Hiro development journal

Assessing readiness for controlled internet observation

Validated and published Machine-readable JSON

Executive summary

Assessed whether Hiro is ready to begin the proposed controlled internet-observation probation and concluded that it is ready for implementation but not yet safe to start network observation.

The main prerequisite exercise is complete: Hiro finished an end-to-end offline autonomous candidate cycle with frozen evidence, benchmark retention, strict rejection, no promotion, and Stage 6 still disabled.

The live Hiro checkout is not yet a clean pinned baseline, and no dedicated internet-observation adapter, egress policy enforcement, sanitizer, immutable snapshot store, or snapshot-to-evaluation conversion pipeline is currently implemented.

No internet access was enabled and no external request was made during this assessment.

Work completed

Prerequisite verification

Partially ready
  • The first complete offline candidate cycle successfully exercised construction, target and regression tests, public and held-out evaluation, frozen evidence, benchmark-ledger retention, and the no-promotion boundary.
  • The live Hiro HEAD remains c9148f1ec4cf9164e4256cc3ce93efdd6794282a and the Stage 6 DISABLED sentinel remains present.
  • A clean isolated evaluation baseline exists, but it is not a complete replacement for the live checkout because ignored development-cycle modules were absent from that snapshot.
  • The live checkout contains numerous existing modified and untracked files, so the autonomous clean-repository guard would block a routine probation launch from it.

Internet-observation control plane

Not implemented
  • Source inspection found design documentation for separating internet observation from evaluation and promotion, but no dedicated observation implementation.
  • There is no current allowlisted GET/HEAD-only client that enforces domain, request-count, elapsed-time, response-size, content-type, and redirect budgets.
  • There is no probation-specific private-network and local-address denial layer or redirect revalidation path.
  • There is no implemented sanitizer and timestamped immutable snapshot packet format for treating retrieved content as untrusted evidence.

Replay and evaluation integration

Not implemented
  • No pipeline currently converts a real-world observation failure into a reviewable, deterministic evaluation case.
  • No runner currently binds baseline and candidate evaluations to the same frozen web snapshot hash.
  • The existing sandbox evaluator can provide the downstream construction, public and held-out evaluation, benchmark retention, and no-promotion boundary after snapshot-derived cases are safely introduced.
  • The benchmark dashboard already has the autonomous-candidate review surface, but internet-observation provenance and snapshot details still need their own review fields.

Decisions and reasoning

Validation and evidence

CheckStatusResult
Live Git and Stage 6 state passed Confirmed live HEAD c9148f1ec4cf9164e4256cc3ce93efdd6794282a, extensive existing working-tree changes, and the Stage 6 DISABLED sentinel.
Observation-lane implementation search passed Repository search found the previously published architecture but no dedicated network observation, sanitization, immutable snapshot, or snapshot-derived replay implementation.
Hiro journal generation and frontend build passed The timestamped-entry test and production build completed successfully before publication.

Current state

Next steps