The question, the evidence, the decision, and the limit.

The problem

SAGE contained many components described as learning, recursive, adaptive, replay-based, experimental, or self-improving. The load-bearing question was not whether those components existed. It was whether a verified result from an external worker changed how the next worker was treated.

The method

The P0-RL pass classified parameter learning, meta-learning, structural learning, closure-derived learning, experimental learning, and repair/policy learning. It then traced whether each mechanism merely recorded an event, changed internal SAGE behavior, changed a later worker invocation, produced independently verified improvement, or transferred retained learning across providers.

L0  record, classify, queue, or preview
L1  result is consumed by an internal SAGE mechanism
L2  retained result materially changes a later worker invocation
L3  the change improves an independently verified held-out result
L4  the retained improvement transfers across workers/providers

The finding

SAGE already has a substantial active internal learning ecology. Hebbian Paths A and C, SAGERecursive._meta_update, SSM, DDSS modulation, MML, Discovery, G1, shadow learning, KAS replay, repair logic, SCAL, ProofKit, and related processes had current evidence or current output.

But the programme found no current mechanism that learns how to treat the next external worker differently. Internal learning was real; worker-steering learning was absent.

The engineering decision

The correct reuse decision is to compose Track-2, G1, and KAS with one deterministic host-owned steering consumer. No new learning database or general learning engine is required.

Rejected architectural shortcuts

• Do not connect worker closure directly into Hebbian learning.
• Do not use SSM or DDSS parameter modulation as worker-task steering memory.
• Do not use MML or Discovery as a substitute for tool, context, budget, and verification policy.
• Do not call the replay queue worker steering merely because it stores replayable events.
• Do not put another AI model in charge of the first steering consumer.

What this demonstrates

Complex-system re-derivation, semantic discrimination between different kinds of learning, multi-worker programme direction, architecture reduction, and the ability to preserve a zero result instead of converting activity into an improvement claim.

The problem

A verification package can easily become a persuasive-looking wrapper around evidence it does not actually establish. The objective was to create a candidate that could verify its own internal consistency, detect tampering, reject unsupported verification profiles, and refuse the broader claim when external evidence was missing.

The result

A qualified exact-SHA candidate passed 315 tests in a clean worktree. Its P1 archive reproducibly verifies package internal consistency. The broader full_record result remains UNESTABLISHED.

Why the refusal matters

The strongest property is not that the package always returns a green result. It can produce a typed negative boundary when the evidence needed for a claim is not present. The verifier does not silently upgrade package consistency into chronology, external effect, correctness, transferability, compliance, or market value.

What this demonstrates

Exact-artifact qualification, test and hash discipline, adversarial claim reduction, refusal semantics, and the separation of package integrity from broader truth.

The problem

Could a machine-to-machine attestation be issued only after a real settled testnet payment, cryptographically signed, bound to its mandate, recorded in the ledger, and independently re-derived without trusting the issuer?

The result

The first mandated paid issuance landed at ledger sequence 2000. An independent verification pass recomputed the content hash, verified the Ed25519 signature against the published key, correlated the payment and ledger row, checked the hash-chain predecessor, and rejected four altered variants.

Security boundary

The payment key remained operator-side and never entered the agent context. Issuance and verification were deliberately separated so the issuer did not bless its own output.

What this demonstrates

Cryptographic receipt design, payment-to-evidence correlation, operator/agent key separation, independent re-derivation, and scoped closure criteria.

The problem

High raw directional accuracy had been interpreted as predictive performance. Later governed evaluation showed that the result was substantially explained by persistence and did not demonstrate an out-of-sample edge.

The decision

The optimistic performance narrative was withdrawn. The predictor was held. The machine-facing surfaces were corrected so raw observational numbers could not be read as validated predictive performance without the adjacent governed status.

What remained valuable

The public, timestamped record; daily history; external Nord Pool outcomes; typed correction; tamper evidence; and the method for letting an outsider re-grade the result without trusting the operator.

What this demonstrates

Willingness to falsify the original product claim, preserve the negative result, correct public interpretation, and retain the evidence mechanism that survived.

The problem

Agent activity, governance decisions, closure events, and consumed substrate context were distributed across domain-specific surfaces. Cross-agent comparison required a generalized record without allowing the record itself to widen authority.

The implementation

The design was reduced to four append-only record families: agent_invocation, governance_admission, closure_event, and substrate_consumption. Runtime hooks were added at energy, trading, governance, and engineering dispatch points with non-blocking writer semantics and explicit authority_granted separation.

Why it matters now

P0-RL determined that Track-2 is the correct factual base for worker steering. It is not itself operational steering memory. A later deterministic consumer must interpret qualified facts and alter the next worker envelope.