MILESTONE 5 FINAL PROOF
=======================

WHAT THIS PROVES
  The engine selects a real mission, determines safely where to work, provides
  minimal context, runs an agent, validates the result and produces a persistent
  isolated change -- without human intervention during the run.

    orchestrator -> mission selection -> target resolution -> workspace isolation
    -> AgentRunner -> structured outcome -> baseline-aware testing -> validation
    -> verdict -> commit -> persistence

WHAT THIS DOES NOT PROVE
  Anything about a language model. The agent is a DETERMINISTIC process applying
  a declared edit. No headless coding agent exists on this machine, and a
  reproducible run is what proves a harness. Swapping it for a real coding agent
  is a configuration line and touches no Core code.

  This is an AgentRunner integration proof, not a CoderAgent autonomous proof.

TARGET REPOSITORIES
  Real git repositories with real tests. In both paths one test failed before the
  run, for a reason stated in the task. Path B additionally carries a plausible
  distractor repository that the engine had to examine and reject.

============================================================
 MILESTONE 5 FINAL PROOF
============================================================

PATH A: DECLARED target
------------------------------------------------------------
  verdict:      READY_FOR_REVIEW
  test result:  PASSED
  commit:       ce09c7342e4c9f37b9bbc189e6c7541f8027cf2d
  branch:       regente/slug-1
  OK    source main unchanged                      27e672daeab61007e5b90648ea919ff3128b4ebf
  OK    source working tree clean                  0
  OK    no branch leaked into source               main
  OK    no worktree in source                      1
  OK    area base branch unmoved                   27e672daeab61007e5b90648ea919ff3128b4ebf
  OK    task file untouched                        TO DO
  target source: DECLARED  (DECLARED never promotes: nothing was discovered)

PATH B: DISCOVERED target
------------------------------------------------------------
  verdict:      READY_FOR_REVIEW
  test result:  PASSED
  commit:       da215dc09de3698baa9452f7fd0708b39b626380
  branch:       regente/sg-91
  OK    source main unchanged                      fe638f1f06aff1dd2f9b2d253169f91acdc0eab6
  OK    source working tree clean                  0
  OK    no worktree in source                      1
  OK    distractor repo untouched                  0
  OK    area base branch unmoved                   fe638f1f06aff1dd2f9b2d253169f91acdc0eab6
  OK    task file untouched                        TO DO
  target source: VALIDATED   confidence OBSERVED  strength 40
    evidence:    sibling_branch: 'feat/SG-90-related-work' belongs to sibling SG-90
    evidence:    module: contains 'src/normalize.py', named by the task
    rejected:    acme/billing: examined; no evidence linked it to this task

NEGATIVE 1: agent edits, validation fails
------------------------------------------------------------
  verdict:      REGRESSED
  OK    agent DID edit the file                    1
  OK    NO COMMIT                                  0
  OK    source main unchanged                      fe638f1f06aff1dd2f9b2d253169f91acdc0eab6
  OK    task file untouched                        TO DO
  target NOT promoted: DISCOVERED
  commit events: 0

NEGATIVE 2: agent tries to escape its workspace
------------------------------------------------------------
  verdict:      FAILED
  OK    source file byte-identical                 114cd886e87eacaafb831629b85e26fb
  OK    no OWNED marker in source                  0
  OK    source main unchanged                      fe638f1f06aff1dd2f9b2d253169f91acdc0eab6
  OK    NO COMMIT                                  0

GLOBAL: forbidden mutations, across every run above
------------------------------------------------------------
  OK    no push (no remote reachable)              https://github.com/acme/slugger.git
  (the remote is a non-existent URL; any push would have failed loudly)
  OK    no PR / merge / deploy                     0

============================================================

READING THE TWO PATHS
  Path A used a DECLARED target and was never promoted -- nothing was discovered,
  so there is nothing to confirm.

  Path B discovered its target from two independent signals, recorded the
  rejected alternative with its reason, and was promoted to VALIDATED only
  because the repository's own tests went from red to green after a change in it.
  Had the baseline been green, promotion would have been refused: staying green
  demonstrates nothing about whether the target was right.

  Negative 1 is the important one. The agent edited the file AND reported
  "ready to commit". The engine still refused: validation failed, so no commit,
  no promotion, no external mutation. The agent's opinion never entered the
  decision.
