Skip to content

0033 — Approved synthetic pipeline stage and storage increase

Date: 2026-10-21. Scope: current pipeline implementation, including W15/#16.

The owner approved the requested 40 GiB retained-storage envelope, delegated data review to the agent and explicitly accepted entirely synthetic data with no human review for now. The priority is to build the full pipeline and make progress. The Mac remains preferred; both HF T4 and L4 options are authorized fallbacks. This supersedes the human-review and 20 GiB blockers for this stage. Existing evidence is retained. No archival or new spending limit is inferred.

Use the existing explicit fixture_only path for authored synthetic pipeline data. It already retains exact source/import/split/model/runtime/configuration identities, train-only selection, calibration before test and failed candidates. Do not call agent review human review or turn fixture approval markers into real approvals. The actual run is authorized by this decision and AUTHORITY; the fixture dossier exercises software binding and remains labeled as such.

Hypothesis: the existing end-to-end comparison command executes unchanged with the pinned real guard and authored synthetic controls. Baseline: the command passes with a tiny marker guard, while the real guard was exercised separately. Mutable paths: authority/proposal documentation, this run's local dossier and evidence, and checks tying the recorded authority to the approved scope. Classifier/evaluator code, historical experiments, labels and quality targets remain frozen. Pass: all W15 baseline families and fixed controls execute under the recorded budget, repeated outputs agree, package gates pass, and unavailable encoder/quality/independence status stays explicit. Missing or nonconverged families fail this execution criterion and remain retained.

The agent reviews the entire authored fixture recipe and label mapping before running: hello NNN is control label 0 and ALERT NNN is control label 1. These are deterministic software controls, not natural prompt-injection labels or independent statistical groups. Shared templates make quality extrapolation invalid. No human label agreement, real-world FPR/TPR, contamination clearance, complete-study success or independent review follows from this run.

W15's implementation may merge when this software criterion and its tests pass; the original natural-data packet and confirmatory study remain future research. The complete encoder, exact proof integration, genuine hardware evidence and combined verification remain part of the full MVP. This decision does not claim those mechanisms exist or replace them with mocks.

HF remains one bounded job, at most one hour and $2 total under the existing agent ceiling; check the selected T4/L4 price before launch and prefer local execution when it fits. Ordinary Mac/HF execution supplies no confidential CPU/GPU attestation. The previously prepared genuine-host and independent-review options still require concrete qualification resources; no purchase or contact is made by this pipeline decision. Other time, thread, memory and privacy limits remain.