FIN Weekly Dispatch: 2026-W26 (Exemplar #1 candidate)
The heal loop closed for real this week: central execute armed in production
FIN moved from guarded proving into live, gate-first operation: more autonomy under tighter invariants, better observability, and reversible rollout mechanics.
This week’s work pushed FIN’s agentic stack from guarded proving into live, gate-first operation: the biggest step was arming central Hermes execute in production while simultaneously standing up real invariant scoring, qualifying-read protections, and proof trails around it. In parallel, the team turned “heal” from a placeholder into an actual closed loop, with genuine tenant self-heal and a central content-heal path reaching real production targets. The through-line was clear: more autonomy, but only with tighter invariants, better observability, and reversible rollout mechanics.
What this exemplar is
The rendered dispatch is the sibling HTML artifact: 2026-W26-dispatch.html. This MD twin is its wiki-linkable counterpart (every HTML artifact in the vault ships one). The dispatch covers the full anatomy of DISPATCH_ANATOMY.md sections 0 to 8: brief, verdict, lede, numbers-that-matter, story beats by channel, foundations, outlook, take-action, and provenance with a gap note. Editorial copy is harvested from the weekly wire digest github_activity_digest_weekly_2026_W26; every number is read from the period’s digest payload.
By the numbers
- PRs opened (work started): 217 (prior period 147, as of 2026-06-21)
- PRs merged (work completed): 213 (prior period 132)
- Issues closed (work completed): 0 (prior period 1)
- Commits (supporting signal): 213 (prior period 131)
- Repositories touched: 2 (prior period 2)
Reproduce: pnpm digest:gh --type weekly --date 2026-06-22 and the digest payload at docs/digests/github-activity/weekly/2026-W26.md.
Highlights (harvested)
- Central execution crossed an important threshold with #1613, which armed
FIN_CENTRAL_HERMES_EXECUTE_ENABLEDin production in a gate-first, fail-closed, reversible way, the clearest sign this period that autonomous central actions are moving from shadow mode into controlled reality. - The safety case around that move got materially stronger through #1587, a real invariant scorer for
resolver_routing_replay, and #1589, which flippedHERMES_EVAL_GATE_MODEfrom shadow to enforce in prod, backed by proof work in #1622. - Loop closure became substantive rather than symbolic with #1618, replacing hollow tenant repair with genuine tenant knowledge self-heal, then hardening and extending that path in #1623.
- Central healing also matured quickly: #1625 reworked central heal into a true source→profile JOIN/fold, and #1626 activated central content-heal against 125 real production targets alongside tenant heal, still under gate-first controls.
- The central knowledge base gained a much firmer health contract through #1606, which introduced a real-invariant
central-kb-health-v1scorer, followed by fixes in #1608 and #1610 to land live-harvest end-to-end and get that health check to persist a pass. - Data-engine reliability work kept pace with the autonomy push: #1602 added a backlog-adaptive sharded projection-rebuild drain, and #1604 armed the production batch path with the reaper, sharded drain, and scheduled crons to close the drift loop.
Provenance
- Renderer:
newsroom-dispatch-html@1.0.0withnewsroom-dispatch-tokens@1.0.0 - Orchestrator:
scripts/newsroom/build-exemplar-w26.ts(newsroom-build-exemplar@1.0.0) - Pinned origin/master SHA:
87b808f8781c44502f565258236c27333e846289 - Editorial copy harvested from:
docs/digests/github-activity/weekly/2026-W26.md(editorial: true) - Weekly digest regenerated by workflow run: https://github.com/fin-platform/fin-agentic/actions/runs/28621678558
- Payload generated at:
2026-07-02T21:58:11.253Z
Gap note: what we don’t know yet
- Editorial provenance is single-pass harvest: the lede and highlights were harvested verbatim (typography-normalized only) from the editorial: true weekly MD digest at docs/digests/github-activity/weekly/2026-W26.md, regenerated by workflow run https://github.com/fin-platform/fin-agentic/actions/runs/28621678558. The multi-pass editorial pipeline (N-10) has not run over this dispatch; the headline and dek are an operator-supplied condensation of that harvested lede with no new facts or numbers.
- Editorial judge scores pending: no scorecard run (N-13) has graded this dispatch; control separation has no measurement yet.
- Serving-surface rendering unproven: this artifact has not yet been captured from the intranet renderer readers use (open finding HTML-SERVING-PATH-UNPROVEN); pixel and click proof on that surface is coordinator scope.
- Owner acceptance pending: this dispatch is the Exemplar #1 candidate (status pending-owner-acceptance) until the N-07 owner gate accepts it.
- Issues closed in this window: 0. The period ran PR-only; there is no issue-tracking signal in the wire data for this week.
- Capabilities-affected pillar mapping is empty for this window: the checkout brief reader has no FROM snapshot to diff tracker rows against, so row deltas and their pillar mapping honestly report no measurement yet (brief-repo-reader.ts contract).
- No outlook was derivable verbatim from the harvested copy, so section 6 renders the honest placeholder instead of authored prose.
- The computed read-time counts every visible word, including the collapsed evidence appendix and the full take-action list, so it overstates the prose reading load; the anatomy’s length targets remain design intent to be judged against this exemplar at the owner gate.