docs/validity-envelope.md

FUTURE-TS validity envelope

This page states what the current repository supports and where claims should stop. It is intended for reviewers comparing the design paper, empirical paper, and code.

Supported today

FUTURE-TS v0.1.0 is a runnable benchmark package and integrity scaffold. Its strict benchmarks/v1 surface implements:

• task cards with issue times, horizons, delayed sources, revision metadata, resource budgets, anchors, metrics, and adaptation budgets
• submission and actual validation, including full visible-label coverage at score time
• leakage auditing through available_at, first_published_at, and last_updated_at
• archived prediction scoring with deterministic prediction_hash and manifest_hash
• strict pretraining manifests for benchmarks/v1; manifest source entries are included in the manifest hash and carry evidence/confidence metadata
• local sealed-runner MVP execution with platform-stamped platform_issued_at and platform_received_at, CPU/wall-clock limits, best-effort memory limits, and Linux network namespace isolation where available
• prequential cutoff scoring for tasks that declare cutoff_schedule, emitted as a backward-compatible aggregate task score
• capability-vector reports plus tier-weighted and rank-style aggregate leaderboard surfaces

This supports the claim that FUTURE-TS is an executable future-aware benchmark protocol and local evaluation package.

Not supported yet

The repository does not by itself establish a fully hosted, externally attested live benchmark. These remain service-layer or data-expansion milestones:

• immutable remote submission deadlines and live waves where labels physically do not exist at submission time
• signed external timestamps, container digest attestation, and artifact-bucket immutability
• production-grade no-egress execution on every host; non-Linux local runs are warning-only unless a Docker/Kubernetes --network=none backend is used
• fixed-hardware, platform-measured per-prediction runtime/memory telemetry for cross-model efficiency claims
• PEFT and full fine-tuning under sealed training/evaluation
• broad decision-utility coverage; v1 grounds capD in one newsvendor task, so grid redispatch, hospital staffing, and capacity-planning tasks are needed before making broad operational-utility claims
• frozen anchor prediction artifacts and alternate-anchor sensitivity tables
• a source ontology with canonical IDs and aliases for stronger pretraining overlap matching

Claim wording

Use:

FUTURE-TS introduces a runnable future-aware benchmark protocol for TSFMs, with task-card semantics, leakage auditing, strict submission validation, archived predictions, manifest-based contamination flags, multi-dimensional scoring, and an empirical TSFM.ai run demonstrating the protocol on real data. The current release is a local package plus sealed-runner MVP; hosted attested live evaluation is the next milestone.

Avoid:

FUTURE-TS structurally solves live benchmark integrity.

The latter requires the hosted attested service, immutable submission windows, and label release controls.

Covariates and multimodal context

Task cards may declare known covariates or multimodal context. Those fields are eligibility metadata unless the actual task-window payload handed to a submission contains the corresponding inputs. A task should only be described as operationally measuring covariate-aware or multimodal use when its runner payload includes those typed inputs.

Empirical scope

The empirical paper should be read as a first real-data slice: real hosted TSFMs, temporally constrained tasks, archived predictions, and diagnostic family differences. It does not establish a final ordering of TSFMs. Wider waves, more tasks, stronger manifests, external attestation, and full multi-budget adaptation are required before presenting a definitive public leaderboard.