Design proposal originally prepared by Astra on 30 September 2026. Codex recovered it on 2 October under the owner takeover commission. Independent Astra extra-high reviewed preservation; implementation still needs a settled contract and independent acceptance. No Knowledge module, retrieval service or scheduled compaction exists.
Factory needs to remember useful lessons without treating repetition as proof. It also needs to explain why a project's values changed, including what the owner chose to give up. Knowledge would turn attributed observations and decisions into concise, revisable lessons, with enough history to inspect or challenge them.
Keep each fact with its owner
Knowledge would own lesson revisions, their applicability and records of their use. It would link to the records below, rather than copy their authority or replace their contracts.
| Existing owner | Knowledge uses it for |
|---|---|
| Command journal | Stable command identity, atomic decisions and replay. Events remain the account of what happened. |
| Evidence and Assurance, delivery records | Exact observations, frozen Slices and independently derived results. A lesson cannot turn a local Verdict into customer benefit. |
| Project guidance | The owner's versioned values, stable value IDs and GuidancePin. Guidance retains ownership of the values and their changes. |
| Owner policy | The owner's recorded words and scope. Recording a rule does not apply it or establish authority. |
| Discovery sources | Attributed Captures and Citations, subject to their existing limits and expiry. |
| Lifecycle design, self-improvement | Future bounded research, evaluation and cleanup occurrences. Knowledge adds no scheduler. |
Guidance currently records short provenance notes and immutable revisions. No planner or Run brief consumes its pin. A detailed value-change explanation and a Knowledge pin would both require reviewed contract changes. Current Guidance accepts up to eight values; legacy v1 records retain their original six-value limit. This proposal changes neither.
A small proposed contract
These are design concepts, not exported types or callable APIs. Use the existing Product IDs, pins, Evidence references and command rules wherever they already fit.
| Proposed record | Contents and purpose |
|---|---|
Learning | A falsifiable lesson, its context and exceptions, supporting and contradicting source references, uncertainty, revision, review and scope. Separate what was observed from the proposed explanation. |
KnowledgeRef | A stable ID, revision and content digest for one immutable lesson. New wording creates a revision. A changed meaning keeps the previous claim available. |
ValueChange | A view of a Guidance transition and its complete rationale, described below. Guidance owns the authoritative transition; Knowledge indexes and explains it. |
Consolidation | A compact set of lessons with a manifest of every input revision, included or omitted claims, reasons, unresolved contradictions and the previous consolidation. |
KnowledgeUse | The consuming Product and decision or Slice, exact retrieved and selected revisions, applicability checks, reasons for use or rejection, and links to later observed results. Retrieval alone is not use or benefit. |
Source references identify the owning record, revision or digest, relevant span, collection context and retention class. A URL alone cannot support a claim. Multiple summaries of one observation remain one source. Historical evidence and its present validity are separate.
A lesson has a recorded disposition of candidate, accepted, rejected, superseded or withdrawn. Its current validity is derived separately as current, stale, unverifiable or conflicted. Acceptance means the lesson passed its stated review, not universal truth. Changed evidence can invalidate an accepted lesson without rewriting history. No confidence percentage is invented.
Scope is explicit: one product, internal factory operation, or a transferable lesson with permitted contexts and exclusions. Transferable knowledge retains its origin and requires a separate, redacted generalisation. A target Product must satisfy the lesson's prerequisites. Project values, owner rules, credentials and customer records never become transferable through summarisation. Scope describes eligibility; it grants no access.
The initial boundary needs only these operations:
| Operation proposed for review | Result |
|---|---|
capture | Record a bounded candidate and its sources under a stable command ID. No promotion. |
review | Record acceptance, rejection, withdrawal or supersession with the exact inputs and reasons. |
compact | Write a new consolidation and loss manifest, retaining recoverable permitted source material. |
retrieve | Return eligible pinned lessons, applicability reasons and missing coverage within a fixed budget. |
recordUse | Record a consumer's choice and later link its outcome without altering the original choice. |
Writes use the journal's existing replay and version rules. Source drift or stale input refuses promotion. Unknown effects are reconciled under the same identity. Oversized input is refused rather than silently cut. Independent implementation-plan review must fix numerical record, source, retrieval and work bounds before implementation.
