Strategy application context — when each strategy applies, and when it must not fire

[AetherAI3]

Gives every library strategy the context needed to decide when it applies and when it must not fire.

One commit, 15 files, comments only — no IR, no source, no packaging changes.

The problem

Every library strategy carried two or three lines restating its own entry condition:

// Classic RSI oversold reversal (Pine: ta.rsi(close, 14) < 30).
// Buy when the 14-period RSI dips below 30.

That has zero selection value. It says what the strategy does, never when it applies — and a consumer reading this corpus already has the condition from the compiled IR. So the header was spending tokens to repeat something the reader already knew, while the thing it actually needed to decide was absent.

The change

Each header now states:

• the regime it belongs to, and the regimes where it must not fire
• conditions that must hold before it is considered
• invalidation — what makes it wrong, stated before entry
• its shape on the chart and the timeframe it operates on
• the near-neighbour it gets confused with
• the instrument class its thresholds were calibrated on

That last field is not documentation. A consumer retargets a published rule to whatever instrument is live but carries its thresholds across unchanged, so the hazard is wrong-regime numbers transplanted between instrument classes — and this header is the only place a rejection can be expressed.

Concrete cases the headers now name:

• roc_momentum — ROC > 5 is an ordinary hour for BTC and a limit-move event for an index future.
• bb_squeeze_breakout — BB_WIDTH < 4 is an absolute percentage; on a quieter instrument it is permanently satisfied and stops discriminating at all.
• atr_volatility_halt — the worst case, and counterintuitive. Carried across unchanged this breaker does not misfire, it never fires, and the volatility brake silently disappears.
• golden_cross — marked a regime filter, not an entry. It is true for months, so firing it as a trigger fires on every bar of a bull market.
• volume_spike_confirmation is a strict subset of rsi_oversold_reversal, and williams_r_reversal is the inverted stochastic_oversold. Both pairs agreeing is one idea counted twice, not confirmation — now stated where the selection is made.

Verification

Comments cannot change compiled output, so the source/IR conformance guard was never at risk — but it was verified rather than assumed:

• 0 non-comment lines changed; strategy bodies are byte-identical
• source/IR conformance guard passes; the corpus loads through a consumer's catalog reader — 21 strategies (15 library + 6 examples), 17 directional
• Nano suite 339 passed, 2 skipped
• 15 of 15 library entries carry both REGIME: and CALIBRATED ON:
• ASCII-clean — em-dashes introduced while authoring were normalised, since the library had been pure ASCII

Cost

The rendered catalog grows to roughly 4.9k tokens for a consumer that loads the whole library each cycle. That is the price of a selection filter and is the intended trade, but it is a real per-cycle cost worth watching.

🤖 Generated with Claude Code