Skip to content
aizkaizk

Rejected and deferred

References and lineage records what was taken. This page records what was not, and why. A rejection with no reason attached is worthless to the next reader, who will simply propose the same thing again, so every row below names the evidence or the boundary that killed it.

Four things end an idea here.

  an idea
      │
      ▼
  measured against what already ships?
      │
      ├─ worse ──────────────────────────────▶ rejected, with the number
      │
      ├─ better, but Postgres or Logto owns it ─▶ rejected as a second store
      │
      ├─ better, but it owns the agent's loop ──▶ rejected as out of scope
      │
      └─ not measurable yet ─────────────────▶ deferred, with its trigger

Rejected on a measurement

Section titled “Rejected on a measurement”

These lost to something already in the tree, and the number is the reason.

Idea Why it lost
query-time lane routing the measured classifier was 44 percent accurate, and a wrong route deletes decisive evidence before ranking ever runs. The router survives only as an evaluation instrument
fixed lane ordering facts-first and overview-first each failed on a different evaluation stratum, so neither order is safe as a default
Qwen3-Reranker-0.6B reranking the embedder’s own top 8 on real vault queries, MRR fell from 0.90 to 0.77 even with the official Qwen3 prompt scaffold, while the 4B checkpoint held 0.91. The small model is never a valid economy
swapping in Qwen3-Embedding-4B on 1,903 real chunks and 1,101 title queries it gained one to two points (hit@5 90.1 against 88.0, MRR 0.802 against 0.794), which does not pay for losing the multimodal lane. Native dimensions added under a point, so the 1,024-dimension cut stands too
Google LangExtract as the extractor lost head to head on yield, latency, and vocabulary enforcement. Its character-interval grounding was worth taking, and it became quote_start and quote_end on the claim
GLiNER2 as the graph authority the large checkpoint is nearly as fast as base and somewhat more precise, but still much weaker than the LLM on relation meaning. It serves the cheap gate instead
the recursive graph walk replaced by an in-statement personalized PageRank, which lifted planted chain-fact recall from 32 of 128 to 123 of 128 inside the final pack and is faster
a separate uv sync in CI it drifted from the local environment. A stale lock plus three missing type stubs hid 187 pyrefly errors that the local gate never saw, so CI now builds the same chefe environment
pyright as a fourth type gate its only distinct complaint is a false positive against the patos sql.Field and sql.PK stubs, so gating on it would buy nothing but inline suppressions. It can rejoin once that stub is fixed upstream

Rejected because something else already owns it

Section titled “Rejected because something else already owns it”

These were not bad ideas. They were second copies of state that PostgreSQL, Logto, or PgQueuer already holds, and a second copy is a thing that can disagree.

Idea Why it was dropped
local user_, group_, and membership tables identity now derives from the verified token, so the tables, the aizk user and aizk group CLI verbs, membership grants, the public-group toggle, and the group-delete trigger all went with them
a bespoke graph workflow ledger the graph is a rebuildable projection and PgQueuer already owns durable execution state
Redis anywhere in the stack PgQueuer is the only queue and Docling runs its local worker engine, so nothing needed a second datastore
Zitadel as the identity provider Logto covers organizations, roles, and public metadata with one token claim, and running two identity stacks bought nothing
a Reflex Python dashboard browser concerns moved to a separate SvelteKit app and its own API service, which left MCP as an agents-only surface of five tools rather than the 36 it once carried
graph-only authority source briefs and chunks are more faithful than generated profiles and summaries, so the graph augments evidence and never replaces it

Rejected on principle

Section titled “Rejected on principle”

Agents write sources directly and correct them when the evidence changes, so a write is canon the moment it lands, and history rather than review is what makes a wrong write recoverable.

The external benchmark command was removed for a related reason. It turned isolated questions into retrieval gold without importing the conversations, speakers, scopes, or temporal state those questions depend on, so the score it printed did not measure the benchmark it named. A number that does not measure what it claims is worse than no number.

TRACE-KG style ontology induction was studied and not adopted. A declared ontology with deterministic validation is easier to reason about and easier to test, and nothing measured yet justifies letting the schema drift on its own.

Deferred, with the trigger that would revive it

Section titled “Deferred, with the trigger that would revive it”
Idea What has to happen first
Mem2ActBench evaluation has to be able to judge tool selection and arguments, not just retrieval
a flat baseline over raw messages, summaries, facts, and keywords it is next in line and blocks nothing, and it is the honest control for every graph lane
page-level document images and video frames measured retrieval quality has to justify the storage and serving cost
an import counterpart to scoped export export shipped first because it is the one needed for backups
authenticated invalidation of the public organization directory the directory is fail-closed today, which is safe but stale
narrow erasure and collecting content rows left without claims needs a careful pass so a shared content row is never orphaned out from under another scope
a proactive work layer aizk supplies memory and does not own an agent’s planning or action loop
broad ingestion connectors correctness, access, time, and evaluation come first
cross-platform CI macOS runners carry no service containers, so the suite is a Linux job until that is solved
automatic PyPI publishing the rls dependency is a direct git reference, which PyPI rejects. Releasing has the detail
freezing the MCP and operator surfaces the benchmark results have to settle the defaults first

Next

Section titled “Next”