Calibrating Cortex with feedback

How calibration works

Calibration is the process of teaching Cortex the difference between what it scored and what you believe the correct score should be. The workflow is straightforward:

1. Cortex investigates an account and assigns a confidence score.
2. A controller or reviewer examines the score and findings. If the score is wrong (too high, too low, or based on a misinterpretation of the data), they click Correct Score.
3. The controller enters the corrected score (a value between 0.00 and 1.00) and selects a reason from the correction categories.
4. The correction is stored as a calibration event, linked to the account, the period, the original score, and the corrected score.
5. In subsequent investigations, Cortex checks the calibration history for matching patterns and adjusts its scoring accordingly.

The correction workflow

Navigate to the account's reconciliation and open the Cortex panel. The current confidence score is displayed at the top with the color band (GREEN, AMBER, or RED). Click Correct Score to open the correction form.

Correction fields

• Corrected score: the score you believe is appropriate, entered as a decimal between 0.00 and 1.00. You can also click the GREEN, AMBER, or RED band to set the score to the midpoint of that range (0.92, 0.68, or 0.25 respectively).
• Reason category: select the primary reason the score was incorrect:
  • Expected variance: the variance Cortex flagged is a known, expected difference (e.g., a timing difference that clears in the next period).
  • Account characteristic: the account has a structural characteristic that Cortex does not understand (e.g., the account always carries a large reconciling balance by design).
  • False positive finding: Cortex flagged something as an issue that is actually normal activity for this account.
  • Missed issue: Cortex scored the account too high and missed a real problem that the reviewer identified.
  • Threshold mismatch: the materiality threshold used by Cortex does not match the threshold your team actually applies.
  • Other: free-text explanation for cases not covered by the categories above.
• Notes: optional free-text field for additional context. Useful for documenting why this account behaves differently than Cortex expects.

Who can correct scores

Score corrections require the Controller or Administrator role. Staff accountants and senior accountants can view scores and findings but cannot modify them. This ensures that calibration data comes from experienced reviewers whose judgment the system should learn from.

Correction patterns

Cortex does not treat each correction as an isolated event. It looks for patterns across corrections to build reusable rules. The pattern matching works on three dimensions:

Account-level patterns

If the same account receives the same type of correction in consecutive periods, Cortex recognizes the pattern. Example: the Prepaid Insurance account is corrected upward in January, February, and March because its large reconciling balance is by design (monthly amortization against an annual payment). By April, Cortex applies the adjustment automatically and notes “calibrated based on 3 prior corrections” in the investigation log.

Account-group patterns

Corrections on one account can propagate to similar accounts. If three different prepaid accounts all receive “account characteristic” corrections pointing to the same behavior (large expected reconciling balance), Cortex recognizes the pattern and applies the adjustment to all accounts in the Prepaid Expenses group, even accounts that have not been individually corrected.

Entity-level patterns

Some correction patterns are entity-specific. A foreign subsidiary might have a structural FX timing difference that does not apply to domestic entities. Cortex keeps entity-level patterns separate from cross-entity patterns to avoid incorrect generalization.

Threshold overrides via Cortex rules

For situations where repeated corrections indicate a systematic misalignment, you can create explicit Cortex rules instead of relying on automatic pattern recognition. Rules are deterministic overrides that take effect immediately.

Navigate to Settings > Cortex > Rules and click New Rule. Each rule has:

• Scope: the accounts the rule applies to. Options include a specific account, an account group, an entity, or all accounts matching a filter (e.g., all accounts with type = Asset and balance > $1M).
• Condition: the situation that triggers the rule. Common conditions:
  • Variance below a custom threshold (overrides the default materiality threshold for scoring purposes).
  • Reconciling item age below a custom limit (overrides the default 90-day stale threshold).
  • Finding type suppression (tells Cortex to stop flagging a specific finding type for these accounts).
• Action: what the rule does when the condition is met:
  • Adjust factor weight: increase or decrease the weight of a specific scoring factor for these accounts. For example, reduce the variance factor weight from 35% to 20% for accounts where variances are structurally expected.
  • Set factor floor: set a minimum sub-score for a factor. Useful when a factor would otherwise penalize the account for expected behavior.
  • Suppress finding type: prevent Cortex from generating findings of a specific type. The data is still analyzed, but the finding is not surfaced.
  • Override threshold: set a custom materiality or stale-item threshold that differs from the entity-level default.

Rules are version-controlled. Every change is logged with the user, timestamp, and the previous rule definition. You can revert a rule to any prior version.

Getting smarter with each close cycle

The calibration loop is designed to require less human input over time. Here is how Cortex improves across successive close cycles:

Cycle 1: Baseline

Cortex runs its first sweep with default settings. It has no historical data and no calibration patterns. Scores may require significant correction. This is expected. Cortex is learning your organization.

Cycle 2: Pattern detection

Cortex has one period of corrections to reference. It applies account-level adjustments where corrections were made and flags accounts where it is uncertain whether the prior correction should recur.

Cycle 3: Group propagation

With two periods of corrections, Cortex begins recognizing account-group patterns. Adjustments propagate to similar accounts. The number of corrections typically drops by 40-60% compared to Cycle 1.

Cycle 4+: Steady state

By the fourth close cycle, most structural patterns are captured. Corrections are limited to genuinely new situations: a new account, a process change, or an unusual transaction. Controllers report spending 70-80% less time on score corrections compared to the first cycle.

Viewing calibration history

The full calibration history is accessible in two places:

• Account level: on any account's reconciliation, click Cortex > Calibration History. This shows every correction for this account across all periods, the patterns Cortex detected, and any active rules.
• Organization level: in Settings > Cortex > Calibration Dashboard, see the aggregate correction rate across all accounts, the most-corrected account groups, the correction trend over time, and the rules in effect.

The calibration dashboard is a key metric for Cortex health. A declining correction rate means Cortex is learning. A flat or increasing correction rate suggests that rules need attention or that the underlying data quality has changed.

Best practices

• Correct early, correct consistently: the more corrections Cortex receives in the first two cycles, the faster it converges. Do not ignore a wrong score because “it's close enough.” Every correction teaches Cortex something.
• Use reason categories, not “Other” structured reason categories enable pattern matching. A free-text note under “Other” is logged but cannot be matched automatically. Use the specific categories whenever possible.
• Create rules for known structural patterns: if you know that all intercompany accounts carry a timing variance by design, create a rule immediately rather than waiting for Cortex to learn the pattern from corrections. Rules take effect instantly.
• Review the calibration dashboard quarterly: check the correction trend. If certain account groups are still being corrected after four cycles, consider whether a rule would be more effective than relying on pattern detection.
• Involve the right people: calibration is most effective when corrections come from controllers and managers who understand the nuances of each account. A correction from someone who does not know the account's history can teach Cortex the wrong lesson.