# Modular Runtime Phase 5 — Release and Comparison Evidence Gate

Date: 2026-08-01

Status: evidence gate accepted; authority transition deliberately not executed

## Accepted outcome

Phase 5 now has the evidence required to recommend the modular runtime as Nyx's successor:

- the isolated release recipe accepted one exact frozen commit, package, publication, rollback proof, restart, and terminal graph replay;
- the modular runtime won the matched direct lane without elapsed, call, retry, or orphan regression;
- the modular runtime won the matched complex lane, accepted all four integrated behaviors, and completed in 51.94% of legacy elapsed time;
- both comparisons were content-addressed before results, preserved exact model routes and process outcomes, used zero human checkpoints, and passed independent offline revalidation;
- production comparator code derived `evidence_supports_switch` from the two immutable lane verdicts.

This accepts the release and comparison evidence. It does not silently change which runtime owns production dispatch.

## Gate matrix

| Phase 5 requirement | Evidence | Result |
|---|---|---|
| Isolated release recipe | [`MODULAR_RUNTIME_PHASE5_RELEASE_CANARY_01.md`](./MODULAR_RUNTIME_PHASE5_RELEASE_CANARY_01.md) | accepted |
| Matched direct work | [`MODULAR_RUNTIME_PHASE5_DIRECT_SHADOW_01.md`](./MODULAR_RUNTIME_PHASE5_DIRECT_SHADOW_01.md) | modular wins |
| Matched complex work | [`MODULAR_RUNTIME_PHASE5_COMPLEX_SHADOW_02.md`](./MODULAR_RUNTIME_PHASE5_COMPLEX_SHADOW_02.md) | modular wins |
| Direct no-regression threshold | 0.9983× elapsed, equal calls, equal retries, equal orphans | accepted |
| Complex material-improvement threshold | 0.5194× elapsed and accepted versus legacy `needs_review` | accepted |
| Independent evidence replay | both comparison documents and aggregate authority derivation recomputed | accepted |
| Human checkpoint count | zero in release, direct, and complex canaries | accepted |
| Authority switch and legacy archive | no production mutation performed | pending explicit transition authority |

## Comparative facts

| Lane | Legacy | Modular | Delta | Verdict |
|---|---:|---:|---:|---|
| Direct | 259.806 s, 2 calls, not integrated | 259.372 s, 2 calls, accepted | −0.434 s | `modular_wins` |
| Complex | 467.083 s, 4 calls, not integrated | 242.608 s, 4 calls, accepted | −224.475 s | `modular_wins` |

Neither lane retried or leaked a process. The distinction was authority design: legacy assigned Git integration to a workspace-restricted model worker; the modular runtime kept Git sealing, verification, promotion, replay, and state transitions in deterministic trusted code.

Subscription-backed model calls do not expose a trustworthy per-call dollar amount, so cost is recorded as `subscription_unpriced` rather than fabricated. Equal call counts and exact routes are preserved for later priced comparison.

## Current-revision regression

Immediately before the accepted complex canary:

- 235 deterministic tests passed across 49 active suites;
- 20 gated live tests were skipped in the deterministic lane;
- TypeScript passed with no diagnostics;
- 72 runtime source files averaged 183 logical lines, with p50 144 and p90 409;
- no runtime source file exceeded 494 logical lines;
- `git diff --check` passed.

The comparison harness additionally preflights all complex implementation packets before launching any model. Revision 01 remains in evidence as the reason that guard exists.

## Authority boundary

The evidence satisfies the declared technical switch criterion: match or win direct work and materially improve complex work. The next operation is nevertheless a migration, not another test. It changes the production dispatcher, rollback target, and legacy lifecycle. That operation must have explicit transition authority and must preserve:

1. a frozen legacy snapshot and read-only audit path;
2. an atomic selector or equivalent reversible authority pointer;
3. a documented rollback command and rollback proof;
4. one post-switch direct smoke task and one graph smoke task;
5. automatic rollback if either smoke task fails;
6. no deletion of legacy source or evidence.

Until that bounded transition is authorized and succeeds, Phase 5 is not described as fully complete and legacy Core remains the production authority. Phase 6 Hermes remains deliberately later.