P15 · Learning · Rendered from source

P15 — Continuous corpus and production learning: first-principles synthesis

Continuous corpus and production learning

251 lines15,902 bytessha256 3490ece60db7

P15 — Continuous corpus and production learning: first-principles synthesis

Part: P15 · Lane: S1-L5 · Run: 2026-08-27-sprint-1-fable Author: ACTIONIST-S1-L5-RUNTIME (Opus 5[1m]; wave-1 census under Fable 5[1m]) Status: research only · DESIGN_ONLY · UNEXECUTED · NOT_ADMITTED · admitted_blocks 0


1. The objective, and the honest starting position

P15's thesis is that every new component, repository, conversion and production outcome should improve the shelf, taxonomy, recipes and future recommendations.

The honest starting position must be stated first, because it determines the entire design:

There is currently no production learning data of any kind. Zero accepted plans, zero recorded client edits, zero incidents, zero maintenance outcomes, zero admitted blocks. The local AutoSaaS loop is procedural, not measured — its own maturity section records no machine-readable template registry, no SaaS built by the method, no automated freshness check, no command runner. Ledger entry A37 ("production learning will improve retrieval and reuse decisions") is a hypothesis.

Therefore P15 is not a system to be tuned. It is a system that must work correctly on its first observation and remain correct at ten. Cold start is not an edge case; it is the only case that currently exists. Any mechanism requiring thousands of observations per arm is disqualified regardless of how well it performs elsewhere.

2. The regime, stated as a constraint

PropertyTypical recommender/bandit assumptionActionist reality
Observations per arm10³–10⁶10⁰–10¹
Feedback latencyseconds to hoursdays to months (maintenance outcomes)
Feedback costfree (a click)expensive (a build, a client, an incident)
Armsthousands, cheap to trydozens, each with real adaptation cost
Wrong choice costone bad recommendationa client project
Data sharingone operator's own usersseparate clients, contractually private

Invariant I-P15-1: the sparse, expensive, privacy-partitioned regime disqualifies most of the external prior art's machinery, and the design must select for estimators that are honest at n=10. Bayesian and preference-learning methods (Bradley-Terry/Luce, Thompson sampling over posteriors, hierarchical priors) survive this test. Frequentist A/B testing on tens of builds does not.

3. The invariant imported from the local estate

The Universal Block Framework's evidence ladder carries an anti-averaging rule: evidence_tier may never exceed the minimum tier across the nine families — "A single T1 family pins the whole block at T1 regardless of eight T4s."

This directly contradicts how recommender systems work. Recommenders compute weighted scores, and a weighted score is designed to let strength in one dimension compensate for weakness in another. For ranking films that is correct. For ranking a capability that will be embedded in a client's production system, it is catastrophic: an asset with excellent adaptation cost, excellent workflow success and unresolved rights must not outrank a merely-good asset with clean rights. Averaging makes exactly that mistake.

Invariant I-P15-2: asset ranking is min-gated on the weakest evidence family, never a weighted average. This is the clearest place where first-principles reasoning from the local estate diverges from the entire external census, and it should be treated as a deliberate design divergence rather than an oversight.

4. What the shelf actually contains today

Ranking presupposes something to rank. Current state:

  • 8,515 joined UI identities, of which 3,506 are source-bearing; much of the newer 7,949-entry

harvest is bundle/metadata oriented.

  • 500-candidate GitHub corpus; 270 of 1,700 industry-repository pairs complete at the measured

baseline (1,430 deficit).

  • OpenConnector generated 1,445 provider definitions and 15,156 actions locally, but its global

connection storage failed the tenancy model.

  • The remembered ~1.3M-repository index and the 850k/80k Mini corpus have no authoritative path

in current receipts (A20/A21 unknown).

  • Zero admitted blocks.

Invariant I-P15-3: catalogue size is not supply. The OpenConnector case is the cleanest demonstration: 1,445 providers and 15,156 actions, and the storage model made the whole thing unsafe to inherit as-is. A ranking signal derived from catalogue breadth would have ranked it top.

Consequence for refresh design: P15 must be designed against an unknown-size, unknown-location corpus and must not assume the large feed exists.

