Evidence boundary / recompute before trust

Do not trust the story. Check the record and the claim state.

The public evidence layer is designed so that a visitor can distinguish a re-derivable fact from an operator attribution, a provisional result, a historical implementation, and a claim the current evidence does not establish.

Publicly recomputable

The energy record is a test of evidence integrity, not a performance pitch.

Directional calls were placed on a public timestamped record before the market outcome, resolved against Nord Pool prices, and preserved in Git history. The later governed evaluation did not show an advantage over persistence.

git clone https://github.com/Nordvei/sage-track-record
cd sage-track-record

# Inspect the public inputs
head predictions.csv
head actuals.csv
head accuracy_log.csv

# Inspect history for retrospective edits
git log --follow --stat predictions.csv

# Verify an OVL receipt when one is supplied
python3 ovl-verifier/ovl_verifier.py receipt.json
Honest result: the public record supports prospective grading and visible correction. It does not establish predictive edge, profit, or a better-than-persistence model.

Claim matrix

What can be checked, what is bounded, and what remains unestablished.

Public prediction chronologyRecomputable Prediction and outcome files plus Git history are public.
Energy predictive edgeNot established Governed result approximately 46.5%; no demonstrated advantage over persistence.
DecisionBundle packageBounded verification Package internal consistency is verified; full_record remains UNESTABLISHED.
OVL testnet mechanismClosed on testnet Payment, signed receipt, ledger correlation, re-derivation, and tamper rejection passed.
OVL external demandNot established The closure was self-paid and did not demonstrate a customer transaction.
P0-RL primary resultProvisional Execution reported complete; artifact review and adversarial challenge remain.
Worker-steering improvementZero verified loops No L2, L3, or L4 loop is currently claimed.
Internal closure ledgerDescribed property Internal append-only behavior is not yet exposed through a complete public chain export and verifier.

Evidence status vocabulary

A machine should not infer certainty from confident prose.

Verified / closed

Exact scope passed

The named acceptance criteria closed. Adjacent correctness, value, authority, and legal claims do not inherit that status.

Provisional

Primary finding accepted for review

The execution produced a credible result, but formal artifact review or independent challenge remains.

Bounded

Verified only inside a defined envelope

The result holds for the exact package, test, network, or evidence surface named.

Held

Operation or claim intentionally stopped

Evidence is insufficient, a negative result controls, or the next authorization has not been issued.

Historical

Valid dated context

The implementation or observation was real at its date; current runtime freshness is not inferred.

Unestablished

The evidence does not prove it

This is a typed result. It is not an invitation to silently fill the gap with general knowledge or marketing language.

Corrections remain visible

A defect is appended beside the original, not edited out of history.

The method separates chain integrity from source re-derivation, and verified facts from operator-attributed cause. A clean chain does not prove the source interpretation was correct. A plausible causal explanation does not borrow the chain's authority.

Examples of retained negative evidence

  • The previous 65% energy headline is superseded.
  • The public predictor is held as FALSIFIED_HELD.
  • DecisionBundle refuses the broader record when evidence is missing.
  • The OVL rail states self-payment and testnet status prominently.
  • P0-RL reports zero L2-L4 loops despite substantial internal learning activity.

Public standard

A green integrity result must never silently become a green correctness result.

Receipts attest records. They do not make the underlying decision correct. Governance can admit an artifact. It does not prove the artifact created value. Improvement can be verified. It does not grant authority.