[RAGE-1][P0] Freeze downstream observation, provenance, scope, and redaction records #47

New issue

Open

opened 2026-09-20 01:40:02 +00:00 by nsaspy · 0 comments

nsaspy commented

2026-09-20 01:40:02 +00:00

Owner

Parent: #6
Upstream runtime: lost-rob0t/prolog-rlm#167

Goal

Define the product-owned closed-data records that RAGE uses to observe operator work without turning raw lifetime transcripts into model context.

This is the downstream evidence boundary for AgentProlog. Generic skill runtime/eval/lifecycle state remains upstream in prolog-rlm.

Slice

Define typed records for:

source -> observation -> evidence cluster -> inferred pattern

Each record must carry stable identity, source kind, source locator/reference, project/environment scope, timestamps where available, provenance, redaction state, and content fingerprint.

Support bounded source categories including AgentProlog traces/sessions, repository evidence, imported session archives, external agent histories, and explicit user-provided transcripts.

Requirements

closed deterministic serialization suitable for fixtures;
project/global/environment scoping;
provenance survives every downstream transform;
redact/filter secrets and unrelated private content before candidate synthesis/eval;
normalized records contain distilled evidence, not authority or capabilities;
adapters may not inject executable policy;
raw source retention and distilled observation records are separate concepts;
no duplicate generic trace/artifact/lifecycle database.

Acceptance

schema covers at least AgentProlog session evidence + repository evidence;
stable fingerprints are deterministic;
scope/provenance/redaction are explicit and queryable;
secret fixture proves excluded material cannot reach synthesized candidate text;
malformed/stale records fail structurally;
fixtures are usable by later RAGE slices without live providers.

Non-goals

no pattern mining yet;
no model-driven promotion;
no generic skill runtime implementation downstream.

Parent: #6 Upstream runtime: lost-rob0t/prolog-rlm#167 ## Goal Define the product-owned closed-data records that RAGE uses to observe operator work without turning raw lifetime transcripts into model context. This is the downstream evidence boundary for AgentProlog. Generic skill runtime/eval/lifecycle state remains upstream in prolog-rlm. ## Slice Define typed records for: ```text source -> observation -> evidence cluster -> inferred pattern ``` Each record must carry stable identity, source kind, source locator/reference, project/environment scope, timestamps where available, provenance, redaction state, and content fingerprint. Support bounded source categories including AgentProlog traces/sessions, repository evidence, imported session archives, external agent histories, and explicit user-provided transcripts. ## Requirements - closed deterministic serialization suitable for fixtures; - project/global/environment scoping; - provenance survives every downstream transform; - redact/filter secrets and unrelated private content before candidate synthesis/eval; - normalized records contain distilled evidence, not authority or capabilities; - adapters may not inject executable policy; - raw source retention and distilled observation records are separate concepts; - no duplicate generic trace/artifact/lifecycle database. ## Acceptance - [ ] schema covers at least AgentProlog session evidence + repository evidence; - [ ] stable fingerprints are deterministic; - [ ] scope/provenance/redaction are explicit and queryable; - [ ] secret fixture proves excluded material cannot reach synthesized candidate text; - [ ] malformed/stale records fail structurally; - [ ] fixtures are usable by later RAGE slices without live providers. ## Non-goals - no pattern mining yet; - no model-driven promotion; - no generic skill runtime implementation downstream.

nsaspy referenced this issue

2026-09-20 01:40:38 +00:00

[EPIC] Personal skill refinery: observe workflows, synthesize candidates, RAGE-evaluate, and import external skills #6