5. Constraints

#ConstraintSource
C1Cross-client aggregation requires explicit contractual permissionG7 client/legal gate
C2Learn from structural facts only — never client content, schema names or copyEvidence + privacy discipline
C3Models cannot author ranking changes; the deterministic pipeline owns writesAuthority invariant
C4Zero-tolerance events are never averaged into a scoreZero-tolerance classes
C5Every signal carries source, date and evidence classEvidence discipline
C6Unknown stays unknown; absence is not a negative signalLedger vocabulary
C7Revocation propagates to every composing assemblyFramework invariant 6
C8Regulated industries excluded from cross-client aggregation by defaultContract-pressure column

6. The signal inventory — and where it comes from for free

The most efficient finding of this lane is that P13 and P14, designed properly, already emit P15's inputs. No separate telemetry programme is needed for the primary signals.

SignalEmitted byFeeds
Accepted AssemblyPlanP13 approval / composition lifecyclePositive fit evidence
Edit volume and verb mix per capabilityP13 operation logFit signal — heavy editing means poor default fit
Rejected offersP13 offer logCounterfactual, prevents selection bias
Refusals (out-of-algebra requests)P13 ABSTAIN logDemand/roadmap signal
Escape-hatch/fork rateP13 state transitionsBounded-editing health
Adaptation hours by categoryBuild instrumentationPrimary ranking signal (replaces stars)
Glue size per bindingSolver predicted_glue_loc vs actualIntegration cost
Workflow defects post-releaseP14 acceptance suiteQuality
Incidents attributed per capabilityP14 attribution (needs stable capability ID)Demotion
Maintenance and upgrade-break eventsP14 upgrade pathBurden + upgrade-safety evidence
UNDERDETERMINED resultsCompatibility solverContract defects, not model failures
Falsified assumptionsEvery buildAssumption-ledger maintenance

Two of these deserve emphasis.

UNDERDETERMINED aggregates are contract defects. The composition architecture already names this as "the single most valuable output" for the contract lane: each such result names a field the Block Contract is missing. Aggregating them redirects effort from burning model budget to fixing the contract layer, which is where the defect actually is.

Falsified beliefs are a first-class output. AutoSaaS §8 lists "failed assumptions" among the things each run must update. No system in the external census records which belief an outcome falsified — they all rerank items. This is a genuinely local idea and a candidate differentiator: it converts the assumption ledger from a static document into a maintained artifact.

7. Exploration in a client-delivery business

Pure exploitation freezes the shelf: the first-ranked asset stays ranked because nothing else is ever tried, and the ranking becomes a ratchet rather than a measurement. Some exploration is structurally necessary.

But exploration in this business means trying a less-proven capability on a paying client's project. That is only acceptable inside a safety envelope:

  • exploration candidates must pass every zero-tolerance gate (rights, tenancy, security) — exploration

is over quality uncertainty, never over safety uncertainty;

  • never on regulated workloads (health, finance, legal, HR);
  • bounded fraction of builds, declared in advance;
  • always with the human admission decision retained.

Invariant I-P15-4: exploration is permitted over quality, never over safety. This is what makes having any exploration budget at all defensible.

8. Assumptions

IDAssumptionStateFalsifier
AP15-1Production outcomes predict future reuse quality better than metadatahypothesisOutcome-ranked shelf performs no better than metadata-ranked over N builds
AP15-2Tens of observations suffice for useful rankinghypothesisPosterior remains too diffuse to separate candidates after N builds
AP15-3Industry/archetype priors transfer to new clientshypothesisCold-start priors perform no better than uniform
AP15-4Structural facts suffice; content is never neededhypothesisUseful signal proves inseparable from client content
AP15-5Clients will permit anonymised cross-client aggregationunknownContracts routinely forbid it, collapsing the cross-client premise
AP15-6Adaptation cost is measurable consistentlyhypothesisHours-by-category proves unrecordable in practice
AP15-7Min-gating produces usable rankings, not universal floorshypothesisEvery asset pins to the same low tier, making the ranking uninformative

AP15-7 deserves attention: min-gating is correct but could be degenerate if every asset has one weak family. If so, the ranking's job changes from ordering assets to naming the missing evidence, which is arguably more useful at this stage anyway.

9. Contradictions

X-P15-1 — A flagship contextual-bandit service in this category was retired. Azure AI Personalizer appears in the census and its retirement (status flagged for verification in wave-2) is a negative signal about the durability of buying this capability. The implication is to own the mechanism and treat vendors as replaceable, not to conclude the technique is unsound.

X-P15-2 — The cross-client learning premise may be contractually unavailable. P15's value proposition assumes learning aggregates across clients. If client contracts forbid it, P15 degrades to per-tenant learning, where the sparse-data problem becomes far worse (tens of observations become ones). This is the largest commercial risk in the part and it is unknown, not merely unmeasured.

