| name | sea |
| description | Plan abi SEA (Sparse Evidence Attention) self-learning work — evidence-augmented completion, 8-signal scorer, EMA modulator persistence, and constitution audit. Use when working on ai_learn / complete --learn / evidence recall. Routes to sea-learn-loop, abi-superpower-sea, abi-superpower-constitution, and sea-learning-controller. |
sea
Entry point for abi's SEA self-learning loop (src/features/ai/). Routes:
| You want to… | Use |
|---|
Drive abi complete --learn end-to-end | sea-learn-loop |
| Deep-dive SEA scoring / adaptive modulation | abi-superpower-sea |
| Constitution audit (observability-only) | abi-superpower-constitution |
Toggle SEA in the REPL (/learn) | sea-learning-controller |
Facts that constrain any SEA plan
- SEA = evidence-augmented self-learning completion with an 8-signal scorer +
budgeted greedy selection. It classifies 7 task types; 3 currently apply
non-baseline signal-weight deltas and 4 retain the baseline weights.
AdaptiveModulator weights (EMA, alpha=0.3, key modulator:weights)
persist in WDBX only on the --learn/SEA path. Plain complete re-runs
sentiment each turn with no EMA persistence.- Constitution audit (6 principles) is observability-only, not a gate —
sets
audit_passed / audit_vetoed / escore, std.log.warns on violation,
but complete / run still return the response. Safety+privacy hard-veto
only when either < 0.5.
- The live EMA path is
completeAdaptive / completeWithStoreAdaptive via
runLearnLoop only; there is no separate routeInputAdaptive wrapper.
Honest boundary
No accuracy / energy / learning-gain claims without a repo benchmark.
