B07 · Outcome learning architecture

The Learning Loop

How the organism turns real outcomes into governed behavior change: a forecast sealed before the outcome, a calibration pair, a versioned proposal, an independent gate, and a promotion that can be rolled back. No silent self-modification anywhere in the chain.

B07 conclusion

The learning architecture exists. The real missing seam is feeding it truthful pre-outcome predictions and production outcomes.

Design specification board. Chain links, adapters, and protocols below are the documented architecture and its current source owners, not live telemetry from a running system.

01

The canonical chain

One shared outcome schema in. Promoted behavior out. Every link owned.

Every producer in the organism writes outcomes in one shape, neuruh.outcome-record.v1. From there the record moves through calibration, proposal, gate, and governance before any worker, recipe, or routing behavior is allowed to change. Step through the chain.

02

B07 preflight

Sep-5 court. Approximately 14 outcome producers mapped.
F1

No new outcome schema required. The existing v1 record is sufficient.

F2

Several producers already adapted. Part of the estate already writes the shared shape.

F3

Additional adapters required. The remaining producers need bridges, listed in section 04.

F4

DeedSonar, LandOS, and commercial outcomes must bridge into the shared OutcomeRecord.

F5

N-Lab evidence feeds promotion evidence. It never silently mutates production.

03

The starvation defect

Calibration cannot be retroactively invented.

A real learning loop needs a forecast recorded before the outcome: the prediction, its confidence, the model and recipe versions that made it, the evidence available at that moment, and the predicted result class. When reality arrives, the observed outcome is captured against it. Only paired observations calibrate. Try both paths below.

Forecast

SEALED BEFORE OUTCOME
  • predictionWhat is expected to happen
  • confidenceStated probability, fixed at seal time
  • versionModel / recipe / version that predicted
  • evidenceEvidence available at prediction time
  • cohortCohort and context of the prediction
  • classPredicted economic / result class

Observed outcome

CAPTURED WHEN REALITY LANDS
  • outcomeWhat actually happened
  • timingWhen it landed, time-to-outcome
  • valueEconomic value where observable
  • stateSuccess, failure, or partial state
  • evidenceEvidence for the observed result
  • caveatsCausal caveats on the pairing

What paired observations make measurable

Without the pair, learning is narrative hindsight. With it, the organism can measure:

Recipe calibration
Source quality
Confidence calibration
Routing quality
False positives / negatives
Time-to-outcome
Economic value
Cost vs value
Failure concentration
04

Required adapter families

The bridges between producers and the shared record.
A0

Fold / unit normalization. Bring every producer's units and shapes into the shared record.

A1

Founder Runtime execution projection. Project runtime executions into outcome records.

A2

LandOS bridge. lead_outcomes bridge plus scheduler into the shared record.

A3

Opportunity Intelligence receipt. Receipt to forecast seal and prediction log.

A4

DeedSonar forecast capture. Seal the forecast at decision time, inside the product.

A5

Commercial terminal states. Paid, delivered, refunded, repeated: the states that end a commercial arc.

A6

N-Lab to promotion evidence. Experimental evidence packaged for the gate, never a silent production change.

A7

Calibration dimensions. The axes along which pairs are scored and sliced.

05

The governance boundary

What automatic learning may touch, and what it may never rewrite.

Automatic learning may propose

Adjustments that stay inside versioned, reversible scope.
  • Routing weights
  • Recipe calibration
  • Retry / backoff behavior
  • Context selection
  • Tests

It may not silently rewrite

These stay behind proposal, tests, and independent court.
  • Governance rules
  • Privileged capability
  • Destructive authority
  • Public exposure
  • Legal / financial commitments
  • Canonical mutation semantics

Changes on the right-hand side remain versioned proposals with tests and an independent court. The loop proposes. Governance disposes.

06

Topology-to-outcome protocol

Sep-7 experimental protocol. Does changing the organization around the same mission change measured results?

