Preprint · Independent Research

Extractive Residuals After Compaction

Handle catalogs, last-wins story, and session-scoped vault retrieve
Cary Palmer — Independent Researcher
GitHub · Marionette v0.9.245

Read the PDF Fable 5.1 continuity note Receipts and screenshots Download Code Tables

Abstract

Long-running agent chats compact older turns so the next prompt fits a context window. The usual residual is a paid language-model paragraph. We ask a narrower question than “should agents stop summarizing?”: can a bounded extractive handle catalog, plus a last-wins selected story and a session-scoped SQLite FTS vault, serve as a factory-default residual without pretending that summarization is obsolete? A claim-grade 11×4×3×2 factorial (264 trials, $1.23) shows complementarity: the summarizer keeps designed catalog-miss nonces and drops generic policy stems; the catalog keeps stems and drops nonce prose. Hybrid-plus-stems is the only compact residual that keeps both. Dump-and-query uniquely recovers buried prose when the later ask overlaps and peek has already dropped the middle; it does not select. A pre-registered factory-flip gate then passes on Luna and Sol. Factory residual in Marionette v0.9.244 is catalog. v0.9.245 paints last-wins compact receipts and vault-cite chips on the transcript; that is a surface, not a residual algorithm change. Summary and hybrid remain Settings opt-ins. We do not restamp claim_ready on the experimental cells, and we do not claim a universal “stop summarizing.”

Key findings

Complementarity, not a knockout. On the claim-grade merge, fact-bearing residual_recall_round1 is A 42/60, B 60/60, C 54/60. claim_ready=true on that receipt. That ranking is not a default flip.
Dump-and-query is complementary. Vault-on recovers peek-evicted mid-history prose 3/3 on both models; vault-off is a refuse or peek miss. Recap and paraphrase failed catalog+vault until compact wrote a selected story.
miss_plan died on purpose. Empty FTS dumping the compact-time user extract poisoned unrelated asks. Empty FTS is now empty. Paraphrase C is 3/3 because the story is already in the residual.
Factory-flip gate passed on both models. 16 experimental cases, A vs C, 96 trials each (Luna $0.082, Sol $0.830). Factory residual is catalog as of v0.9.244. v0.9.245 then paints last-wins receipts and vault cites. Last-wins and ack hygiene are part of the residual, not a later prompt trick.
An honestly localized limit. We did not loosen the oracle to force summarizer last-wins to 3/3. We did not add experimental cells to live_cases(). Summary stays a Settings opt-in.

Figures

Three planes after compact: residual, vault, and peek sidecar — Figure 1. Three planes after compact. The residual is always in the next prompt. Vault retrieve is lexical. Peek is a dashed fallback.

Topic last-wins: later go-ahead evicts the older don’t-write; the one-word ack never enters — Figure 2. Topic last-wins. Polarity words are stripped before overlap, so the later go-ahead evicts the older don’t-write. The one-word ack never enters the comparison.

Claim-grade round-1 recall bars: A 42/60, B 60/60, C 54/60 — Figure 3. Claim-grade complementarity. Hybrid wins the ranking. Catalog still misses designed nonce prose. Not a default-flip argument.

Abstract

Key findings

Figures

Paper