings, an expert scorecard, then a learned model if audited data supports it. Keep features and labels within their historical knowledge cutoff. Train earlier and test later. Compare with a lines-added-only baseline, reporting incident prevalence, false alarms and recall within a review-effort budget. Qualgen's separate learned-risk library does not establish that this Pyx benchmark is complete.

Our example's record now reads: permitted research note, Head A evaluated, Head B found on re-read, action refused. Retain those revisions, the evidence states, named reason and route. Keep the rule version and authority with the record. After any future activation, measure outcomes and the cost of additional review; a score's accuracy alone does not prove an intervention helped. Use the article for the detailed source boundary and examples.

For more Assay explainers, like and subscribe. You'll find the complete guide at Assay dot finance.