X-P15-3 — Ranking requires attribution that does not exist yet. Demotion on incident requires knowing which capability caused the incident, which requires the stable capability identifier and instrumentation contract that P14 must impose. P15's most valuable signal is blocked on a P14 deliverable.

X-P15-4 — Corpus refresh cannot be designed against a corpus with no path. A20/A21 remain unknown. Refresh cadence must therefore be specified for the assets actually in hand.

X-P15-5 — The external census is methodologically the wrong shape. The strongest documented- methodology analogues are registry-health scorers (OpenSSF Scorecard, ecosyste.ms, SourceRank, Renovate merge-confidence), not recommenders — because they score software assets on evidence rather than users on preferences. The recommender literature contributes cold-start technique; the registry-health literature contributes the actual scoring shape.

10. Top 10 design hypotheses

Full register: innovation-register.jsonl (100 entries, 10 ranked).

  1. Min-gated ranking, never weighted average — imported from the anti-averaging rule.
  2. Rank by observed adaptation cost, not popularity — attacks A22 directly.
  3. Sparse-regime estimator selection — honest at n=10 or disqualified.
  4. Cold-start priors from industry archetype — because cold start is the only current case.
  5. Falsified-belief records as a first-class output — local idea, absent externally.
  6. Explicit exploration budget — without it the shelf is a ratchet.
  7. Exploration only within a safety envelope — quality uncertainty, never safety.
  8. Client edit volume as a fit signal — P13's log is P15's richest input at zero extra cost.
  9. UNDERDETERMINED aggregates drive contract evolution — defects belong to the contract layer.
  10. Cross-client aggregation is opt-in and contractual — precondition, not feature.

11. Falsifiers

IDClaimFalsifier
F-P15-OUTCOMEOutcome ranking beats metadata rankingNo measurable improvement over N builds
F-P15-SPARSETens of observations sufficePosteriors too diffuse to separate candidates
F-P15-PRIORArchetype priors transferCold-start priors no better than uniform
F-P15-MINGATEMin-gating is informativeAll assets pin to the same tier
F-P15-PRIVACYStructural facts sufficeUseful signal inseparable from client content
F-P15-CONSENTClients permit aggregationContracts routinely forbid it
F-P15-ATTRIBIncidents attribute per capabilityAttribution unachievable (inherits F-P14-ATTRIB)
F-P15-LOOPThe loop improves buildsLater builds show no reduction in adaptation cost or defects

12. Experiments (designed, not authorized)

  • E-P15-1 — Retrospective ranking. Rank existing corpus assets by available evidence under

min-gating; check whether the ranking is degenerate (tests AP15-7) before any live use.

  • E-P15-2 — Adaptation-cost instrumentation. Record hours by category on the first pilot build;

produces the first real point of the A07 distribution.

  • E-P15-3 — Prior transfer. Build archetype priors from the first N builds; test on N+1 against

uniform.

  • E-P15-4 — Signal sufficiency. Determine whether structural facts alone separate good from bad

assets, or whether content is required (tests F-P15-PRIVACY).

  • E-P15-5 — Consent probe. Establish with the client owner whether anonymised cross-client

aggregation is contractually available. This is a conversation, not code, and it gates the entire cross-client premise. Cheapest high-value experiment in the part.

  • E-P15-6 — UNDERDETERMINED harvest. Run the solver over existing contracts; aggregate missing

fields into a prioritised contract backlog. Runs today with no model and no execution.

13. Decision gates

GateQuestionPassFail
GP15-AIs min-gated ranking informative?Assets separate into usable tiersRanking's job becomes naming missing evidence instead
GP15-BIs cross-client aggregation available?Contractual permission obtainablePer-tenant learning only; re-scope P15's value claim
GP15-CIs adaptation cost recordable?Hours by category captured on a real buildPrimary ranking signal unavailable; fall back to defect counts
GP15-DDoes attribution work?Inherits GP14-CIncident-driven demotion unavailable

14. What this document refuses to claim

  • No learning loop exists. Zero accepted plans, zero client edits, zero incidents, zero

maintenance outcomes, zero admitted blocks. Nothing has been ranked or reranked.

  • No claim that production learning improves outcomes — A37 is a hypothesis and stays one.
  • No claim about the 1.3M/850k corpus; its path remains unknown (A20/A21).
  • No claim that clients will permit cross-client aggregation (X-P15-2 open, and it is the part's

largest commercial risk).

  • The commercial denominator was 5 records at wave-1 and is being rebuilt in wave-2; the

not_applicable determination for a pure "learning-loop vendor" category was not made at wave-1 and is a wave-2 deliverable. The OSS denominator is complete at 97 records.