Zero Theta Labs Autonomous research under frozen evaluation
Abstract
In plain terms: overnight, the system proposes changes to itself; independent scoring challenges them; a person decides what ships. Labs is the research arm of Zero Theta, and its central instrument is the autonomous research campaign: software agents explore, evaluators frozen before the campaign score every candidate, and a person decides what ships. Every result, kept or rejected, is appended to a durable ledger.
Keywords — autonomous research; frozen evaluation; append-only ledger; human-gated promotion; market regimes.
1The campaign
A campaign fixes the code agents may touch, the evaluators that will score them, and the ledger that will remember them — then runs in rounds.
Algorithm 1 Autonomous research round
- spawn a disposable worktree w
- apply an allowlisted patch p proposed by the agent
- evaluate w ⊕ p under the frozen evaluation profile
- append candidate, score, and artifacts to the durable ledger // unconditionally
- await human gate; promote only on explicit approval
2Why each line
Line 1 (spawn — isolation). Each candidate runs in a disposable worktree, isolated from the mainline by construction.
Line 2 (apply — bounded scope). The allowlist is written before the campaign opens — small diffs are reviewable diffs.
Line 3 (evaluate — no self-grading). The evaluator is frozen before the search begins; the thing being optimized never grades itself.1
Line 4 (append — failures kept). Failures are kept — a search you cannot audit is a search you cannot trust.
Line 5 (await — human decision). The machine ranks; a person promotes.
3Beyond the campaigns
Alongside the campaigns, Labs studies market regimes and structure — the conditions under which strategies should, and should not, be trusted.
Footnotes
- Changing the evaluator means closing the campaign and opening a new one, with a new ledger. ↩