Update the card to the Cascade-edition white paper

Add the Cascade-style title/deck, opening, contents, one-page claim table and three headline results. Rename the sections, reorder the encoder subsections, add explicit limits/what-is-not-claimed and a separate Reproduce section, sync the canonical-paper and repository links, and correct the 35B verdict-set and 93.9% replay recall statements.

CITATION.cff

@@ -15,8 +15,8 @@ authors:
 version: "0.1.0"
 date-released: "2026-09-22"
 license: MIT
-repository-code: "https://github.com/sky-is-green/splinter-memory"
-url: "https://github.com/sky-is-green/splinter-memory"
+repository-code: "https://github.com/sky-is-green/hive"
+url: "https://github.com/sky-is-green/hive"
 keywords:
   - context-management
   - memory-management

README.md

@@ -12,23 +12,75 @@ tags:
 pipeline_tag: sentence-similarity
 ---
 
-# Long-Context Memory Curation Management
-
-**Splinter Memory — a managed-decay context curation architecture for long-horizon LLM conversations.**
-
-This card is the Hugging Face edition of the Splinter Memory white paper. It
-is written for the people who build LLM applications and want the context
-window to *stay useful* as the conversation grows: the memory layer, the
-mechanisms, the measured results, and the code that implements them.
-
-The canonical paper lives with the code
-([sky-is-green/hive/tree/main/splinter](https://github.com/sky-is-green/hive/tree/main/splinter));
-this edition adds the encoder measurement taken on the harness's own traffic
-(§7) and the operational notes a practitioner needs.
-
----
-
-## The problem, in one paragraph
+# Splinter Memory: A Managed-Decay Context Curation Architecture for Long-Horizon LLM Conversations
+
+**A managed-decay context curation layer for long-horizon LLM conversations, with falsifiable predictions and the failures kept.**
+
+**sky-is-green** · Hugging Face edition · November 2026 \
+Canonical paper: [hive/splinter/SPLINTER-WHITE-PAPER.md](https://github.com/sky-is-green/hive/blob/hive_test/splinter/SPLINTER-WHITE-PAPER.md) · Code: [github.com/sky-is-green/hive](https://github.com/sky-is-green/hive)
+
+A model that has forgotten a fact is rarely uncertain. It answers from
+whatever is still in the window with the same confidence as a model that
+remembers. Every long conversation inherits this: the window fills, the
+oldest decisions fall out, and the model keeps talking.
+
+Splinter Memory is a context-curation layer outside the generative model. A
+small fleet of bidirectional encoders scores the stored conversation; a
+membrane deduplicates and detects topic drift; a retention pass applies
+managed decay; and a focal assembler builds a bounded, relevance-ranked
+window. The primary model sees only what matters, at a fixed budget, every
+turn.
+
+This is the Hugging Face edition of the paper. The canonical paper lives with
+the code; this edition adds the encoder measurement taken on the harness's own
+traffic (§8) and the operational notes a practitioner needs. The test corpora
+ship with the repository; the live run bundles are recorded (R), not
+published.
+
+**Contents**
+
+- [The claim in one page](#the-claim-in-one-page)
+- [1. The problem](#1-the-problem)
+- [2. What it does](#2-what-it-does)
+- [3. The measured results](#3-the-measured-results)
+- [4. Use it](#4-use-it)
+- [5. How it works](#5-how-it-works)
+- [6. The harness and pipeline](#6-the-harness-and-pipeline)
+- [7. The evidence, hypothesis by hypothesis](#7-the-evidence-hypothesis-by-hypothesis)
+- [8. The encoder: measured on live traffic](#8-the-encoder-measured-on-live-traffic)
+- [9. Limits and what is not claimed](#9-limits-and-what-is-not-claimed)
+- [10. Reproduce](#10-reproduce)
+- [11. License and community invitation](#11-license-and-community-invitation)
+
+## The claim in one page
+
+| the claim | the measurement |
+|---|---|
+| Bounded context keeps throughput flat | decode 14.5 → 15.5 tps over 308+ turns (+6.7%, P1 PASS) |
+| Curation beats FIFO | 85.1% of fact-retrievable turns on the 27B (P3 PASS); 88.4% at-least-as-good on the 35B-A3B, a strict-win tie |
+| Recall is real; precision is the ceiling | 90.3% deterministic recall (P2 recall PASS), 10.7% precision (P2 precision FAIL, encoder ceiling) |
+| The score | post-run PES 80.0 GREEN vs rolling 12.2 / FIFO 11.6 |
+| The encoder thresholds are the trade | dedup at 0.92: precision 1.00, recall 0.03; a tuned encoder halves fact preservation (32.66% vs 60.51%) |
+| The context is flat; the curation cost is not | 322 → 332 tokens while the store grows 7.6×; 22.8 ms at 212 chunks → 156.8 ms at 1,612 |
+
+Three results carry the paper:
+
+1. **Bounded context is the lever.** The window stays fixed while the store
+   grows, decode throughput stays flat (P1), and facts survive when history
+   contains them (P2 recall, P3).
+2. **Selection is the product; precision is the ceiling.** The small encoder
+   fleet and deterministic membrane beat FIFO (P3) and keep the densest
+   duplicate (P9). P2 precision, P6 and P10 fail; the failures and the
+   threshold trade are kept (§8).
+3. **The curation's answer advantage is generator-dependent.** On the 27B it
+   won 115 turns to 26; on the 35B-A3B it is a tie. What survives for a strong
+   generator is the bounded context, the fidelity and the cost (§8.3).
+
+What is not claimed: that retrieval precision is solved, that curated context
+always wins on answer quality, that P6/P10 work, or that the auditor is
+independent. The evidence and caveats are in §9.
+
+## 1. The problem
 
