Skip to content

Futures — Things We Considered But Haven't Built

This section is for ideas that came up during the project, are worth preserving, and would be valuable to someone — possibly future-Ken, possibly someone else picking up gemma-forge — if and when there's time, motivation, or external pressure to pursue them.

These are not journey entries. They didn't happen. They are not improvements either, because they aren't fixes for things that broke. They are speculative engineering notes: enough detail that a competent engineer could pick one up and start work, but no implementation, no commitments, no schedule.

If a futures entry ever does get implemented, it should be promoted to a journey entry that documents the actual build. The original futures entry stays as the record of when the idea first surfaced and what the thinking was at that point.

Entries

  • The Harness Is a Training Data Factory — the open-weights case study nobody is running yet. The accumulated structured (context, action, outcome) data from agentic harness runs is, by construction, a fine-tuning corpus. With open weights you can close the loop locally; with closed weights you cannot.
  • Detection Engineering as a Third Skill — proposed weekend-project skill for tuning Sigma rules against labeled threat-telemetry corpora (DARPA OpTC, BOTS, MITRE Evaluations). Surfaced while looking for additional use cases that exercise different problem shapes than the two existing skills, captured with honest unknowns. Statistically-graded evaluator instead of binary pass/fail; corpus is per-run config so the tip pool accumulates corpus-independent Sigma authoring patterns. Tests the architectural claim that "skills are pluggable" against a structurally different problem shape than STIG/CVE.
  • Detection-Tuning Skill — Pre-Flight Checklist + Build Sequence — the operational companion to the proposal above. Three pre-flight investigations (~2 hours total) with concrete green/yellow/red signals, followed by an hour-by-hour build sequence (Friday night minimum slice → Saturday real skill → demo arc). Reads as the execution plan that a fresh Claude session can pick up cold.
  • Gemma 4 12B vs 31B — Capability-vs-System Experiment — operational companion to ADR-0024. Reframed from an L4 throughput bake-off into a portable test of the project's core thesis: can the system (the loop + accumulated tips/learnings) lift a smaller 12B to the capability of a bigger bare 31B — especially when the 12B inherits learnings the 31B built? Both models official bf16; capability (STIG rules fixed, Architect quality) is the metric, throughput the kicker. Pre-registered conditions (31B cold, 12B cold, 12B warm←31B-tips) with a falsifiable gate; the cross-model transfer is the headline. Hardware-agnostic on purpose — the result survives the next box.