Skip to content

Roadmap v10 — forecast evidence closeout

Status dated: 2026-09-27

Roadmap v10 completed its four implementation and audit workstreams (#397–#400). The closeout is blocked at empirical evaluation: the required point-in-time BTC/USD corpus is unavailable, the prospective evidence ledger has not accumulated enough eligible mature forecasts, production-parity replay has no metrics, and no candidate was preregistered or scored. This is not evidence for or against an accuracy improvement. No winner was identified, and there was no production configuration or forecast change.

Workstream outcomes and dependencies

Issue Deliverable and evidence Outcome / dependency
#397 Corpus acquisition audit and canonical machine-readable audit. Blocked: no admissible licensed, immutable, single-venue BTC/USD archive with point-in-time publication and revision provenance. Coverage is 0/32,136 target hours and 0/4,320 warm-up hours. Must pass before canonical replay.
#398 Prospective, append-safe shadow evidence ledger in src/btc_timesfm/research/shadow_deployment.py, schema v6. Forecast predictions and exact-target matured outcomes are write-once; reports account for eligible forecasts, maturity, failures, and versions. Runtime SQLite and status report belong under ignored .state/ (default DB: .state/shadow_deployment.sqlite); no private runtime database is committed. Implemented; current ledger evidence unavailable: no runtime database or status report was supplied to this closeout. The ledger is a prospective source of evidence, not an already populated empirical result. Its eligible, mature evidence is a prerequisite for #399.
#399 Production-parity replay readiness contract and blocked readiness report. Blocked: #397 corpus gate fails and no #398 ledger report was supplied. The machine-readable report has null ledger hash/counts and metrics: null; no replay or canonical skill estimate is available. It depends on both #397 and #398 passing.
#400 Candidate confirmation protocol and blocked candidate report. Blocked: parity diagnosis is unavailable; no candidates were preregistered or scored, and candidate_winner and metrics are null. Candidate selection depends on valid #399 evidence, followed by a separately frozen prospective confirmation.

The dependency sequence is #397 corpus + #398 accumulating prospective ledger → #399 production-parity replay → #400 candidate preregistration, scoring, and (only if selection criteria pass) disjoint prospective confirmation. The parallel #397 and #398 workstreams do not substitute for one another: a historical admissible corpus is required for replay, while the ledger accumulates contemporaneous, production-policy evidence.

Remaining gates

  1. Corpus: obtain and preserve the license, original files, acquisition manifest, and hashes for a single-venue BTC/USD hourly archive spanning the exact target interval 2023-01-01T00:00:00Z through 2026-09-01T00:00:00Z (32,136 hours) and contiguous warm-up interval 2022-07-05T00:00:00Z through 2023-01-01T00:00:00Z (4,320 hours). Every candle needs close time, first-publication time no later than close, and revision-0 provenance. Run the canonical audit and satisfy every gate; do not relax coverage, infer vintages, or splice venues.
  2. Ledger accumulation: operate the #398 prospective shadow ledger and preserve its status report and immutable database evidence. Accumulate eligible, provenance-complete forecasts with exact matured outcomes at 2h, 4h, 8h, and 16h, and account for all failures. No runtime database or status report was available for this closeout; #399 consequently records the ledger path, hash, and counts as unavailable. This closeout does not assert ledger coverage or maturity counts. The downstream replay gate requires all expected pairs to mature and no failed attempts.
  3. Parity replay: after both prior gates pass, rerun the #399 readiness audit, attest production-policy parity, freeze the corpus/ledger hashes, policy identity, code revision, command, and evaluation cutoff, then execute the preregistered exact-key comparisons and publish metrics with dependence-aware uncertainty and family-wise multiplicity adjustment.
  4. Candidate confirmation: only after valid parity evidence, freeze and hash the bounded candidate-family preregistration before any scoring. Retain every attempt and failure. A selected candidate would still require disjoint prospective confirmation for at least 30 calendar days and at least 200 exact mature pairs per supported horizon, plus human review; selection or shadow success does not automatically promote or alter production.

Final evidence disposition

All four workstreams have auditable implementation/status artifacts, but the empirical evidence chain did not reach a scoreable evaluation. The only quantified corpus result is absence (0/32,136 target and 0/4,320 warm-up hours);

399 has metrics: null; #400 has no scored candidates and a null winner. No

empirical accuracy conclusion—positive, negative, or inconclusive—is supported. In particular, this closeout makes no accuracy-improvement claim and authorizes no production configuration change.