DRIFT v0.9.1

Tier.LIVE grounded-chat evidence pass — 2026-07-18

Codex Session ID: 019f7278-ee77-7f02-bafd-6eba8bf046d2

• Executed a small, explicitly bounded Tier.LIVE (gpt-5.6-terra) grounded-chat
  capture against the reviewed store through the DRIFT Manual Run notebook
  (Section 7). One question was asked per configured primary source (eight
  total), each running the same retrieve_live_insights and
  answer_question_with_model code paths the live /chat endpoint uses. The
  run read the five reviewed Tier.FINAL Insights (IDs 10, 11, 13, 15, 16) and
  wrote nothing to the database.
• Seven of eight questions were answered with a grounded citation; the eighth
  (PyTorch) was declined outright. Terra grounded each answer only in the
  Insights that actually supported it — TensorRT 11.1 → 16, vLLM v0.25.1 → 13,
  Transformers v5.14.1 → 11, JAX v0.11.0 → 10, NCCL v2.30.7-1 → 15 — and cited
  exactly IDs 10, 11, 13, 15, and 16 across the run, never an ID outside the
  reviewed store and never a fabricated release.
• Verified the retrieve-first boundary directly. Retrieval ranks by vector
  distance with no cutoff, so it surfaces the nearest reviewed Insights even for
  sources with no reviewed Insight (PyTorch, Triton, CUTLASS). For PyTorch,
  Terra cited nothing; for Triton and CUTLASS it answered that no such Insight
  exists and grounded that negative answer only in the specific reviewed
  Insights it had inspected. The notebook records Terra's true grounding rather
  than the /chat UX fallback to the whole retrieval window, so "cite only the
  evidence used" is inspectable.
• Local spend for the eight-question run was approximately $0.076, within the
  configured spend guard (settled ledger $1.50 → $1.58 against the $5 cap); no
  per-attempt reservation exceeded the per-call cap and none was left reserved.

Added

• backend.evidence_archive.archive_chat_capture() and a Section 7c archive
  cell in notebooks/drift_manual_run.ipynb, the retrieve-first analogue of the
  reviewed-capture archive. It preserves each question, the retrieved and
  grounded Insight IDs, citations, structured answers, provider metadata, and
  per-question spend deltas as a scrubbed JSON record with a SHA-256 manifest
  and no-overwrite protection, refusing credential- or review-note-shaped
  metadata. The shared serialize/hash/manifest path is factored out so both
  archivers guarantee byte-exact LF hashing.
• A frozen, Markdown-only notebooks/drift_manual_run.terra.results.ipynb
  results record (no executable cells, operator configuration, provider/budget
  logs, or review notes), a companion to the gpt-5.6-luna and gpt-5.6-sol
  reviewed-capture records covering the retrieve-and-answer path. The
  display-only regression guard now covers all three results notebooks.
• Section 7 of the Manual Run notebook: a preflight-safe, spend-gated Terra
  grounded-chat capture over the reviewed store, documented in
  notebooks/README.md.
• Archived the bounded run's inspectable evidence and SHA-256 integrity manifest
  to assets/evidence/2026-07-18-all-sources-terra.json
  (sha256 b606fe31f22449de0d404446a9249ddd4c2397dc7d258c7c204a177539dec00d).
• Refreshed the tracked README screenshot gallery into a numbered
  Luna → Sol → Terra flow: landing (01-landing), branded API docs
  (02-api-docs), briefing and frozen claim evidence at the gpt-5.6-luna
  dev tier and gpt-5.6-sol final tier (03.1–03.4), the Ask DRIFT box
  (04-ask-drift), and two gpt-5.6-terra grounded answers (05.1 NCCL,
  05.2 TensorRT). Every gallery reference resolves to a checked-in file.
  The two Terra answer frames preserve their capture-time grounding labels;
  the authoritative current reviewed-store IDs and Terra grounding are the
  scrubbed archive and manifest above.

Release boundary

• This is a source, grounded-chat-evidence, and documentation release. It does
  not draft, publish, or retract any Insight: the live Railway store continues
  to serve the same five reviewed Tier.FINAL Insights (IDs 10, 11, 13, 15, 16).
  The tracked README screenshots were refreshed to the current Luna/Sol briefing
  and Terra Ask DRIFT UI. Hosted app verification remains at v0.8.0 pending an
  independent recheck; the deployed Railway /health and Vercel frontend are
  not being described as v0.9.1. The configured spend guard remains
  authoritative over any reported balance.