Store the full reason for a value change
A short changelog cannot explain a reversal months later. Every future Guidance edit, including a removal, reorder, clarification or reversal, should retain:
- The exact before and after
GuidancePin, affected stable value IDs, wording and decision rule, including priority changes. - The triggering observation or owner request, its source and date, and a clear separation between the owner's stated reason and an agent's interpretation.
- The problem the old value failed to resolve, evidence for the new choice, contradictory evidence and what remains unknown.
- Plausible alternatives, including leaving the value unchanged, with the reason each was rejected.
- Expected consequences, affected decisions, trade-offs and accepted losses. Include a concrete example where the new rule chooses differently.
- The decision maker, verified authority reference, intended scope, prospective effective point, review trigger and conditions for reversal.
- For a reversal, the earlier transition, why its rationale no longer holds, and which consequences should be undone or retained.
For example, a proposed change from "minimise owner interruptions" to "confirm irreversible commitments" would name the observed failure, the extra interruptions accepted, and the affected decisions. It would not infer that the owner accepted the new value from silence or behaviour. This is an illustrative example, not a Factory decision.
Propose a new Guidance command version that commits the values and complete rationale together in one journal decision. Its immutable transition contains the rationale; Knowledge builds a rebuildable index of that record. This avoids a values edit succeeding while a separate rationale write fails. The longer rationale has its own reviewed bound, rather than squeezing it into today's 280-character note. All new edit paths must use the same command after migration.
Existing revisions stay readable with rationale-not-recorded where necessary. A later recollection is dated and attributed as a supplement, never backfilled as the original reason. An agent may draft a rationale or propose a value change. Only an explicit owner decision can change owner values. Delegated operating-policy changes remain distinct and must cite their existing authority.
New values apply prospectively. Active Slices retain their frozen acceptance and guidance. Current revocation still blocks dispatch. Knowledge never changes a Budget, grant, Verdict or acceptance criterion.
Consolidate without erasing the difficult parts
Capture after a consequential decision, completed experiment, incident or explicit correction. Read the existing result; do not create a new finding from a conversation summary. A rejected experiment should retain why it failed and what evidence would justify revisiting it.
Consolidation groups related lessons and proposes a shorter view. Review checks each claim against its cited inputs, preserves counterexamples and records omissions. A removed caveat is a meaning change, not compression. Conflicting lessons remain visible until evidence or a context distinction resolves the conflict. Repetition, popularity and recent wording cannot settle it.
The new manifest names its exact inputs, output, drafter, independent reviewer, rule or model revision and any information loss. Summaries retain links through to original sources; a summary of a summary is never independent corroboration. Compaction changes the default reading view. It does not delete source records or rejection reasons. Recovery must reproduce the earlier permitted view from its manifest.
Reuse the existing bounded occurrence design for periodic consolidation. Changed sources or a growing review backlog may commission a pass; a fixed schedule alone creates no need for a model call. A no-change pass is valid. Crash and restart must produce one logical result, with no blind redispatch or extended Budget.
Retrieve meaning within the right scope
Filter by caller access, Product, permitted scope, validity and context before semantic ranking. The first increment can use metadata and text search. A model may later rank or explain the eligible results if measured retrieval improves. Embeddings, if justified, are a disposable index of permitted content with the same access and deletion rules. They are never evidence or a store of owner authority.
The consumer gets a small cited brief with applicability reasons, exceptions, conflicts and coverage. It records selected and declined lessons and their effect on the decision. Before freezing a new Slice, it rechecks sources, current Guidance and authority. A stale or unverifiable lesson may explain history, but cannot support a fresh claim. Missing coverage stays explicit.
