Skip to content

Summary

TOMIC's learning loop was a journal, a coach, and review. The 2026 framework mechanizes it: every day writes machine-readable records, every research item resolves, every decision cites precedent, and the framework's rules — the regime→strategy mapping and the risk posture tables — are amended only through a quorum process on accumulated evidence. Learning is what turns 200 days of operation into an asset instead of 200 days of opinion.

The Accumulators (96-journal)

Three append-only, machine-readable logs (JSONL, one record per day/event; schema frozen at v0 — see 96-journal):

Log Record Purpose
Regime journal date, state variables, regime label, transition flag, data-missing flag Builds the ex-ante labeled regime series that will one day validate (or kill) the mapping
Decision log date, candidates, gates passed/failed, final proposals, human verdicts The audit trail of the decision function
Outcome log date, playbook id, realized result vs the attested EV, slippage, assignment notes Closes the loop between forecast and reality

Markdown narratives accompany the JSONL for human reading; the JSONL is what processes consume.

Forced Resolution

Research only counts when it resolves. Every inbox item acquires a resolution: link to a decision-log entry — including explicit "no action, because X" outcomes. An item without a resolution after its stated urgency window is flagged stale and surfaces in the daily review until resolved. This is the anti-hoarding rule.

Consistency by Precedent

Day-to-day LLM drift is the framework's quiet failure mode. Mitigation is procedural: before the intelligence function classifies an item or the decision function gates a candidate, it must cite how similar prior items/candidates resolved (a journal lookup). Divergence from precedent is allowed but must be stated as such — consistency becomes checkable rather than assumed.

Rule Amendment — The Concordance Quorum

The mapping and posture tables are versioned artifacts (v0, v1, …). An amendment requires:

  1. Concordance quorum: ≥3 journal-cited, concordant events (same regime context, same predicted direction, same realized direction) — or one named historical case-study analogue with explicit reasoning (the analogue path supports the proposal; activation still requires any stated calibration and human approval);
  2. A written amendment note: what changed, the evidence, and the predicted consequence of the change;
  3. Effective date: the new version applies from the next session onward, never retroactively.

This is the framework's defense against noise-chasing — the learning-side analogue of backtest discipline for live rules.

What Learning Is Not

  • Not per-day parameter tweaking (that is overfitting to noise).
  • Not silent drift: if behavior diverges from the charter, the divergence is documented in the journal, not absorbed into practice.
  • Not self-approval: rule amendments are proposals to the human operator, like trades.

Links

References