# Native Session Learning Acceptance

## Scope

The built-in DSH RSI settings now support selecting a live or persisted source Session, manually starting archive/extraction/independent screening, and optionally selecting a target Session for automatic baseline/memory text replay. A known outcome and benchmark adapter are not required. Configured workspace consent still applies. This is a general conversation-learning entry, not automatic production release or full environment reconstruction.

## Actual Run

- Date: October 17, 2026.
- Isolated profile: `artifacts/native-session-learning-s5yjomaw`.
- HTTP base: `http://127.0.0.1:52283`.
- Trigger: actual settings-page source/target selectors, checkbox and Start Learning button.
- Run: `8a1bd500-5dc8-4dd5-a861-511f1e4543cc`, completed.
- Source and target: `a0199841-7a46-4119-8284-551c81dd6e51`; explicitly recorded as in-sample.
- No benchmark adapter configured. Six pre-existing Sessions discovered through the native persisted/live corpus.
- Extracted three hypotheses; independent screening retained one and rejected two with recorded reasons.
- Extraction: 77,816 tokens, one model response, 24.14 seconds.
- Screening: 92,991 tokens, one model response, 82.57 seconds.
- Baseline: 1,856 tokens, one model response, 721.58 ms.
- Memory: 2,081 tokens, one model response, 755.63 ms.
- Both success outcomes remain unknown; no positive-benefit verdict or automatic publication.

## Evidence

`session-learning-audit.json` passes 14 checks: completed state, truthful unknown outcomes, no publication, source identity, distinct extraction/screening, equal paired user inputs, five distinct native phase/source Sessions, no task tools, logged native system prompts, absent baseline memory, actual candidate text delivered in memory system assembly, frozen mapping, no generation overrides, and usage reconciliation.

`session-learning-evidence.json` retains the complete source/target log snapshots, extraction/screening traces, accepted candidate and both original replay traces. The settings page displayed completed state, the accepted candidate, both metrics, native Session navigation buttons and full raw comparisons. Browser inspection confirmed baseline/memory system messages and `rsi:memory` only in the memory arm. A Harness `PAGE_SCRIPT_ERROR` warning occurred during navigation; the workflow completed and the page rendered, but this is not a clean-console acceptance claim.

## Limits

The selected finance source requires tool-provided business records. Text-only replay intentionally disables tools and omits original assistant replies; this run proves the general plugin workflow and delivery instrumentation, not successful finance-task replay or better outcomes. Attachments are rejected before learning when automatic replay is selected. Later user messages referring to original replies can make multi-turn replay non-equivalent. A trustworthy success verifier and controlled tool/environment replay remain separate follow-up work.

## Verification

Six focused launcher suites passed 42 tests, including source-only learning, unloaded persisted Sessions, paired fixed mapping, rejected attachments/workspace scope, durable restart, cancellation/awaited close and settings selection RPC. Launcher Host/Client type programs and both build faces passed. The fresh native profile exposed and exercised the real RPC path. The daily profile was not touched.