Pre-outcome record

  • Mission archetype and measurable success condition
  • Evidence available at decision time
  • Predicted outcome and confidence
  • Expected cost and latency
  • Model / provider / version
  • Recipe / strategy / version
  • Active organ / team topology or N-Cell signature
  • Tools and capabilities available
  • Important alternatives and known unknowns

Post-outcome record

  • Actual outcome and time-to-outcome
  • Economic value where observable
  • Actual cost and latency
  • Errors, reversals, and human interventions
  • False-positive / false-negative class
  • Causal caveats
  • Customer / operator acceptance
  • Promotion, demotion, or retest recommendation

Comparison protocol

For repeatable mission classes, N-Lab compares five configurations and measures outcome quality, calibration error, completion rate, intervention rate, cost, latency, policy failures, and economic contribution.

V1Simple baseline workflow V2Current canonical topology V3One-organ ablations V4Alternate provider, same topology V5Alternate topology, same provider
Hard rule: do not infer a systems advantage from one successful run. The comparison only counts across repeatable mission classes with paired records.
07

Research maturity ladder

Eight rungs. The commercial objective arrives before any broad research claim.

The ladder runs from recoverable state to independently confirmed transfer. Climb it in order; each rung is a claim the system has to earn with paired evidence. Select a rung.

08

The second loop: demand learning

OutcomeRecord teaches whether executed work was correct. Demand telemetry teaches what to build next.

Buyers repeatedly asking for something the organism cannot yet serve is a learning signal of its own. It is captured privacy-safe: only the bounded fields needed for commercial learning, never unnecessary raw prospect text.

Privacy-safe demand signal

F1Problem class F2Requested outcome F3Buyer / market class F4Existing offer / recipe match F5Missing ingredient or blocked capability F6Price / urgency band, when voluntarily supplied F7Conversion state F8Reason lost / rejected, when known

The missing-ingredient loop

Repeated unserved demand becomes a ranked build candidate instead of another idea backlog. Product development turns into an evidence-backed response to market pull.

Demand maturity

Demand is not proof of willingness to pay. Searches, questions, likes, and comments are weak demand. Productization should rise with this ladder. Select a rung.

09

Reasoning compilation

Successful intelligence should get cheaper. An Offer Foundry law, promoted.

Repeated successful reasoning should be compiled downward into cheaper logic. The preferred execution ladder tries the cheapest reliable rung first and escalates only when ambiguity, consequence, or economic value justifies it.

1
Deterministic software / cached known resultNo model call at all when the answer is already proven.
2
Deterministic Neuruh rules / schema / queryTyped logic over the organism's own structures.
3
Local or low-cost modelCheap reasoning for bounded interpretation.
4
Inexpensive remote reasoningHosted models for work that outgrows the local rung.
5
Frontier reasoningOnly when ambiguity, consequence, or economic value justifies the cost.

How a pattern compiles down

Frontier reasoning→ Reusable prompt / worker contract→ Recipe→ Typed decision rule / deterministic transform→ Cache / precomputation, where valid
Do not compile ambiguity away. Keep model reasoning where the task still requires interpretation, novel evidence, or adversarial judgment. The objective is not minimum token spend. It is the lowest reliable cost for the required evidence quality and consequence level.

Economic learning fields

Every outcome record carries the economics, so compilation decisions are made on cost against value, not on token counts.

Model / provider Topology Recipe / ingredient versions Quoted vs actual price Estimated vs actual time Model / tool cost Founder / operator effort Support burden Success / refund / rework Customer outcome Repeat / expansion
10

The economic north star

The first real paid tape is the proof.

The first real paid, customer, or transaction tape should create one complete pass through the loop. That pass is the proof the organism can learn from reality rather than merely remember conversations: learn which evidence, worker, model, tool, and topology should handle a real objective, then prove the choice improves outcomes.

Forecast→ Bounded action→ Real customer / transaction outcome→ OutcomeRecord→ Calibration→ Learning proposal→ Governed promotion→ Better next mission