Process audit & findings · plain-language guide
A certificate grades one closed trade against the exit you named at entry. It cannot tell you whether you have a habit. The process audit is the roll-up that can — and when a pattern in it holds, the app opens a finding: a written note that waits for your verdict and never closes itself.
01 · The idea
A pharma lab does not stop at the batch record. It also runs a periodic review across every batch of a product — pass-rates by test, trends, anything drifting. The batch record is a released, signed document about one lot. The trend review is a live document about the process that makes the lots. They answer different questions and they are never the same page.
Same split here, and it is enforced in the data, not just in the layout:
Certificate
One closed trade against the spec you declared for it. Graded, then signed out by you. After the signature it is a record — corrections mint a new version, they never edit the old one.
Process audit
Every graded certificate, rolled up. No verdict stamp, no signature, no certificate status. It is recomputed from scratch every time you open the tab, and it changes as the book changes.
Finding
A stored note the app opens on its own when the roll-up crosses a stated line. It has a status, a lifecycle, and an owner — you. It is working paper, not a released record.
That last distinction is deliberate and written into the design: findings are notebook work, not release work — one app, two strata, kept apart by where each row comes from. A certificate is what you certified. A finding is what the system noticed. Mixing them would let an observation wear the authority of a signature.
02 · The loop
The audit reads certificates. It never writes to them. When a stated line is crossed it opens a finding — and even then, the finding cannot change your rules. Only you edit a rulebook. The app records that you did, and why.
03 · The audit
Every number on the tab depends on which trades are counted, so this is the first thing to read. The corpus is closed trades that carry an annotation, oldest to newest. From there, three groups are removed for three different reasons — and they are kept apart on purpose, because pooling them would report a decision you made as an accident of the data.
04 · The audit
The certificate runs a battery of tests on each trade. The audit runs the same battery across the whole book and reports, for each test, the share that came back on plan. This is the spine of the tab.
| # | Criterion | What a low rate means |
|---|---|---|
| 1 | Named an exit at entry | You are entering without a written condition to leave on. Nothing downstream can be graded properly. |
| 2 | Exit pre-committed, before the outcome | The plan is being written after you know how it went. The words may be true; they are not evidence. |
| 3 | Exited at or by the named condition | You name levels and then trade around them. |
| 4 | Override was pre-named, not rationalized | When you leave the plan, the reason is being invented at the exit. |
| 5 | Exit reason recorded | Capture is leaking. Nothing to read later, on any of the rows above. |
One line leads the whole tab: the criterion with the lowest pass-rate — the single process habit to fix. Two guards sit on it, and both exist because a headline that lies is worse than no headline:
Each criterion also carries a trend arrow: the last ten certificates against everything before them. Direction, not magnitude — at this sample size the arrow is the only honest part of the comparison.
05 · The audit
Separate from any grade, the audit asks a data-integrity question: was the plan written down before the outcome was known? Three buckets, and the honest one is the biggest.
An earlier version treated any annotation created on the entry date as witnessed. That counted machine-created shells as pre-commitments and reported roughly twelve where the defensible number was somewhere between one and twelve. The clock, not the author, is what makes a record trustworthy.
06 · The audit
Process grading invites one fair objection: fine, but does following my rules make money? The audit answers it out loud rather than assuming the answer is yes.
Discipline dividend
Average realized P&L of the trades that finished on plan (including flagged) minus those that finished off plan. Only finalized verdicts take a side — anything still awaiting review is pending evidence, not a side.
Honor dividend
The counterfactual: where the price went after each exit, split into three arms — exits that honored the plan, exits that were rationalized, exits that were operational. Two horizons per exit (the declared hold, and the follow-through), aggregated separately.
Per-rulebook dividend
Once you run more than one rulebook, the pooled figure can no longer say which spec is mispriced. This slices the same comparison by the rulebook each certificate was graded against, so the arrow points at a named spec.
Outcomes
Profit factor, conviction, alert win-rates. Collapsed by default, on purpose — they flip on a handful of trades. A sanity check, not a lesson.
07 · The audit
Every time you change your own call on a certificate, the audit counts it. Once per certificate, not once per test — revising both judgment tests on one trade is one changed mind.
This is not a mistake tally. Re-judging after the system surfaces a fact you did not have is the process working exactly as intended. What matters is direction: revisions that move a verdict up, and revisions that move it down. The audit also separates two kinds — a call changed after new evidence appeared, and a call that overruled the machine's proposal outright. A drift downward is the number worth acting on.
08 · The audit
Read back over everything above and one thing is missing. Every rate on this page is an exit rate. Named the exit, exited where you said, deviated on purpose, wrote it down in time — all of it is about the way out. The audit has never had a number about the way in, because the certificates it reads have never carried one.
That is being fixed at the source rather than here. From 2026-08-31 the app files, on every trade it captures, what the market actually was on the day of the entry: price against the 200-day, RSI, where in the 52-week range the fill landed, volatility, distance from the high. Pulled from the tape for that date. Nothing to type.
Why it wasn't there already, which is a finding about the instrument
The app has been able to fetch that reading for months. It fetched it when you expanded a trade row on the Performance tab, and at no other time. Across 1,142 annotations it had run seven times. No failure, no outage, nothing to repair — the measurement was wired to a click, so what accumulated was a record of curiosity rather than a record of the book.
