The audit loop: how coefficient choices get graded and the ladder stays true.

The log

What gets written down: the decision, its source, and its context.

The Coefficient Audit Log is the one VIP surface built for daily writes: record real trade and cycle coefficient decisions and outcomes. Each entry captures the final coefficient used and — critically — its source: static, VIE, or override, alongside the context the comparison was made in. The source field is what elevates the log from bookkeeping to calibration data, because it partitions every outcome by which instrument made the call. Months of entries answer questions no single decision can: does the VIE model out-select the static profile for this symbol? Do the operator's overrides beat the model they overrode? Is one rung of the ladder persistently associated with a failure pattern? None of those are answerable from memory, and all of them are answerable from a disciplined log.

The grading metrics

Four numbers grade a coefficient: outcome R, MAE, MFE, and giveback.

The maintenance doctrine names the review explicitly — run periodic reviews comparing coefficient source versus outcome R, MAE, MFE, and giveback — and the four metrics each catch a distinct mispricing. Outcome R is the headline but the least diagnostic alone. MAE against the trail distance reveals the too-tight error: trades whose adverse excursion consumed the trail and died inside what should have been ordinary noise, on moves that then continued. MFE with giveback reveals the too-wide error: excursion peaks generously reached, then donated back through breathing room the conditions didn't require. A well-priced rung shows the signature of neither — trails surviving their regime's noise while giveback stays inside the structure's design tolerance. The MAE/MFE lab computes these per trade already; the audit review just slices them by coefficient and source.

Reading the review

Findings route by pattern — rung, source, or key.

The review's findings sort into three families with three destinations. Rung findings — a specific tier persistently showing the too-tight or too-wide signature across symbols — are ladder-calibration candidates, and they travel the heavyweight path: the maintenance rule forbids casual ladder changes, so the finding gets documented, the proposed adjustment versioned, and the change validated before live use. Source findings — VIE persistently out-selecting static for a symbol, or overrides persistently underperforming the model — recalibrate trust: a drifted static profile routes to the Instrument Master update rule, while weak override performance is a discretion finding the operator owes themselves honestly. Key findings — the pattern living in one profile key's trades only — usually indict the baseline, and route back through the sample-status and Distance Matrix maintenance machinery rather than the ladder at all.

• Sample patience applies with full force: rungs fire at different rates, and the upper tiers' verdicts ripen slowly.
• One brutal giveback trade grades nothing — the review reads signatures across the slice, never verdicts from a row.
• Custom authority add-ons get the same treatment: used, noted why, and reviewed against the default they replaced.

Why the loop must exist

A fixed ladder in a moving market needs a maintenance contract.

The ladder's fixedness is a feature — production rules that bend per trade aren't rules — but markets genuinely change: symbols shift character, sessions re-price, and a rung calibrated against last year's conditions can drift quietly mispriced. Without the loop, the only correction mechanisms are catastrophe or vibes, both of which arrive late and teach wrong. The loop supplies the third option the whole system is built on: evidence, accumulated at decision time, reviewed on cadence, converted into documented change through a path that can't be walked casually. It's the identical architecture as tier attribution grading the risk ladder and the refresh rule keeping the benchmark honest — a fixed structure, a live record, and a gated road between them.

The key idea

Log the decision when it's cheap; the answer arrives when it matters.

Each audit entry costs seconds at a moment when the information is free — the coefficient, its source, the context, all sitting on screen. The return compounds silently: a year of entries is a coefficient dataset no amount of retrospective effort could reconstruct, and it's the difference between 'the ladder feels about right' and 'T5 has been giving back 0.3R more than design across forty London trades, here's the versioned proposal.' The loop's whole demand is that small discipline, every trail trade. Everything it gives back arrives later, labeled, and true.