The probability bridge: manual hit rates in, automated EV out.

The design decision

Why the scorecard takes probabilities, not trades.

The scorecard could have been built to ingest raw trade rows and compute everything — that's what CP3's rollup engine does. Instead it takes one deliberate step of manual mediation: the operator counts the week's branch outcomes from the journal record, computes the observed probabilities — P(1R) per branch, the conditional runner rates, the Overflow conversions — and enters them as decimals. The mediation is a feature with two payoffs. It keeps the scorecard's role clean: a weekly EV cockpit, not a second journal, with raw trades explicitly belonging to TradeZella and the CP3 Journal. And it forces the operator through the evidence weekly — the person entering 0.58 into the Normal P(1R) cell has personally counted the week's Normal trades, which is a form of engagement no automated pipeline provides.

The bridge's spans

From counted inputs to the week's verdict, in one traceable line.

The bridge runs in fixed spans. Inputs: the branch probability cells and the risk/context inputs, entered as decimals with notes alongside. First computation: per-branch EV — Normal from its 1R and 2R chain, Trend Partial and No-Partial from their 1.6R structures, Overflow from its 0.75R and 1R conversions — each branch's payoff structure encoded in its formula, so the probability meets the exit architecture it actually traded. Second computation: the two blends — Trend Blend combining the trend family, then Overall Blended EV under the live profile weights. Verdict: the blended number against the thresholds, tagging the week. Every span is inspectable, which means any strange verdict can be walked backward — status to blend, blend to branch, branch to probability, probability to the journal count it came from.

The failure modes

A manual bridge fails in known, checkable ways.

The troubleshooting table reads like a field guide to this design's specific hazards, and every entry traces to the same root: the bridge trusts its inputs. EV cells blank means a required probability cell is empty. EV impossibly large means someone entered 60 instead of 0.60 — the decimals rule is the bridge's one strict convention, and violating it inflates every downstream number a hundredfold. Weight sums off 100% mean the legend's profile weights were edited incorrectly, which silently corrupts every blend until caught. And the subtlest one — schema drift, where a probability lands in the wrong branch's cell — produces plausible-looking nonsense that only the walk-backward audit catches. None of these are exotic. All of them are five-minute checks, which is exactly what a well-designed manual system's failure modes should be.

• Decimals, always: 0.60, never 60. The single most consequential convention in the workbook.
• Probability cells that are formula-driven from the summary stay formula-driven — typing over the bridge is how schema drift starts.
• A weird verdict gets audited backward before it gets interpreted: status → blend → branch → probability → journal count.

What the mediation buys

The weekly count is a discipline disguised as data entry.

The unautomated step earns its keep beyond data integrity. Counting the week's outcomes by branch is a forced confrontation with the actual evidence: the operator who tallies four Normal trades knows the week's Normal P(1R) rests on four trades, and reads the resulting EV with exactly that much confidence. The same operator receiving a fully automated verdict knows only the color. The count also surfaces classification honesty at the moment it can still be fixed — a trade whose branch label looks wrong during Saturday's tally gets corrected in the journal before it corrupts a probability, rather than discovered months later in an attribution anomaly. The bridge is slower than a pipeline by perhaps ten minutes a week. What it buys is an operator who has personally touched every number the verdict stands on.

The key idea

Narrow, manual, and locked is a deliberate architecture, not a limitation.

The scorecard's design refuses two tempting expansions: it won't ingest raw trades (that's the journal's job) and it won't let its formulas be adjusted casually (that's what the R&D lab is for). What's left is a bridge with one job — observed probabilities become profile-weighted expectancy becomes a comparable weekly verdict — done identically every week, auditable end to end, failing only in ways a five-minute check catches. In a stack full of sophisticated engines, the scorecard's simplicity is its contribution: the one place where the week's evidence and the operator's hands meet the system's arithmetic, with nothing hidden in between.