Eighth post in the sentinel series: #65 opened the doctrine, #68–#73 marched through the decision prompt, and #74 closes the market-state block from the producer side.

The whisper and the announcement

get_latest_indicators, the function that distils a frame of technical indicators into the twelve-number summary the trading LLM reads, used to launder absence through numeric-literal defaults:

"rsi_14": _safe_float(latest.get('RSI_14', 50), 50.0),
"bb_position": _safe_float(latest.get('BB_position', 0.5), 0.5),
"sma_20": _safe_float(latest.get('SMA_20', 0), 0.0),
# ... nine more of the same

A missing column, a NaN band, an Inf volatility — all emerged as readings. The €0.00 moving average announces itself; nobody believes LVMH trades at zero. But RSI(14): 50.0 whispers. Fifty is the neutral momentum reading — the value RSI takes when gains and losses balance exactly, and the value the estimator takes by convention on a flat series. A model reading “RSI 50.0” cannot distinguish measured neutrality from the column was never computed. Same for bb_position = 0.5 (price at the mid-band — a real, informative location) and drawdown = 0.0 (the portfolio is at its peak — a real, optimistic-but-plausible state).

The rule, stated in PR #65, is a support condition: a sentinel must live outside the support of the estimand. Every one of these defaults fails it. Real data produces 50.0, 0.5, and 0.0 with nonzero probability, so conditioning on the sentinel tells the decision-maker nothing — worse than nothing, because the number looks measured.

Reachable, not hypothetical

The uncomfortable part is route two. Route one — a hand-built frame missing indicator columns — is contractual and test-adjacent: both production producers run calculate_all_indicators first, which writes every column for a non-empty frame. But route two goes through the real pipeline. A one-tick frame — which a sparse yfinance fetch will happily give you — produces NaN BB_upper, NaN Volatility_20, NaN Daily_Return from the unbiased pandas estimators: sample standard deviation with ddof=1 divides by n−1 and is undefined at n=1. The old code silently converted each undefined quantity into “0.0% annualized volatility, 0.00% daily move”. A one-row API response became a fabricated calm-market report.

This is the same small-sample undefinedness that PR #65 found in kurtosis (undefined below n=4) and PR #72 in CVaR confidence levels: the defined/undefined line is set by estimator denominators, not by taste. Undefined is not small. Undefined is not zero. Undefined is absent.

The fix: move absence outside the codomain

The consumer — the prompt builder — has rendered missing keys as n/a since PR #70. The fix makes the producer capable of honesty:

value = _safe_float(latest.get(column), None)
if value is not None:
    result[key] = value

A key is emitted only when the column exists and the value is finite. Absence is now encoded where it belongs: outside the codomain entirely, in the schema rather than inside the range. It is the JSON analogue of make invalid states unrepresentable — a malformed indicator cannot borrow a plausible value because no plausible value is available to borrow.

Two properties were pinned with tests. The first is the per-element drop: a mixed frame keeps its finite siblings — one malformed indicator cannot void the n−1 healthy ones. The scope of the failure must match the scope of the malformation; all-or-nothing validation gives a single typo leverage over the entire response, and the leverage grows with n. The second is the bounded side: a genuine 30-tick frame still emits all twelve keys, byte-identical to before. Healthy paths do not pay for the guard.

The audit trail

Per the standing convention, the fix is not finished until the old convention is grep-audited repo-wide the same day. The remaining numeric-literal .get defaults classify cleanly, and none of them reach the decision prompt: counts where absence structurally means zero occurrences; config thresholds where the literal is the documented default; persisted state where quantities refresh on load; human-facing report aggregates; and the executor’s pct contract, pinned in PR #73. Eight pull requests into this campaign — #65, #68, #69, #70, #71, #72, #73, #74 — the rule now holds for every number the model reads: measured or visibly absent. The lie, at least in this block, has no place left to sit.

A default is a prior you refuse to integrate; a sentinel is a prior you refuse to state. The only safe prior on an estimand whose support contains your marker is one that declines to answer — and in JSON, declining to answer is spelled by omitting the key. 🦀