intelligence-radar/SCOPE.md

# Scope

## Implemented v0.2

- Compendium, research program and historical explorations with explicit provenance.
- Bounded HTTPS source capture, retained content-addressed artifacts, change receipts,
  review-needed observations, deduplication and per-source ingestion health.
- Task-specific five-instrument selection, exact profile/deployment identity,
  constraints, benchmark score normalization, fitness ranking and workload cost scenarios.
- Empirical outcome recording with unique runs, exact-task/profile Wilson intervals,
  provisional sufficiency thresholds and explicit missing evidence.
- Frozen recommendations, deterministic historical replay and Markdown reports.
- Reviewed publication records. An included claim must name a capture already in
  the ledger. A generated draft is not that record. Reviewed publications now
  cover the week-40 through week-43 briefs and the week-44 daily briefs through
  30 October.
- Configured Europe/Berlin daily/weekly job runner, retries, locking, health receipts,
  and user systemd timer units for the local workstation.
- Retention of unarchived ChatGPT and Grok automation briefs from the local drop folder,
  and of Grok mail routed through inbound-library. Website accounts are not polled.
  Mail seen so far is a teaser, not the full brief.
- Behavior tests and isolated ledger/advisor demonstrations. `make check` is the current count.
- Coulomb workplans and Statehub integration. The hub holds work projections, not research evidence.

## Evidence boundary (2026-10-31)

Five catalog sources are captured through 30 October. SWE-bench has not changed.
Reviewed readings, status still reported, cover the 30 October Anthropic pricing
page and OpenRouter catalog membership and per-token fields from 22–30 October.
The pricing page presents Max at $100 a month and Ultra at $200. No Sonnet, Opus,
or Haiku token-price change was found in the visible text. The Sonnet page and the
Sonnet 5 launch note changed a footer sentence only. `frontier/ledger-2026-09.md`
remains the September import. Week 40 publication
`b7a2614df227142c610e1402f2e0f7d84c62ef48bbb1a503af1d072159149d0a` keeps Sonnet 5.5
prices and OpenRouter rows for Sonnet 5.5 and GPT-6.1 Sol. Week 41 publication
`ae08625ddf94184bbbd716be73d5e2debac28948a6fe6817bbb302fc278942c7` keeps the
7 October Haiku 5.5 rates, the Sonnet 5.5 cache-read change from $0.20 to $0.10,
and the OpenRouter rows for Haiku 5.5 and Mistral Large 4. Week 42 publication
`db9e3a7068329a605daee94cc0429706d0aedfaeb42c4f167b88d21483783339` keeps the
OpenRouter rows for Muse Wonder, Gemini 4 Argon, Muse Spark Open, and DeepSeek
V4.1 Flash 1017, and the 18 October Haiku 5.5 and Sonnet 5.5 rates. Week 43 publication
`2181566caba12b0325c604b53168a6c02acaeddcc88cefaf3963d1a46ee60f23` keeps the
OpenRouter row for Kimi K3.1 Pro and the 25 October Opus 5.5, Sonnet 5.5, and
Haiku 5.5 rates. The pricing page lists Max 5x and Max 20x through 25 October
and adds the navigation label Claude Duet on 21 October. Week 44 publication
`43113feeff26834ad3d7585469141fe2c13a4652b7c650500d4e4e23afba1630` keeps the
OpenRouter rows for GPT-6.1 Terra and the 29 October Mistral Large 4 fields,
and the 30 October Max and Ultra plan text. No week-44 weekly brief is retained,
and there is no 31 October capture. The rest of each Grok brief stays excluded.
Price readings are in `frontier/readings-2026-W40.md`, `frontier/readings-2026-W41.md`,
`frontier/readings-2026-W42.md`, `frontier/readings-2026-W43.md`, and
`frontier/readings-2026-W44.md`. A reviewed two-profile comparison is retained, and
it withholds Frontier, Efficient, and Budget winners because requester access and
exact benchmark results are not evidenced.

Grok daily briefs for 1–30 October and weekly briefs for weeks 40–43 are retained
as unverified external digests. The 31 October daily file is empty and still in
the drop folder. ChatGPT has no mail path. Pulling website runs without an operator
remains IR-WP-0009-T01 (wait). A week of full briefs with no operator step remains
IR-WP-0010-T05 (todo). The October files were placed in the drop folder.

The key check in IR-WP-0005 is done. Paid field runs still wait on IR-WP-0006-T03:
the USD 0.023712 reservation has no authoritative provider billing artifact.
IR-WP-0007-T01 (comparable profiles) is todo. T02 and T04 stay wait until that
reconciliation and matched outcomes exist. Real task outcomes are absent. Full
ranking and calibration behavior are demonstrated with isolated synthetic fixtures.

The implementation is a local CLI and curator workflow, not a hosted service or UI.
A budgeted OpenRouter work-sample runner and portable recommendation bundles are
implemented. It does not learn cross-task calibration, extract trustworthy facts
automatically, optimize decomposition, or estimate empirical cost/runtime quantiles.
Published costs are explicit workload scenarios. Artifact hashes prove local
integrity, not truth. Scheduler operation depends on this workstation and its user
service manager. A draft or an external brief becomes a belief, a price, or a
recommendation only after editorial review.

## Generated work-record index

This repository opts into staged-source WORK-RECORDS.md generation through
`.generated-work-records.json` and the isolated `.generated-work-records.mk`
recipe. `make generated-work-records` reuses the State Hub renderer through
the Custodian adapter; `make generated-work-records-check` verifies byte-exact
agreement with staged Markdown. The pre-commit hook uses the isolated recipe,
preserves unstaged index edits by refusing them, and retains the generated-only
commit guard. It does not contact the hub, alter workplan statuses, or grant
deployment authority. This consumes shared tooling; it does not make this
repository the provider of the generator capability.