AFMR Lift Evidence / public-card consumer / Phase 1

Phase 1 · Configuration only · No results

Lift evidence without the universal claim.

Future public cards will report measured differences between matched governed and wild cohorts inside one exact model–harness–protocol configuration cell. This route publishes no lift result today.

CARDS: 0 · UPDATED: NOT YET

Current status

No lift results are published. The machine index is intentionally empty with evidence_maturity: "configuration_only" and cards: []. It declares the consumer contract without inventing evidence.

AFMR.ai consumes the redacted public-card projection from the canonical AFMR LiftComparisonManifest. It links to that versioned schema and does not copy or redefine it. The canonical codebase owns artifact validation, signatures, provenance, challenge, and supersession logic; this site is a read-only publication and discovery surface.

What lift means

Lift is a cohort-level comparison between matched conditions inside one configuration cell:

lift(cell, metric) = outcome(comparison, cell, metric) − outcome(baseline, cell, metric)

The baseline and comparison labels are defined by the pre-registered manifest. A card must disclose cohort counts, metric direction and bounds, observed values, the lift method, uncertainty where available, review state, and provenance.

Lift is adoption evidence for one declared configuration. It is not individual agent reputation, a portable guarantee, or proof that the same result holds for another model, version, harness, task, evaluator, protocol, time period, or AFMR release.

The configuration cell is the claim boundary

modelProvider, name, and exact model version.
harnessName, version, and configuration digest.
governance protocolContent digest of the exact protocol.
taskTask identifier and version.
evaluatorEvaluator identifier and version.
AFMRThe exact AFMR version used for context.
cell identityStable cell_id plus sha256 cell_digest.
observation windowStart and end times for the bounded cohort.

Within-cell lift can be computed only when baseline and comparison observations share this complete identity. Cross-cell aggregation is descriptive unless a pre-registered method explicitly supports it.

Stable public-card projection

The canonical public-card contract is version 1.0.0. AFMR.ai consumes these top-level fields exactly:

FieldRole
schema_id, schema_versionBind the artifact to the canonical versioned schema.
card_id, manifest_id, artifact_digestStable identity and content-addressed integrity.
evidence_maturityConfiguration-only through independently reviewed evidence maturity.
configuration_cellThe full model, harness, protocol, task, evaluator, and AFMR claim boundary.
preregistrationCommitment, metric IDs, cohort definition, and timestamp fixed before outcomes.
cohort, metrics, within_cell_liftBounded counts, declared measurements, and within-cell comparisons.
review, challengeIndependent scrutiny and dispute state.
lifecycle, supersessionCurrent validity and links between replaced cards.
provenance, signatures, verification_linksSource digests, redaction policy, signing evidence, and resolvable verification.
limitationsMachine-enforced configuration, individual-reputation, universal-lift, and safety boundaries.

Phase 1 permits exactly one item in within_cell_lift. Additional metric comparisons require a later versioned contract rather than an unannounced shape change.

Evidence, review, and lifecycle states

DimensionAllowed statesInterpretation
Evidence maturityconfiguration_only, preregistered, measured, review_ready, reviewedA staged evidence record. “Measured” is not “reviewed.”
Reviewunreviewed, in_review, accepted, changes_requested, challenged, resolvedThe current independent review state.
Challengenone, open, upheld, rejected, withdrawn, resolvedWhether a result is contested and how the contest ended.
Lifecycledraft, active, suspended, superseded, revoked, expiredWhether the card is currently available for reliance.
Supersessionnone, supersedes, supersededExplicit previous/next card links preserve history.

“Preliminary” corresponds to evidence that has not reached independent reviewed maturity. A replicated result requires separately declared supporting evidence and does not erase its configuration boundary. A superseded card remains part of the audit trail but is not current.

Future verification path

01Resolve schema
02Verify digest & signatures
03Inspect provenance & review
04Check lifecycle & scope

When cards exist, verification_links can resolve the source manifest, canonical schema, current status, signed evidence, and any challenge record. Missing required provenance or verification material must not be treated as confirmed lift.

Every future card must preserve these machine-readable limitations:
  • configuration_scoped: true
  • individual_reputation: false
  • universal_lift_claim: false
  • safety_guarantee: false

A card never guarantees future performance, universal lift, agent safety, market admission, or authority to transact.

Machine resources