Overview · the whole book, unfiltered
Where the money went
Price variance and margin impact are the same number. Standard cost
cancels out of the subtraction, so every dollar of price variance is a dollar of
margin, one for one — there is no volume offset and no cost recovery. The identity
is reconciled on every matched line by validate.py check 5, and derived the
long way in conform.py so it is verified rather than assumed.
Relationship: change over time
Relationship: magnitude
Now filter it
Calibrate before you alert
Design point five: run the exception report visibly for a few cycles, watch where the count falls as the threshold rises, and find the line that surfaces real money without burying the team. Alerting on a rule nobody trusts yet only teaches people to dismiss it. Move the threshold and watch what survives.
Relationship: ranking
Relationship: ranking
Relationship: distribution
Relationship: ranking
The exception report
Ranked by margin given up, descending. This is the artifact a deal desk would actually work from on a Monday morning. Approved exceptions are excluded from the totals above and shown here separately — some were priced above the agreement, and folding them into leakage would both overstate it and destroy the calibration story.
Data & method
The matching rule
Every quote line is joined to the agreement register on customer and product, keeping only agreements whose effective window contains the quote date. Where more than one survives, three tiebreaks apply in order: an agreement that had already expired or been superseded as of the quote date is excluded outright; a part-scope agreement beats a family-scope one covering the same part; and where two equally specific agreements both cover the date, the one with the later effective start wins. If a tie survives all three, the run fails rather than guessing — that is a data-quality defect, not a judgment call.
A line counts as priced to agreement when the quoted price is within ±0.5% of the agreed price. Real quotes carry rounding, and a tolerance band is the honest way to handle it. The threshold control above operates on materiality in dollars, not on this tolerance.
What an off-agreement line actually is
Not random noise around the agreed price. The failure being modelled is a rep quoting the customer's previous price instead of the newly negotiated one — so the variance is the renegotiated premium, which is why every unapproved off-agreement line in this dataset is negative and why the variance is larger where the premium was larger. Off-agreement lines cluster on specific reps, on recently renegotiated products, and in the weeks just after a renewal lands, when the pricing sheet has not caught up.
Why "no governing agreement" is a category and not a gap
261 lines have no agreement covering them, and dropping them would hide the risk rather than report it. Three distinct causes, all present: the customer never signed for that product (99), the product launched after the agreement was written (95), or the agreement lapsed and was never renewed (67). Note that a lapsed agreement is not the same event as a lost quote — the first means coverage ran out, the second means the line never converted.
Realized, open, and closed lost
Every line carries a funnel position, but exposure dollars exist only on off-agreement lines. Realized means the line booked at the wrong price and the money is gone. Open means it is quoted and not yet booked, so it is still fixable before it reaches the customer — which is precisely what a deal desk exists to do. Closed lost means the quote never converted, so there is no exposure, but it stays in the coverage audit rather than being quietly dropped.
Open exposure is small, and that is honest. A twenty-four month retrospective only has a few months of undecided funnel in it. The number is persuasive because it is actionable today, not because it is large.
A register-hygiene defect worth naming
The register contains two kinds of overlapping coverage. Most are linked renewals, where the predecessor carries a pointer to its successor. Nine are unlinked duplicate registrations — the same customer and part registered twice with no supersession link between them. The recency tiebreak resolves them deterministically, but the fact that it has to is itself a finding about the register, and a real deal desk would fix the entry process rather than lean on the tiebreak.
Honest limits
- The data is synthetic. It demonstrates a design; it measures nothing real.
- The generator imports the matcher, which guarantees the pipeline is consistent but is not independent evidence that the rule is correct. Correctness rests on the worked examples and the identity checks, which do not depend on the matcher.
- Family-scope agreements carry a single unit price. That works here only because part prices within a family are generated within ±12% of a family base; a real register would more often hold a discount schedule, which needs a different scope model.
- Register
statusdescribes the register as of 2026-08-01. The matcher judged status as of each quote's own date, so an agreement labelled superseded today may legitimately have governed a 2024 quote.
Full rule, worked examples and tie handling:
governance/matching_rules.md. Generator assumptions and realized mix:
governance/generator_assumptions.md. All twelve validation checks:
governance/validation_report.md. Chart review:
governance/chart-review.md.