 A long conversation is not a long document. Facts that mattered twenty turns
 ago fall out of a rolling context window; the model answers from what
@@ -38,7 +90,7 @@ exactly the old fact the user came back for. Keeping everything is worse -
 the prompt grows, the KV cache grows, the cost grows, and the model
 "loses the middle" while the interesting fact sits somewhere in the middle.
 
-## What this does
+## 2. What it does
 
 Splinter Memory keeps a **managed** memory outside the model and hands the
 model a **bounded, relevance-ranked** context every turn:
@@ -65,23 +117,23 @@ answers and the wrong tool for selecting context; measured live, using the
 the hot path, while the 17M-parameter drone did the same job in
 milliseconds.
 
-## The measured results
+## 3. The measured results
 
 | Metric | Naive FIFO (baseline) | Target | Measured |
 |---|---|---|---|
 | Throughput over 500 turns | decays | flat ±10% | **flat** (14.5 → 15.5 tok/s over 308+ turns), 0 OOM events |
-| Retrieval recall (facts the model actually stated) | n/a | ≥75% / ≥90% | **90.3%** live, 93.5% on the isolated-store replay |
-| Retrieval precision (sentence proxy) | near-chance past the window | ≥70% / ≥85% | **10.7%** - *the encoder ceiling* (§7) |
+| Retrieval recall (facts the model actually stated) | n/a | ≥75% / ≥90% | **90.3%** live, 93.9% on the isolated-store replay |
+| Retrieval precision (sentence proxy) | near-chance past the window | ≥70% / ≥85% | **10.7%** - *the encoder ceiling* (§8) |
 | Context sufficiency vs. an equal-size FIFO window | — | ≥80% of turns | **85.1%** of the fact-retrievable turns (4.4:1) |
 | Routing accuracy | — | ≥80% / ≥92% | **100%** |
-| Added per-turn latency | 0 ms | <50 / <30 ms | **≈18 ms** on a small store (3.4 ms assembly + ~15 ms scoring) - **and it grows with the store**: 22.8 ms at 212 chunks, 156.8 ms at 1,612 (§7.4) |
+| Added per-turn latency | 0 ms | <50 / <30 ms | **≈18 ms** on a small store (3.4 ms assembly + ~15 ms scoring) - **and it grows with the store**: 22.8 ms at 212 chunks, 156.8 ms at 1,612 (§8.4) |
 | Peak RSS over 500 turns | — | small | **34.7 MB** |
 | Pipeline Efficiency Score | ~30 | ≥65 / ≥80 | **80.0 (GREEN)** live, 73.1 on the 20-conversation evidence run |
 
 Full protocol, per-hypothesis logic chains and falsification conditions:
 §5 of the canonical paper.
 
