DEPLOYMENT_BEHAVIORS.md

Deployment Behaviors

This repo is one local-first artifact machine with explicit deployment kinds. memory remains the canonical baseline. question, repair, oracle, prompt, and witness now all differ through explicit policy rather than copy alone, though memory, question, and repair remain the most developed.

What Is Actually Distinct Now

Deployment	Room feel now	Selection / weighting	Wear / legibility	Room-loop posture
memory	weathered, layered, composed	baseline lane, mood, age, and featured-return weighting	normal wear	baseline anti-repetition, gap pacing, overlap
question	unresolved, returning, lightly haunted	boosts status=open / unresolved items, recent questions, recent-topic clustering, and explicit topic/status threads	lighter wear keeps questions readable longer	shorter anti-repetition, slightly quicker gaps, plus rare question-chorus layering that can extend into a later echo when a thread persists
repair	practical, recent, useful	boosts recent items, shorter notes, near-term topic recurrence, and explicit work-thread returns	much lighter wear for clarity	shorter gaps, calmer tone bed, less overlap, plus bench-log follow-ons that get lighter and tighter as a practical thread stays active
oracle	sparse, ceremonial, event-like	penalizes brand-new material, favors older absent fragments and featured returns	very light wear	longer gaps, longer pauses, very low overlap
prompt	catalytic, quick, recirculating	boosts recent shorter prompt responses and recent-topic echoes	lighter wear than memory	shortest anti-repetition, quicker gaps, slightly more overlap
witness	settled, contextual, documentary	cools hyper-recent material and favors settled contextual notes	gentler wear than memory	longer anti-repetition, calmer pacing, lower overlap

Metadata In Play

No new large metadata system was added.

Current lightweight fields:

• deployment_kind
• topic_tag
• lifecycle_status

Small compatibility improvement:

• ingest now accepts topic_tag, topic, or category
• ingest now accepts lifecycle_status or status

That is enough for:

• question topic clustering and open/answered behavior
• repair topic recurrence and practical-status bias

Existing artifacts without metadata still work. Empty lifecycle_status still counts as unresolved in question, which keeps older data compatible.

Where The Behavior Lives

Primary backend policy:

• api/engine/deployment_policy.py

Shared pool selection:

• api/engine/pool.py

Room-loop config delivery:

• api/engine/views.py
• api/engine/static/engine/kiosk-room-loop-policy.js
• api/engine/static/engine/kiosk-room-loop.js
• api/engine/static/engine/kiosk-room-loop-playback.js

Participant-facing copy:

• api/engine/static/engine/kiosk-copy.js

Operator awareness:

• api/engine/templates/engine/operator_dashboard.html
• api/engine/static/engine/operator-dashboard.js
• api/engine/api_views.py

Operator metadata editing:

• GET /api/v1/operator/artifacts
• POST /api/v1/operator/artifacts/<id>/metadata

Implemented Policy Summary

memory

Baseline behavior is preserved:

• same shared pool age/cooldown/wear logic remains canonical
• featured returns still matter
• fresh, mid, and worn lanes still shape the room
• normal wear continues to build patina over time

question

Implemented now:

• unresolved lifecycle states are boosted
• answered / resolved states are cooled down
• recent questions are favored, but older open questions can still return
• recent topic_tag history can cluster adjacent questions loosely
• the room loop can carry a short topic/status thread so an unresolved question can return as a deliberate revisit
• some threaded unresolved questions now trigger a rare same-topic layered return, producing a small chorus moment instead of only single-file recurrence
• when that same unresolved topic has already persisted across several recent room events, the chorus can add a later third voice instead of stopping at a simple pair
• wear is lighter than memory so question audio stays more legible
• room-loop anti-repetition is shorter than memory
• room-loop gaps are modestly quicker than memory

Not implemented:

• semantic grouping
• transcript-aware similarity
• explicit moderator workflow for answering / closing questions

repair

Implemented now:

• recent material is strongly favored
• shorter items are boosted, dense/long items are cooled
• near-term topic recurrence is supported with topic_tag
• the room loop can prefer a recent practical thread so related bench notes return while still useful
• some repair threads now trigger a short same-topic follow-on note, so the room can feel like a bench notebook instead of isolated tips
• longer practical threads now force those follow-ons toward lighter-density companions with shorter post-cue breathing room, so the room feels more like an active workbench than a memorial loop
• wear is much lighter than memory to preserve clarity
• room-loop gaps are shorter than memory
• room tone is reduced and overlap chance is lower

Not implemented:

• structured tool / part / issue taxonomies
• repair-specific retention controls

prompt

Implemented now:

• recent prompt responses are favored over old spent material
• shorter prompts are boosted over long dense ones
• recent topic_tag history can echo prompt-adjacent material
• anti-repetition is shorter than memory
• gaps are quicker and overlap is slightly more permissive

Not implemented:

• steward-authored prompt packs
• explicit prompt chains or thread awareness

witness

Implemented now:

• very recent witness notes are cooled until they settle
• settled mid-age material is favored
• longer contextual notes are slightly preferred over clipped fragments
• anti-repetition is longer than memory
• pacing is calmer and overlap is reduced

Not implemented:

• richer witness verification workflow
• witness-specific context fields beyond topic/status

oracle

Implemented now:

• brand-new material is strongly penalized
• older absent material is boosted
• featured-return logic matters more than memory
• wear is very light
• anti-repetition window is long
• room-loop cue gaps and pauses are much longer
• overlap chance is heavily reduced

Not implemented:

• formal ceremony states
• steward-triggered oracle events
• custom visual mode beyond the shared playback surface

Responsiveness Rule

Deployment differences do not change the responsiveness ladder:

1. immediate acknowledgement
2. near-immediate reflection / preview
3. ambient afterlife in the room/archive

The distinct behavior only changes step 3, plus a small amount of truthful copy around what that afterlife means.

Safe Fallbacks

• unknown deployment names still normalize to memory
• if a non-memory deployment has no playable artifacts at all, the pool can still fall back safely
• if a non-memory deployment does have playable artifacts, the selector stays inside that deployment before falling back elsewhere

Operator Editing

/ops/ now includes a small deployment-scoped artifact metadata section:

• only recent artifacts from the active deployment are shown
• only topic_tag and lifecycle_status are editable
• lifecycle_status now uses deployment-specific picker presets, while still preserving older custom values already in the archive
• edits are audited through steward actions

This keeps stewardship inspectable without turning the dashboard into a large artifact management console.

Next Likely Extensions

• deployment-specific room personalities that package tone, movement, and density more explicitly
• room-state transitions driven by recent recording activity, not just playback history