Operational route

Skill evolution

Turn repeated run lessons into candidate skill improvements with evals, shadow runs, human promotion, and rollback state.

Page evidence

Operational mapImprove from evidence without silent mutation.
  1. Reflectrun lesson
  2. Proposecandidate change
  3. Evaluatepositive + negative cases
  4. Shadowhistorical traces
  5. Promotehuman approval
  6. Rollbackversioned snapshot
/guild:evolve

Route comparison

Same job. A visible operating record.

Without a route record

Skill edits can become prompt drift when they skip evals, shadow mode, approval, or versioned rollback.

With Guild Stack

Guild Stack routes improvements through proposals, paired evals, flip reports, shadow mode, promotion gates, and versioned rollback.

Route stops

Request to verified result.

  1. 01

    Reflect

    Learning checkpoints and reflections identify candidate improvements.

  2. 02

    Evaluate

    Current and proposed behavior are compared with positive and negative cases.

  3. 03

    Shadow

    Proposed behavior is tested against historical traces without changing live routing.

  4. 04

    Promote or archive

    Human approval gates promotion; rejected candidates remain useful evidence.

Receipts

What the route leaves behind.

  1. 01
    Path .guild/reflections/
    Role

    Candidate improvements from run evidence.

  2. 02
    Path .guild/evolve/<run-id>/evals.json
    Role

    Working eval set.

  3. 03
    Path flip-report.md
    Role

    Behavior flips, regressions, and fixes.

  4. 04
    Path .guild/skill-versions/
    Role

    Versioned rollback snapshots.

Next

Follow this route through an illustrative run.

The demo traces one bounded task through approvals, specialist lanes, review, verification, and the final run record.