-### The whole protocol, refreshed on the 35B-A3B
+### 3.1 The whole protocol, refreshed on the 35B-A3B
 
 The P1-P11 protocol re-run live against the **35B-A3B ternary** (a
 12-conversation fixture subset), to ask whether the verdict set is a
@@ -90,7 +142,7 @@ property of the architecture or of the generator it was measured on:
 | | verdict | the evidence |
 |---|---|---|
 | P1 constant throughput | **PASS** | the throughput does not decay |
-| P2 retrieval precision | **FAIL** | the precision clause (the known encoder ceiling, §7) |
+| P2 retrieval precision | **FAIL** | the precision clause (the known encoder ceiling, §8) |
 | P3 context sufficiency | **PASS** | the curated context against an equal-size FIFO window |
 | P4 domain-dependent decay | **PASS** | code m90 1.8 vs prose m90 1.2 - a gap of 0.6, beyond the 0.2 band |
 | P5 targeted masking | SKIP | a training experiment |
@@ -101,29 +153,29 @@ property of the architecture or of the generator it was measured on:
 | P10 drift reset | **FAIL** | the reset changes the recall by 0.0 points |
 | P11 comb resurrection | **PASS** | the pressure regime: 100% topic-return recall against 20% without the comb |
 
-**PASS 6/11, 2 skipped - the published verdict set, exactly.** The
-published run (2026-08-23/24, on a 27B, on the full corpus) reported
-P1/P3/P4/P8/P9/P11 PASS, P2/P6/P10 FAIL, P5/P7 SKIP. The same set
-reproduces on a different model and a smaller corpus, which says the
-protocol measures the architecture rather than the generator.
+**PASS 6/11, 2 skipped in the 35B re-run.** The 35B run reproduced
+P1/P3/P4/P8/P9/P11 PASS and P2/P6/P10 FAIL; P5/P7 were not re-run there.
+The canonical paper reports P5/P7 PASS from their separate measurements.
+The reproduced subset is the point: the verdict set is a property of the
+architecture, not of the generator.
 
 **One of the failures now has a candidate mechanism.** P10 (the drift
 reset) measures a detector whose threshold was calibrated for a different
-encoder (§7): forced on, it fires on most turns, so "the reset has no
+encoder (§8): forced on, it fires on most turns, so "the reset has no
 effect" may be measuring a detector that fires far too often rather than
 a reset that does not work. (The exact rate was measured on the
-pre-correction corpus and is withdrawn with the rest of §7.) The verdict
+pre-correction corpus and is withdrawn with the rest of §8.) The verdict
 stands as measured; the mechanism is now in question, which is a
 different claim.
 
 **And the one result that does not carry over is P3's margin.** On the
 27B the curated context won 115 turns outright to 26; on the 35B it is a
-tie (§7.3). The *verdict* holds (the curated context is sufficient, and
+tie (§8.3). The *verdict* holds (the curated context is sufficient, and
 at least as good on 88.4% of the turns); the *margin* does not. The
 curation's accuracy advantage is a function of the generator - what
 survives for a strong one is the bounded context and the cost.
 
-## Use it
+## 4. Use it
 
 Three ways, in increasing order of control:
 
@@ -154,7 +206,7 @@ print(result.assembled.content)     # the curated context, bounded
 diagnostics and the self-contained run bundles are in the repository, so the
 claims can be re-measured on your own hardware and your own backend.
 
-## How it works
+## 5. How it works
 
 | Layer | Function | Core mechanisms |
 |---|---|---|
@@ -169,7 +221,7 @@ chunk → Membrane deduplicates and checks the topic → Retention applies decay
 and the remembrance pass → Focal assembles a bounded context → the backend
 generates → the reply is observed back into the store.
 
