When did our last decision stop being correct?
The platform tells you, by name, in days. Capital Refinery is private-capital software that holds itself accountable for the decisions you make on top of it. Decisions are graded against reality. Theses are aged. Dead positions can't make new commitments. The product refuses to let stale data convene IC.
Investment software that grades itself.
When you commit a decision in Capital Refinery, the system records what you predicted would happen. Then it checks. Did the action you took deliver the improvement you said it would? Did the thesis you committed survive the next two quarters of operating reality? Did the partner who made the call have predictions that tracked the outcome? Most software lets you make decisions. Capital Refinery tells you, on a fixed cadence, which of your prior decisions stopped being correct.
Five mechanisms. One closed loop.
These are not features. They are the load-bearing architecture beneath every other surface in the product — the Continuous IC Memo, the Decision Ledger, the Anchored Exports, Risk Signals, and the IRA artifact ladder all rest on this floor.
Action-Outcome Hypothesis Confirmation
When an operator resolves an action with an expected improvement, the system checks the result. A nightly sweep computes the average KPI delta over the next 45 days and flags whether the prediction was confirmed (net delta > 0.5%). The platform grades its own playbook. No one clicks anything. The record builds itself.
Nightly · Deterministic · Per action · Per owner
Regret-Exposure Days
When a decision's underlying conditions deteriorate past a materiality threshold AND the deterioration is demonstrably newer than the decision itself, the system records the timestamp. Once detected, the anchor is frozen — signal noise cannot reset it. The number grows until the committee reconvenes or the position transitions to a non-monitor stance. A CRO can quote it out loud.
Sticky · Materiality-gated · Semantic-boundary clear
Decision Invalidation
Every deal's investment thesis carries one of three states — holding, weakening, or invalidated. Movement between them is driven by sustained pressure, covenant breaches that don't recover, and decisions whose predictions stopped tracking reality. Once a thesis is marked invalidated, the platform refuses to accept any new commitment against it. The deal must be restructured or exited — those are the only paths out.
Tracked state · Not reversible · System-enforced
Three things the system structurally refuses
War Room won't open on a deal that isn't in defense mode — defense mode is a tracked state, not a button. IC approval is refused when portfolio-company accounting is unlinked, with a typed explanation of what to fix. New commitments are refused against a thesis the system has marked dead. Three refusals — none of them advisory, all of them auditable.
Refused, not warned · Reason named · Audit trail
Owner Accountability Score
Your partners' predictions, tracked against reality. When the action-outcome confirmation runs nightly and grades a closed action's predicted improvement against its measured result, the grade rolls up into an owner-attributed score. Over time the score is the record: how well this decision-maker's predictions have tracked the reality the platform measured. Partnership performance conversations stop being a vibes check.
Per partner · Over time · Adds to the record, never erases
How the loop closes
An action is taken. An expected improvement is recorded. The nightly sweep grades it. The grade rolls up into an owner accountability score — how well the decision-maker's predictions tracked reality over time. When the same owner's next decision lands, the system already knows their prior track record. The loop is closed at the partner level, not just the deal level.
What this means for an LP DDQ
The governance section of an LP DDQ — the one that asks how decisions are made, owned, and tracked — is no longer answered with prose. The answer is a structured record: total IC decisions, percentage documented, average days to decision, ownership coverage, prediction-vs-outcome tracking, regret-exposure exposure, supersession status of every signed export. Every metric pulled from a permanent record that builds itself — not a quarterly hand-assembled deck.
Where each mechanism lives in the product
- Self-grading playbook — visible on the Portfolio Operating Loop dashboard; every closed action shows the predicted impact alongside the measured result
- Regret-Exposure Days — visible on /portfolio/decisions and on each deal home; the count persists across visits and only clears when the committee reconvenes or the position transitions to a non-monitor stance
- Decision Invalidation state — visible on every deal home; once invalidated, the action panel will not accept a new commitment against the dead thesis
- Structural refusals — surfaced as named blocking reasons (with the fix the operator needs to take) rather than soft warnings the team can ignore
What this isn't
This is not an AI risk score. There is no black-box model predicting which deal will break. Every mechanism above is deterministic and explainable: the action-outcome confirmation is a measured delta against a recorded expectation; regret detection is a documented threshold against newer data; decision invalidation has named transitions you can point at. The product refuses to ship predictive scores it can't defend in writing.
The one predictive composite the platform does ship is the Decision Fragility Index — six measurable signals combined into a single number per deal, with the factor breakdown visible on the same screen. Every contributing signal is named, every weight is shown, every change is explainable in plain language. No black box, no “the model says so.” If a partner asks why the Decision Fragility Index moved, the six factors and their contributions are on screen.
Who this is built for
Risk officers and CROs
You finally have a number for "when did our last decision stop being correct." Regret-Exposure Days is sticky, materiality-gated, and survives signal noise. You can put it in front of credit committee.
PE operating partners
You record what an intervention is supposed to deliver, the system checks whether it did, the rollup feeds the next partnership performance conversation. Your playbook grades itself.
IC chairs and partnerships
Owner-attributed predictions, tracked over time, against actual outcomes. The committee decides; the system records the conviction; reality grades the conviction. The record builds itself.
LPs and audit committees
Governance section answers come from a permanent record that updates itself, not a quarterly hand-assembled deck. Every IC decision is timestamped, owner-attributed, and joined to the current state of every artifact it produced.
See the closed loop on a real position.
We'll walk a deal through a full decision cycle — commitment, recorded prediction, operating reality, graded outcome, supersession of prior memos, and the Regret-Exposure clock as it ticks.