Calibration

Most teams estimate once and never learn from the result. Calibration closes the loop: as real effort lands against your committed scopes, Scopingly measures how far its AI estimates were off for your team and corrects the bias over time. Future estimates get more accurate because the system learns from your history — not from a generic assumption.

How it works

Calibration reduces each completed work item to a single dimensionless number — the estimate-to-actual ratio r = actual effort / AI estimate. It aggregates those ratios (in log space, with statistical shrinkage) into a multiplier that corrects future estimates.

The correction is per-team and per feature type, and it shrinks gracefully:

team × feature type  ──▶  feature type  ──▶  your org's global

When your organization has no history to draw on — a brand-new org, or a thin segment — calibration falls back to the raw AI estimate, the cold-start identity k = 1.0. As you accumulate actuals, the correction takes over.

Baselines make it honest

Calibration measures actuals against the approved scope version the work was committed under. This is why an approved scope is superseded, never edited in place — a moving baseline would make every project look on-target and train the model on a fabricated history.

Where actuals come from

Actuals are real effort figures ingested from your tracker — for example, JIRA. A few deliberate stances shape the data:

• Team-level only. Actuals roll up to the team, never to an individual — there is no assignee dimension, by design. An estimation-accuracy tool must not become individual surveillance.
• Unattributed work is surfaced, not hidden. Tickets that map to no scope item roll up into a scope-leakage figure rather than being dropped — hiding them would make calibration systematically optimistic.

The cross-org calibration prior

Opt-in

The cold start — a new or thin org with no history — is exactly where estimates churn most. To help there, Scopingly can consult a platform prior: a base rate learned across organizations, sitting one tier above your org's own history.

team × feature type ──▶ feature type ──▶ your org's global ──▶ cross-org prior ──▶ k = 1.0

What is critical is what crosses the boundary — and what never does:

• Only ratios and one taxonomy label. The pool aggregates the dimensionless ratio r keyed by content-free feature type. Never PRD or scope text, titles, counts, dates, or team identity. The ratio reveals estimation bias magnitude, not what was built.
• k-anonymity gate. A shared bucket is used only when at least 5 distinct opted-in organizations back it (with enough effective sample). If a bucket drops below the threshold, it is removed — nothing under the gate stays queryable.
• Symmetric opt-in. Contribution and benefit are coupled: your org only reads the pool if it also contributes, and vice versa. The setting (calibration_pool_opt_in) defaults off and fails closed in both directions.
• Cold-start only. The prior is consulted only when your org's own history yields nothing. A data-rich org never reaches it, and a non-opted-in org is completely unaffected — its calibration stays strictly local.

Opting in

Opting into (or out of) the calibration pool is an org-admin setting. See roles & permissions.

Deletion and the pool

If your organization opts into the pool and is later deleted, its already anonymized contribution (content-free ratios under a random, unlinkable pseudonym) can be retained so the shared prior doesn't lose that learning — and so it doesn't collapse a k-anonymity bucket for other orgs. This is the default, and it's a clean escape hatch: the deletion request carries a flag, so you can override it and choose full erasure instead, wiping even the content-free ratios.

See data management for the full org-deletion flow.

Shadow first

New calibration factors default to shadow mode — they change no estimate and only report what they would do. A factor goes live per bucket only after a held-out backtest shows it reduces error, and it can be rolled back. Calibration never silently multiplies your estimates the moment it has a little data.

Availability

Calibration is available on all plans. The cross-org prior is opt-in and only ever fills the cold-start gap.