-### The four postulates
+### 5.1 The four postulates
 
 1. **Context curation** - for a fixed token budget, a relevance-ranked
    selection from the full history dominates a contiguous window of the same
@@ -184,12 +236,12 @@ generates → the reply is observed back into the store.
    the time, which is enough to drive parameter tuning without a human
    labelling pass.
 
-## The harness that runs it, and the pipeline that manages it
+## 6. The harness and pipeline
 
 The memory system is a library; the **HiveBench Studio** is the harness
 that runs it, and it is where the management pipeline lives.
 
-### The management pipeline
+### 6.1 The management pipeline
 
 A long-horizon memory has to be *managed*: every knob that decides what
 the model sees used to live in three different menus, and a wrong setting
@@ -209,7 +261,7 @@ The point is not tidiness. It is that **a setting which is not consulted
 is indistinguishable from a setting which does not work**, and the only
 way to know which one you have is to wire them and then measure.
 
-### The harness itself
+### 6.2 The harness itself
 
 | | |
 |---|---|
@@ -219,7 +271,7 @@ way to know which one you have is to wire them and then measure.
 | the cascade console | every runner knob in one form, with the Global tab's Quality block as its starting values |
 | the harvest | every turn (the query, the reply, the curation that was used) written to a jsonl, off by default - this is the data the encoder is trained on |
 
-### Answer quality reporting and API escalation
+### 6.3 Answer quality reporting and API escalation
 
 The curation's effect is not only felt, it is reported. Every cascade run
 writes a per-item report (the verdict, the detector findings, the repair
@@ -237,13 +289,13 @@ answer, the cost and the reason recorded. The companion paper
 and the measured verdicts; the benchmark and the run reports are
 published next to this card.
 
-## The evidence, hypothesis by hypothesis
+## 7. The evidence, hypothesis by hypothesis
 
 | | Hypothesis | Verdict |
 |---|---|---|
 | P1 | Constant throughput over 500 turns | **PASS** |
 | P2 | Retrieval precision ≥85% at ≥90% recall | **SPLIT** - recall PASS (90.3%), precision FAIL (the encoder ceiling) |
-| P3 | Context sufficiency ≥80% of turns vs. an equal-size FIFO window | **PASS** (85.1% on the 27B); on the 35B-A3B the answer-level result is a tie (88.4% at-least-as-good, 13 strict wins to 17) - see §7.3 |
+| P3 | Context sufficiency ≥80% of turns vs. an equal-size FIFO window | **PASS** (85.1% on the 27B); on the 35B-A3B the answer-level result is a tie (88.4% at-least-as-good, 13 strict wins to 17) - see §8.3 |
 | P4 | Domain-dependent decay curve (code vs. prose separate) | **PASS** (code m90 1.8 vs. prose 1.2) |
 | P5 | Targeted masking | **PASS** |
 | P6 | Confidence escalation | **FAIL** |
@@ -257,7 +309,7 @@ published next to this card.
 The failures are kept in the paper with their evidence. A pre-registered
 protocol that only reports passes is not evidence.
 
-## 7. The encoder: what we measured on our own traffic
+## 8. The encoder: measured on live traffic
 
 > **Correction (2026-10-21).** The corpus behind this section included an
 > ingested knowledge-base attachment: 17 of its 82 files were chunked
@@ -285,7 +337,7 @@ all landed on the same top-K curve - so it concluded the ceiling was
 *data-structural* and that the remedy was "stop looking for a better
 encoder".
 
-We re-measured all three jobs on the harness's **own** traffic - 81
+I re-measured all three jobs on the harness's **own** traffic - 81
 conversations, 6,975 stored chunks, 4,056 duplicate pairs and 6,533
 topic-change cases, labelled lexically so the label is independent of the
 encoder being measured (`experiments/encoder_suite.py`). The result
@@ -303,7 +355,7 @@ not re-derived.
 | drift at the shipped threshold (0.6) | precision 0.08, **recall 0.96** - the topic-change test fires on almost every turn |
 | drift as a *ranking* signal | **withdrawn** - measured on the pre-correction corpus (79% an ingested attachment), see above |
 
-### 7.1 The F1 is the wrong target, and the threshold is not the lever
+### 8.1 The F1 is the wrong target, and the threshold is not the lever
 
 The obvious conclusion - "move the thresholds to the encoder's optimum" -
 is **falsified** by a deterministic replay of the corpus through the real
@@ -344,32 +396,20 @@ and the fact-preservation constraint fixes where on it the system sits.
 The clean replay and the three failed fine-tunes are in `AB-RESULTS.md`
 §2C-quater through §2C-septies.
 
