Operating practices

Outcome

A reader understands five practices that apply across every stage of the framework, owned by none of them. Each is a piece of how a run is actually operated day to day. One covers how an orchestrator dispatches and recovers. One covers how a run’s own progress is logged and rendered. One covers how a claim or a citation is traced back to where it came from. One covers how untrusted text is kept from being read as an instruction. One covers how work is handed from one session, or one agent, to another.

Where it fits

None of the five pages this page links to belongs to the five-stage framework or to the parallel certification-alignment and Delivery workstreams. Each sub-page extends, and does not duplicate, an already-published page. Agent orchestration extends Human roles, gates and batching. Run logging and dashboards is new material with only a light connection to that same page. Provenance and verification extends Stage 1 and Stage 3. Prompt-injection screening extends S1.4c and S4.6. Hand-off documents and sessions extends the glossary’s own hand-off document entry and that same Human roles page. Read a sub-page’s own “Where it fits” section for the exact relationship. This page only names the five and says how rich or thin each one’s own source material turned out to be.

Why this way

A cross-cutting practice, by definition, does not belong to one stage’s own page. Five pages, gathered under one parent here, keep each practice separately linkable and separately sized to its own material. This spares a thin topic from padding out a stage page it does not belong to, and spares a rich topic from being compressed past what its own detail deserves.

Five practices, and how rich each one’s own material is

Agent orchestration patterns — the session-level procedure an orchestrator follows: what it reads before acting, how it dispatches and waits, how it validates and merges a wave’s results, how it fails and recovers. Rich: this page goes considerably deeper than what Human roles, gates and batching already publishes.
Run logging and dashboards — a step log, a run record, and a static dashboard that together turn a run’s own progress into a structured, shareable record. Rich: a real, fully-designed mechanism this guide’s own research found mostly undocumented anywhere else.
Provenance and verification discipline — a citation-key check, a document-derivation-chain concept, and a short list of integrity rules for tracing a claim or a number back to where it came from. Thin: mostly a cross-reference to what Stage 1 and Stage 3 already cover in full, plus a handful of rules those two pages do not.
Prompt-injection screening, as a cross-cutting practice — one rule, already applied in three separate places under three different names, named once as a single principle. Thin: a naming and cross-reference exercise, with no new mechanism of its own.
Hand-off documents and sessions — two related but distinct practices. One is writing a structured record for a fresh session to continue a task. The other is a subagent signalling that one delegated unit of work is finished inside a single run. Mixed: the first sense is a real, deeper mechanism than the already-published pages state; the second is a different, real mechanism under the same word, found separately.

What is documented versus suggested

Every rule and script default on these five pages carries its own basis label, stated on its own page. Two scripts are worth naming here because their own gap is unusual: X-OP-02 (run record check) and X-OP-04 (hand-off completeness check). Both check a mechanism that the underlying material describes in full but never backed with any script at all. X-OP-05 (completion signal check, on agent orchestration patterns) is the same situation: a documented lesson, a suggested script written for this guide to encode it, not a stand-in for a broken real one. X-OP-03 (citation key check) instead replaces short, unpackaged text-processing one-liners with one safe, tested script that does the same bidirectional check.

Artifacts across these five pages

A governing document (the contract), a tracking file, a hand-off file, and a per-item worker output file, all described on Agent orchestration patterns.
A step log, a run record, and a dashboard, all described on Run logging and dashboards.
A reference list with evidence annotations, described on Provenance and verification discipline.
Nothing new on Prompt-injection screening; its own artifacts belong to the pages it cross-references.
A hand-off document with five structured fields, described on Hand-off documents and sessions. This is distinct from the tracking file named on the orchestration page above. It is the same practice as that page’s own hand-off file; the two pages cross-reference one artifact, not two.

First actions for a new team

Suggested:

Read Agent orchestration patterns and Run logging and dashboards first if your own team is already running parallel agents; both describe real, adoptable mechanisms in depth.
Read Hand-off documents and sessions before your own team’s first session rotation; the two-sense distinction it draws is easy to blur without a page stating it plainly.
Treat Provenance and verification discipline and Prompt-injection screening as short reads. Neither introduces a large new mechanism. Both point back to fuller material your own team may already be applying without a name for it.

Next: Agent orchestration patterns.