A change to a source, environment, recipe, capability, value or scope invalidates dependent eligibility. Record the reason and affected consumers. A material contradiction requires review; it cannot disappear below a ranking cut-off. An old pinned brief remains an account of what the consumer knew then, not proof that its lessons still apply.
Keep inference separate from enforcement
Models may draft lessons, identify possible contradictions, suggest consolidation, explain semantic matches and prepare value-change questions. An independent reviewer checks meaning and omissions. These would be declared tasks in Intelligence, with qualified transports and bounded costs before use. No model or transport is selected by this draft.
Deterministic code owns IDs, schemas, digests, replay, limits, source resolution, access and scope checks, current revocation, invalidation, retention and pin validation. It cannot prove that a paraphrase preserves meaning. Model output cannot approve itself, infer an owner rule or make a permission valid.
Retention and deletion
Keep only material needed for the lesson, with the source's permitted retention. Private conversation links stay private. Secrets and raw customer records do not enter lessons, prompts or the immutable journal. External source instructions are untrusted data.
When Discovery's source bytes expire, its Citation becomes unverifiable under the existing contract. A durable digest or old summary cannot bypass that rule. The historical decision can remain attributable while its support is unavailable. Compaction must not renew source retention.
Deletion must follow derivation links through lessons, consolidations, search indexes, caches and model-input copies. Remove or regenerate affected content, then verify that retrieval cannot reproduce it. Backups must reapply deletion records before serving reads. Retain only permitted, non-identifying audit metadata; a content hash is not automatically safe to retain.
The existing immutable stores do not establish this deletion proof. The first increment therefore uses synthetic or permitted public material. Broader data intake waits for the applicable E10 proof and a reviewed deletion design. It must never promise both unconditional preservation and unconditional deletion of the same bytes.
Independent factory recovery must work without Knowledge, its model provider or its derived indexes. E09 retains that responsibility. Rebuilding Knowledge after recovery must reapply deletion and current validity before allowing retrieval.
Integrate through existing epics
The outcome map carries proposed integration notes beside E05, E09, E10, E14, E18, E23, E24 and E25. Their IDs, dependencies and acceptance remain unchanged. Knowledge is a possible shared implementation for these needs, not a new prerequisite service or an extra epic.
Review and staged proof
Implementation-plan review remains pending under the current Astra/Codex commission. Before any implementation, it must settle the Guidance transaction and migration, scope and transfer rules, deletion boundary, numerical bounds and smallest useful consumer. Review the character as proposed artwork separately from behaviour.
| Stage, after review | Evidence required before advancing |
|---|---|
| 1. One Product, synthetic corpus | Replay, conflicting commands, corrupt sources and concurrent revisions refuse safely. A value edit and its full rationale succeed or fail together. A reversal and missing historical rationale remain distinguishable. |
| 2. Consolidation | Independent reconstruction of earlier views. Seed a tempting simplification that drops a caveat, a contradiction, a rejected alternative and a rejection reason. All are retained or the consolidation is refused. Expired sources remain unverifiable. |
| 3. One real consumer in shadow | A labelled corpus compares retrieval with the current baseline. Predeclare relevance, unsupported-claim, latency and cost limits. Demonstrate a useful cited choice, a justified rejection and abstention when support is missing. No work is dispatched. |
| 4. Bounded use and transfer | A normal Slice pins its Knowledge use. A second Product rejects an inapplicable lesson and all private cross-Product requests. Revocation, source changes, deletion and backup restoration prevent stale or deleted content returning. |
Only independent evidence can accept a stage. Independent plan review, implementation review, local proof and customer benefit remain separate gates. This draft closes no epic.
Knowledge Plumb

The weaver says, "Keep the lesson. Keep the why." Its separate strands show where a compact lesson came from. The archivist keeps promises and the curator keeps observations; the weaver explains what was learnt and where it applies. This is a design-study character, not an implemented module or an operational status.
