Atlas research notebook 2026-05-19: bounded council workdirs and macro spine state

The FreedomCore AI trading research dashboard reached the 2026-05-19 05:17 UTC daily-context gather with fc-system-spine reporting overall=ok, zero warnings, zero failures, and a council runner that no longer snapshots the entire VPS through stray git add --all when a six-seat scout deliberation is fired.

Spine state at the gather window

The macro spine snapshot at 2026-05-19 05:17:30 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=10 against a today_event_count=2528, shadow_pending_wallets=3, shadow_status.local_open_rows=13.

Queue cardinality has held inside a single order of magnitude since 2026-05-13. That stability is the design intent of the cockpit: a low-noise, high-signal surface that exists to be looked at, not to be streamed through.

Bounded scout workdirs in the council runner

The macro market structure research surface depends on a council layer that can be fired without re-indexing every file under /root. Until 2026-05-19 the opencode-nvidia scout invocation was resolving its working directory back to /root and snapshotting the whole VPS through git add --all, which made every six-seat call hang silently until the wall-clock timeout expired.

The repair was logged into the FreedomCore upgrade ledger on 2026-05-19 and is auditable at the file level:

• Patched /root/FreedomCore_Agents/council_runner/run_council.py to forward --dir <task_workdir> into the underlying runner.
• Patched /root/FreedomCore_Agents/config/tasks.yaml so every task name carries an explicit work directory.
• Live backups taken into /root/backups/fc_council_runner_20260519_001/ before the live edit.

Verification of the patch:

• Direct opencode-nvidia smoke returned OK.
• Six-seat council smoke under /root/FreedomCore_Agents/runs/20260519_022508_custom returned OK for qwen, mistral, nemotron and gpt-oss seats.
• Current-state Stage03 review run /root/FreedomCore_Agents/runs/20260519_023542_shadow-stage03-current-upgrade-review-20260519 returned useful PASS_CONTINUE outputs from mistral_small_4, nemotron_nano_30b and gpt_oss_20b; qwen_coder, qwen_next_instruct and mistral_medium_35 returned empty outputs despite exit code 0.

The honest framing of that last verification matters. Three seats returning empty bytes on a 0 exit code is a partial scout pass, not a six-seat approval. Future deliberations that need full coverage will need to wait until those three providers return non-empty payloads or be re-fired through alternative seats.

Source-coverage discipline

The Atlas symbol coverage surface keeps three behaviours that justify the research notebook framing rather than a market-feed framing:

1. Symbol universes are written into cache/market_universe/ from public source files at every refresh. Today the gather sequence touched otherlisted.txt at 2026-05-19 03:43:03 UTC and nasdaqlisted.txt at 2026-05-19 03:43:03 UTC. These files are the inputs, not the outputs.

2. Free-source filings are cached under Atlas_Backend/cache/free_sources/, with sec_submissions_0001018724.json last refreshed on 2026-05-18 12:51:13 UTC and sec_companyfatcs_0001018724.json on 2026-05-18 11:04:54 UTC. The notebook does not re-publish the body of those filings; it publishes the methodology around them.

3. Dated upgrade-ledger entries record what was changed, which file was patched, which command verified the change, and where the live backup lives. The ledger is the long-term memory of the cross asset market intelligence layer.

The OpenBB DefiLlama RSS market spine continues to feed the daily context. See the OpenBB docs for the underlying provider model, and review the Schema.org BlogPosting shape that every Atlas note serializes against in the published HTML.

Council memory at the gather window

The public council index now carries 34 deliberations. The newest entries at the gather window:

• Pumphouse Cohort Sniper Paper V2, 15-round Supreme Court (20260518_225808_supreme10r_pumphouse-cohort-sniper-15r-20260518). Verdict landed at 2026-05-19 00:00 BST.
• 20260518_194559_custom, a smaller upstream deliberation.
• FCORE Tokenization How-Charter (20260515_130359_supreme10r_fcore-tokenization-how-charter-20260515).
• franchise-seo-keyword-strategy-20260513.
• world-class-wallet-profiler-20260513.

The council index is published as static JSON at /api_data/swarm_council_index.json and is consumable by the public Swarm cockpit. The macro research notebook treats it as an immutable audit log: previous verdicts are kept in place, new verdicts append. No verdict is rewritten after the fact.

Daily Atlas pack contracts

The 2026-05-19 daily-context manifest is a small contract that the rest of the research surface depends on:

• surfaces: four named surfaces (Atlas, Maverick, Pumphouse, SHADOW), each with a context pack and a writer prompt.
• publish_policy: draft_to_command_queue_only. There is no auto-publish path inside the pack itself.
• next_step: drafts written by an LLM are queued through owned_note_publisher.py queue, which writes into /var/lib/freedomcore-command/pending_notes.json for operator review. No draft becomes a public note until the operator approves it from the Command queue.

That contract is the load-bearing reason these notes can stay short and specific. The note does not have to wedge itself into a marketing arc, because the marketing arc is built by the Command operator on top of an evidence-led research feed.

Pending macro work

Two items remain in the macro queue for the next 72 hours:

• Six-seat council coverage on the three currently-empty NVIDIA/Qwen seats. Until those return non-empty payloads, deliberations that require unanimous six-seat sign-off must be deferred or re-run through the Stage A scouts plus three deliberating seats.
• Eval-gate fan-out into the cross-surface notes pipeline. Today's pipeline assumes typed BAML output and DeepEval relevancy at the X-reply layer; the parallel hook for owned-note drafts is wired but not yet hard-gated on the same DeepEval threshold.

Caveats

• Macro spine overall=ok records the absence of breaking failures, not the presence of new market signal. Spine is a heartbeat, not an insight.
• The council partial-coverage line above is a real caveat. Three empty seats on a six-seat run is not a quorum.
• Source-cache files cited above are public methodology endpoints. The notebook does not redistribute the body of any free-source filing.
• All Maverick, SHADOW and Pumphouse surface decisions sit outside this notebook's jurisdiction. See Maverick for trading research and the Atlas notes index for the full research surface.
• The notebook does not predict prices, promise returns, or recommend trades. It records what the macro spine shows, what the council voted on, and what was patched into the runner layer with an audit trail.

Cross-references for the wider stack live at FreedomCore and the public Atlas symbol coverage page.