# Modular Runtime Phase 5 — Matched Direct Shadow 01

Date: 2026-08-01

Status: accepted matched comparison; modular wins

Measured comparison time: 519.178 seconds across both runtimes

## Sealed comparison

Legacy Core and the modular runtime received clones of one content-addressed Git fixture, the same accepted base tree, one exact owned source file, the same product contract, the same acceptance command, and the same unattended budget. The comparison contract was sealed before either result existed.

Canary SHA-256: `db2d41db0a4e99c85135cb3576ac3e97384df2802b24fba870b694b0ba7d63c8`.

The direct-lane threshold was also frozen before execution: modular had to accept the same contract and avoid a material regression. When both systems accepted, modular elapsed time could be at most 1.25× legacy, model calls could increase by at most one, and retries or orphaned processes could not worsen. Correctness remained primary: accepting the sealed contract while the other runtime did not is a win.

## Result

| Measure | Legacy Core | Modular builder |
|---|---:|---:|
| Contract outcome | `needs_review` | `accepted` |
| Elapsed time | 259.806 s | 259.372 s |
| Model calls | 2 | 2 |
| Attempts | 2 | 1 |
| Retries | 0 | 0 |
| Orphaned processes | 0 | 0 |
| Integration promotions | 0 | 1 |
| Final acceptance check | failed on unchanged authority | passed |

Comparator verdict: `modular_wins`.

Measured elapsed ratio: `0.9983295228`; modular finished 434 milliseconds sooner with no model-call, retry, or orphan regression.

## Why legacy did not accept

The legacy implementation worker completed its exact one-file change and returned accepted change-set evidence. Legacy then assigned integration to another Luna worker under `workspace-write`. That worker could read the authoritative repository but could not create `.git/ORIG_HEAD.lock`, so its required `git merge --no-ff` could not promote the accepted dependency. The authoritative fixture remained unchanged and the final acceptance command therefore failed.

This was not inferred from a missing output. Legacy returned a terminal `needs_review` envelope naming the Git-metadata blocker, and the comparison recorded the unchanged authoritative tree.

The modular runtime used one Sol/high planning call and one Luna/xhigh implementation call. Trusted deterministic code sealed the exact unstaged owned diff, ran the acceptance command against the candidate, promoted it through the one integration queue, and left the authoritative repository clean.

## Durable evidence

Evidence directory: [`evidence/MODULAR_RUNTIME_PHASE5_DIRECT_SHADOW_01`](./evidence/MODULAR_RUNTIME_PHASE5_DIRECT_SHADOW_01/)

`run-summary.json` SHA-256: `18dad1ac18f5a76167ff435079c2ac1ba7da64f05d9b010b7469f95cd10dba4f`.

Content-addressed comparison SHA-256: `a17afaba261f3ac6b71bf2eeb1d1883b2f17c94f77a1dc694cee9cad8ca9f1b6`.

An independent evidence-review test reparsed the sealed canary and both run envelopes, recomputed the comparison, and matched the stored comparison hash without running either model again.

## Boundary

This canary proves one matched direct scope. It does not alone justify an authority switch; the Phase 5 rule also requires a material improvement on matched complex work.