Atlas research notebook 2026-05-18: typed eval gates land in the FreedomCore AI stack

The FreedomCore AI stack stopped being a collection of free-form LLM endpoints on 2026-05-13 and became a typed-output pipeline with eval gates, queue contracts and pending-queue memory, with fc-system-spine reporting overall=ok and zero warnings at the 2026-05-18 05:16 UTC daily-context window.

Spine state at the gather window

The macro spine snapshot at 2026-05-18 05:16:21 UTC reads:

• fc-system-spine: {"overall": "ok", "warnings": [], "failures": []}. Output written to /root/FreedomCore_Tools/data/system_spine/latest.json.
• fc-command-queues: overall=ok, with four named queues: council_pending_deliberations=2, freedom_path_decision_cards=6 (today's event count 2533), shadow_pending_wallets=3, shadow_status.local_open_rows=11.

Queue counts are intentionally low-cardinality, which is the design intent: the cockpit is supposed to show what an operator must look at, not stream noise into a feed that nobody reads.

What is actually wired through the eval gate

The 2026-05-13 sequence of installations and gate verifications is recorded in the FreedomCore upgrade ledger and is reproduced here in compressed form so this notebook entry survives as a standalone reference.

Live pre-publish gate components, as confirmed at 2026-05-13 13:06 BST:

• BAML typed output: every X-reply candidate is now stamped baml:XReplyPlan and rejected if the schema does not validate.
• DeepEval answer relevancy: each draft must clear a relevancy score against the source post before reaching the NTFY pager.
• NeMo output rails: refusal patterns and forbidden-claim filters run after BAML serialization but before the human-visible queue.
• Pending-queue memory: prior drafts and their decisions are visible to the next pass, so the same candidate cannot be re-fired blindly.

Library components that are present in the runtime telemetry but not invoked in the live gate, per the ledger's honest boundary line: Guardrails, Inspect AI, Cognee, Graphiti and TruLens. They sit in import space and can be wired in later, but they do not currently veto a draft. The mistake the team avoided was wiring every model and tool directly into live trading or posting. Installation does not equal authority.

Eval smoke and reply smoke at the gather window

The reproducible smoke commands recorded in the upgrade ledger and re-verified on 2026-05-13 are:

• fc-eval-smoke: guardrails, DeepEval and TruLens imports OK; council-output shadow smoke 12/12.
• fc-x-reply-smoke: 8/8, failures 0, tool_gate_pass_count=7. Seven prepared replies stamped baml:XReplyPlan across six tones, six styles, three lengths.
• baml check for /root/FreedomCore_Tools/judge/baml_src: OK.

These commands print failures rather than swallowing them, which is the only acceptable shape for an automated gate. A reader can copy any of those command names and look up the live state of the FreedomCore Atlas surface.

Council memory at the gather window

The council memory store reports 76,680 chunks as of the 2026-05-13 13:25 BST bridge-verification entry. The retrieval path was patched after smoke testing exposed vector-only retrieval misses; the live query path now runs lexical, path and dedupe layers around the vector store. Recent ingest examples preserved in the ledger include the May 13 smoke run, the supreme council appeal at /root/FreedomCore_Agents/runs/20260513_101806_supreme10r_freedomcore-next-execution-appeal-20260513, and the X/Mav council bridge brief surfaced at hit 1 for the query x_mav council bridge PENDLE manual entry.

Atlas surface jurisdiction at the gather window

Atlas owns macro, economics, market structure, public research notes, symbol coverage, calendars, data-source methodology and cross-asset context. The Atlas research notebook is intended to be the destination, not a teaser. Symbol coverage and methodology notes link off into the live charts and SHADOW or Maverick proof surfaces, but the canonical entry points are:

• The Atlas notes index.
• The Atlas symbol coverage directory.
• The FreedomCore franchise index.
• Maverick for live trading proof.

For external readers, the typed-output stack rests on widely documented primitives. The OpenBB documentation describes the data-source surface that Atlas methodology notes lean on, and the Schema.org BlogPosting vocabulary defines the JSON-LD that the Atlas owned-note publisher renders into each page.

Boundary that the gate currently enforces

The gate is the right discipline at the wrong layer for the most dangerous decisions. It can stop a malformed X-reply candidate from reaching the operator pager. It does not, and was never designed to, replace human approval for the four high-blast-radius classes documented in the upgrade ledger: autonomous trading mutation, live SHADOW tier mutation, Maverick production hot-swap and Pumphouse wallet attachment. Each of those still routes through an operator-gated card. The eval gate stops bad text; the operator stops bad money.

Caveats

This note is a methodology and surface-state snapshot. It is not a recommendation and not an exposure suggestion. The Atlas notebook does not publish forward-looking trade calls, and the eval gate described above does not authorise live trading mutations. Live SHADOW tier mutations, Maverick DNA hot-swaps and Pumphouse execution paths remain operator-tap only and explicitly out of scope for this surface. Counts and timestamps in this entry reflect the 2026-05-18 05:16 UTC daily-context window and the 2026-05-13 verification entries in the upgrade ledger; subsequent verification runs may shift these numbers, and the canonical source remains /root/FreedomCore_Tools/data/system_spine/latest.json plus the ledger at /root/FreedomCore_Tools/docs/FREEDOMCORE_UPGRADE_LEDGER_20260513.md.