Algorithmic governance and carrier response

1. Algorithmic determinism and Daubert alignment

SafeNY carrier analyses are deterministic outputs of published statistical procedures applied to public FMCSA records. The system is structured to satisfy the four prongs of the Daubert standard for scientific evidence.

Testability and falsifiability

Every published figure is computed by a documented procedure. Any third party can apply the same procedures to the same FMCSA inputs and obtain identical results. The methodology is therefore empirically testable and falsifiable.

Known error rate

Methodology v6.5.h documents calibration parameters (Bayesian shrinkage prior_strength = 10, exposure floor = 0.005, EMA α = 0.40) and the false-positive rates these parameters produce on the May 2026 FMCSA snapshot. Interval estimates are computed alongside point estimates rather than substituted for them.

Peer-reviewable methodology

The entire methodology — formulas, calibration constants, source data references, version history — is published at /methodology and anchored cryptographically (SHA-256). Each version is immutable once released; calibration changes produce new versioned methodologies, not retroactive edits.

Standards and controls

No human analyst interprets, edits, weights, or overrides algorithmic output at any stage — neither during data ingestion from FMCSA, nor during score computation, nor during page rendering. Analyst discretion is structurally absent from the analysis chain. This is enforced at the data-access boundary: the score-generation pipeline reads only public FMCSA records and the published methodology configuration.

The methodology version applied to a specific carrier analysis is recorded at snapshot generation time and preserved cryptographically in the snapshot's hash. Subsequent methodology updates do not retroactively alter prior snapshots.

Methodology evolution & self-correction

The peer-reviewable standard above is operational, not aspirational. The methodology is revised when review identifies a weakness, and each revision is disclosed. The current release, v6.5.h, removed several constructs from the prior version that could not be defended as testable, federally-grounded measures:

• The statistical settlement-value range and its derived 90th-percentile figure were removed — settlement-value estimation is not a testable property of the federal record and risked reading as legal-outcome prediction.
• The attorney-value ranking formula and its verdict-denominated inputs were removed for the same reason.
• An unsupported conditional-variance attribution was excised; the volatility framework now cites Engle (1982) alone.

Full errata and the version-to-version diff are published at /cite/integrity. Disclosed self-correction is the behavior the Daubert peer-review prong contemplates — not a liability.

2. Correcting FMCSA data

SafeNY republishes public safety data from the Federal Motor Carrier Safety Administration (FMCSA). We do not author, score, or alter the underlying federal records, and cannot change a carrier's FMCSA data on request. If you believe a federal record shown here is inaccurate, the correction is made at the source: file a Request for Data Review through FMCSA's DataQs system (https://dataqs.fmcsa.dot.gov), the official mechanism for challenging FMCSA data accuracy. Once the federal record is corrected, the change is reflected here on our next data ingest. Each record carries an "as of" sync date and reflects the federal data as of that date, not real time. Questions about this page: governance@safeny.com.

3. Methodology governance

SafeNY's methodology is governed by:

• Versioning — Each methodology version is hashed (SHA-256), published as immutable, and anchored at /cite/integrity. Methodology version v6.5.h produces output that can be retroactively verified against the published v6.5.h hash.
• Calibration cadence — Parameters are recalibrated quarterly against new FMCSA snapshots; each calibration event is logged with input snapshot hash, parameter changes, and rationale.
• Archive — Published methodology versions are cataloged at /cite/integrity with their SHA-256 anchors and any erratum trail.
• Limitations disclosure — Known model limitations and miscalibration regimes are documented at /methodology.

Methodology changes do not retroactively alter prior snapshots. A snapshot generated under v6.5.h remains v6.5.h indefinitely and renders as such regardless of future methodology revisions.

4. Scope and limitations

SafeNY analyses are:

• Statistical — estimates of distributions over comparable incident populations, not predictions of specific case outcomes
• Algorithmic — deterministic functions of public FMCSA data
• Decision-support — intended to inform attorney consultation, not to replace legal counsel

SafeNY analyses are not:

• Legal advice (we are not attorneys)
• Verdicts (we do not adjudicate fault)
• Determinations of carrier liability for any specific incident
• Predictions of insurance recovery for any specific case
• Comprehensive (we analyze public FMCSA records; private data not disclosed by FMCSA is not included)

For case-specific legal advice, consult an attorney licensed in the relevant jurisdiction.

5. Data sources and updates

Primary sources (all public):

• FMCSA SAFER (carrier profiles)
• FMCSA SMS (Safety Measurement System: BASIC scores, measures, percentiles)
• FMCSA L&I (Licensing & Insurance, via catalog.data.gov)
• FMCSA Crash Reports (5-year window)

Update cadence: periodic ingest (approximately every 6 hours). Each record carries an "as of" sync date and reflects the federal data as of that date, not real time.

Erratum process: factual errors in the underlying federal records are corrected at the federal source via FMCSA DataQs; federal corrections flow into our next ingest.

Source data captures are preserved cryptographically at analysis-generation time. This serves two functions: (a) it defends against post-hoc carrier record modifications by preserving the federal record state at analysis time, and (b) it enables Daubert-grade audit by allowing any party to verify our analysis against the same FMCSA snapshot we used.