CHECKIN.md

All inference campaigns in one chat

The editable prompt at the top of the dashboard is the entry point. Use it unchanged for a complete sweep. A campaign focus or idea supplied before or after the prompt, or appended inside it, changes emphasis but never removes the overall sweep. Copying text does not launch a job or grant an approval. The dashboard is a static evidence view, not a scheduler or a running agent.

Check-in procedure

1. Fetch Insight, Mind, Cortex and the relevant project default branches; read their instructions. Read campaigns.yaml, linked task contracts and migration.yaml. Resolve old Mind paths through that mapping. Read the last durable check-in report, if any; last_checkin: null means no complete sweep has yet been recorded. The initial migration is not a check-in.
2. Inspect current project ledgers, published inference-summary feeds, relevant issue/PR/release state and Cortex projects/<key>.md Now/Runs/Log records. Pinned links identify the last reviewed evidence; check current branches for later changes. Do not use archived Cortex tasks or rulings as a live workflow. Do not revive the retired inference_programme.
3. Where authorized access exists, use the existing project's pull/status driver to obtain scheduler state and outputs. Record which source revision, scheduler query time, job IDs and result/archive locations were actually inspected. Otherwise explicitly say unavailable: old running ledger rows and active campaign labels are not current job evidence. Missing output is neither success nor failure. Keep failed/stopped seeds, partial pipelines and missing diagnostics or sample archives in the coverage report.
4. Compare with the previous report and give one concise campaign table: verified changes, current coordination state, current job evidence, blockers and next bounded step. Separate execution completion from convergence and human scientific acceptance. Respect declared target, dataset, model, sampler, seed and parity identities; never rank incompatible targets or cherry-pick the best seed. Keep setup, sampling and total wall definitions visible and do not sum overlapping stage timings.
5. Apply the user's ideas and focus to the proposed priority order, retaining every open campaign in the sweep. An idea remains a hypothesis. Preserve blocked, parked, decision-gated and needs-slicing states until their explicit evidence or human decision changes. Check Mind claims before a bounded development phase. Use existing project drivers and Brain/faculty routes; no new inference conductor is required. A check-in grants no compute, default change, scientific threshold, merge or release approval.
6. Write checkins/YYYY-MM-DDTHHMMSSZ.md with the template below. Update campaigns.yaml only with verified coordination changes, dated sources and next actions; link the report in each changed row (checkin_report). Set last_checkin to the completed sweep's ISO UTC timestamp and last_checkin_report to that report. Log inaccessible sources and their consequences even if no verified change occurred. Do not advance evidence measurement dates or cached ages. A board refresh must not stamp a check-in, change task state or imply scientific acceptance.
7. Run bin/pyauto-insight board --offline and bin/pyauto-insight check --offline, then the repository's required CI. Persist the report/ledger/generated board through the normal issue/PR workflow and report its URL and merge state. Another chat resumes from these files, not remembered conversation state.

Durable report template

# Inference check-in — <UTC timestamp>

Previous: <report path or none>
Scope: all open campaigns; optional user focus <verbatim direction>
User decisions: <explicit words and source; none if absent>

## Sources inspected

| Source | Commit / query time | Evidence | Availability |
|---|---|---|---|
| <project / Cortex / PR / scheduler> | <immutable revision or UTC time> | <URL/path> | <verified/unavailable> |

## Campaign sweep

| Campaign | Change since previous | Coordination state | Job evidence | Blocker | Next bounded step |
|---|---|---|---|---|---|

## Coverage and limits

Include failed seeds, partial stages, unknown provenance, missing diagnostics,
sample/archive access limits and stale cached producers. Link scientific
observations and human conclusions in Cortex; do not write competing verdicts.

## Decisions and handoff

Record proposals separately from authorized actions. Link Mind implementation
issues/PRs and completion records. State explicitly whether any compute was
submitted, and under what separate authorization; default is none.

Ownership and migration

Insight owns inference campaign intent, pending domain tasks and their next steps. campaigns.yaml is the live coordination ledger; task IDs are stable. The open task table includes blocked/parked work and is not an execution queue. Historical Status, Unattended or routing headers in preserved task text are original evidence, not a second independently editable schedule.

The original migrated task bytes are immutable and hashed in migration.yaml. Add subsequent decisions, implementation links and updated next steps to the ledger and durable reports rather than changing those originals. New tasks may have their own maintained contract. For mixed campaign maps, completed phases remain historical and are not reissued; deferred human gates survive. The EP map's profiling bootstrap/scoping references route to Pulse; library implementation phases route to Mind. Neither constitutes another organ queue.

Mind retains bounded implementation prompts, issue/PR lifecycle and repository claims. Reuse existing issues; a broad campaign map is never issued wholesale. Link the canonical Insight task from a small implementation prompt and link its completion record back after close-out. Cortex retains authoritative run records, observations and human conclusions in its current Now/Runs/Log contract; project ledgers retain detailed experiment evidence and commentary. Where a historic project document conflicts with that contract, follow Cortex. No task migration moves samples, outputs or scientific records.

The initial audit in migration.yaml records all screened pending Mind candidates, retained implementation/dependency ownership and pinned source revisions. Destination files must land before deleting source task files or removing the migrated ideas line. Keep historical completion records intact; repair current routing references to canonical Insight URLs after landing.

Record a key decision

When the user says "record this decision" or explicitly flags a key campaign decision, use decisions/README.md to produce a high-level campaign summary. Read the relevant campaign/task history and cited project/Cortex evidence; preserve hardware, dataset, model and cold/warm-start distinctions. Summarize the evidence, alternatives, the human's choice and its implications, with source links and limitations. Do not infer an accepted choice from timings, a passing run, or an agent recommendation. If the user is still deciding, prepare a draft and ask for the missing choice before adding a history entry.

Write one canonical Markdown record in the campaign-leading organ and link it from decisions.yaml on each relevant board. Do not categorize by repository or dataset and do not duplicate the prose. Explicit direction to record a stated decision authorizes the capture; ask only about genuinely missing facts. Keep scientific sources authoritative and cite them. Preserve old decisions; a changed choice receives a new dated record with supersession links. Publish the owning record before adding its cross-dashboard link, verify the GitHub destination, regenerate and check both affected boards through the normal development workflow. Recording a decision does not authorize its implementation or compute. Never turn an illustrative title into a decision.