This is the third instance of the same shape in three weeks. The counterfactual that says what an exit dodged computed only when somebody opened a certificate, and 47 of 64 signed certificates had their window still in the future at the moment of signature. The alert rollup fired all summer and was never once read. A measurement wired to an interaction is not a measurement. All three had the same cheap cure: move the trigger off the gesture and onto the calendar.
Two questions, and they are different from each other. The first is per trade and belongs on the certificate: did this entry meet the clauses of the entry rulebook it was graded against. The second is the one only a roll-up can ask — where in the range do you actually buy, across everything? A rule that says bottom third and a book that clusters in the top half is not a bad trade anywhere; it is a habit, and habits are what this page exists to find.
The entry clauses will also be diagnosed the way the exit clauses already are. A clause that has never once returned a readable answer across enough trades is a problem with the wording, not with the trading, and it gets filed as a finding and left open until you reword it or accept that it will never be graded. That is the same treatment the semiconductor rule got, and it is the honest one.
An entry can only be measured against a rule that existed before the entry. So the population is trades entered after you declared an entry rulebook, carrying a filed reading — and on the day it ships that population is zero. It fills at the rate you trade, not faster.
That is the price of refusing to backfill, and it was refused on purpose. The readings for older trades could be bought — the tape is still there — but grading a 2026 entry against a rule declared in 2027 is a retrofit, and this system already refuses that on the exit side. A number that cannot be graded is not worth the fetch.
Where this stands today
Collection is live. No entry rates appear on this tab yet, and none will until there are enough graded trades behind a declared entry rulebook for a rate to mean anything. Until then the honest answer is the empty one — which is the same rule every other gate on this page follows.
09 · Findings
Findings are not generated on a schedule and there is no inbox to keep up with. Each one fires from a stated condition, and each condition was chosen because it means something specific has gone wrong with the process, not with a trade.
| Finding | Fires when | What it means |
|---|---|---|
| Calibration disagreement from the spot check |
you graded blind and your category differs from the lab's on the same test | The instrument and the operator disagree about the same specimen. One of you is misreading the exit — worth finding out which. |
| Discipline dividend negative from the strategy review |
pooled per-trade delta < 0 | Your rule-following trades made less per trade than your rule-breaking ones. Either the rules need work or the sample is noise — your call. |
| Rulebook dividend negative from the strategy review |
a single rulebook's delta < 0 with both sides populated, and 2+ rulebooks exist | Within that spec, following it underperformed breaking it. The spec may be mispriced for current conditions. |
| Weakest link low from the strategy review |
weakest criterion below 40% across at least 3 measured trades | One test in the battery is failing far more than the rest. That test is where the process leaks. |
| Rule that can’t be checked from the strategy review |
one rule in a book came back unknown on 5+ graded trades — never resolved, never sat out | The rule is not being broken. It cannot be read at all — nothing in the record ever says whether you followed it. Rewrite it into something checkable, or take it out of the book. |
| Missing exit doctrine from the strategy review |
your default rulebook has no written exit rule, and 3+ certificates are graded | Exits are being judged with no spec. The most fixable finding on the list. |
10 · Findings
When you grade a certificate without seeing the machine's proposed verdict, that grade is marked blind at the moment it is written. Blind grades are the only ones that can be compared: once you have seen the proposal, your answer is anchored to it and the comparison measures nothing.
Blind grades are paired strictly within one test — your call on a test against the lab's call on that same test. Where one answer is copied across two tests, the copy raises nothing: a finding filed from it would blame you for a judgment you never made, and would double any real disagreement, since the original raises its own.
When you and the lab disagree, you win — your newest call outranks any machine verdict, on the certificate and in every count. The finding is not an appeal. It is the note that the two of you diverged, kept so the divergence can be read as a pattern later.
11 · Findings
A finding has one job: to sit there until you decide something. It never expires, never auto-closes, and closing it is a recorded act with a stated reason.
| Choice | What it does |
|---|---|
| Reviewed — keeping my rules | Records that you examined it and decided no change was needed. Your rules are untouched. This is a legitimate answer, not a snooze. |
| Changed my rules because of this | Records that a rule change followed. It does not edit the rulebook for you — you make that edit yourself on the Strategy tab. This is the paper trail for why. |
| Not a real issue | Marks it a false alarm, so later reviews know this flag was noise. |
| Replaced by a newer finding | Closes it in favour of a more recent finding on the same pattern. |
The one thing the app refuses to do
A finding can tell you a rule is underperforming. It will never rewrite the rule. The spec is yours — the same reason the certificate never hands you a suggested stop. An app that both writes the spec and grades against it is grading itself.
12 · Vocabulary
Worth pinning down, because the same word landed on three unrelated mechanisms.
| Where you see it | What it is |
|---|---|
| ON PLAN · FLAGGED | A certificate verdict. Nothing was rationalized, but something was operational — worth noting, not held against you. |
| A flag on a note | A journal entry you marked as a risk note. Yours, written by hand, unrelated to grading. |
| "Flags" in the tab | Findings, under their other name: the plain-finance vocabulary calls a finding a flag. Same rows, same lifecycle, different label. |
Reading it honestly
Every number on this tab moves week to week, and most of them sit on samples too thin to be precise — that is why the gates render provisional rather than hiding the figure. Read rates as direction, read dividends as sign, and treat a single finding as a prompt to look, never as a conclusion. The one thing here that is not provisional is the record: what you graded, when you graded it, and what you decided to do about it.