-### 7.4 The curation's cost grows with the conversation it curates
-
-The stability run - 2,000 turns of a growing conversation, the real drone,
-no model (`experiments.stability_500.py`):
-
-| turns | the store | the assembled context | the curation's cost (p50) | the peak RSS |
-|---|---|---|---|---|
-| 250 | 212 chunks | 322 tok | 22.8 ms | 519 MB |
-| 500 | 412 chunks | 320 tok | 37.3 ms | 564 MB |
-| 1,000 | 812 chunks | 335 tok | 89.7 ms | 594 MB |
-| 2,000 | 1,612 chunks | 332 tok | 156.8 ms | 667 MB |
+### 8.2 What did not work, and what shipped
 
-**The context is flat and the cost is not.** The assembled context moves
-by 10 tokens across a 7.6x range of store - the curation's promise holds
-✓✓ - and nothing is OOM-killed over 2,000 turns ✓. But the curation's
-per-turn cost is **linear in the store**: about 7x for 7.6x the chunks,
-because every chunk is scored every turn.
-
-So the memory's trade is explicit, and the published "≈18 ms added per
-turn" is the *small-store* number: the prompt stays bounded (the
-expensive part for the model) while the curation gets more expensive as
-the conversation grows. The two costs trade against each other, and the
-crossover - the point where curating costs more than the tokens it saves
-- is a measurement the system does not yet make.
+- The shipped configuration stays: the stock encoder at the shipped
+  thresholds, because that is the configuration the published evidence was
+  produced with.
+- The next fine-tune weights a false merge above a missed merge
+  (precision-weighted at the operating point) instead of optimising F1 -
+  the asymmetry makes F1 the wrong target - and is measured on a held-out
+  split, because the current one was trained and measured on the same
+  traffic (a fit, not a generalisation).
+- The full sweep, the fine-tune runs and the replay are in
+  `docs/ENCODER-NOTES.md` in [`splinter/`](https://github.com/sky-is-green/hive/tree/hive_test/splinter).
 
-### 7.3 The curation's advantage is generator-dependent
+### 8.3 The curation's advantage is generator-dependent
 
 The paired A/B - the same model, the same turns, the curated context
 against the last-4k-token window of the same history - is the memory
@@ -403,20 +443,33 @@ is 4k tokens of pure recency, and the two arms are equal *at that
 budget*. The honest claim for a strong generator is the bounded context
 and the cost, not the better answers.
 
-### 7.2 What we changed, and what we did not
+### 8.4 The curation's cost grows with the conversation it curates
 
-- The shipped configuration stays: the stock encoder at the shipped
-  thresholds, because that is the configuration the published evidence was
-  produced with.
-- The next fine-tune weights a false merge above a missed merge
-  (precision-weighted at the operating point) instead of optimising F1 -
-  the asymmetry makes F1 the wrong target - and is measured on a held-out
-  split, because the current one was trained and measured on the same
-  traffic (a fit, not a generalisation).
-- The full sweep, the fine-tune runs and the replay are in
-  `docs/ENCODER-NOTES.md` in [`splinter/`](https://github.com/sky-is-green/hive/tree/main/splinter).
+The stability run - 2,000 turns of a growing conversation, the real drone,
+no model (`experiments.stability_500.py`):
+
+| turns | the store | the assembled context | the curation's cost (p50) | the peak RSS |
+|---|---|---|---|---|
+| 250 | 212 chunks | 322 tok | 22.8 ms | 519 MB |
+| 500 | 412 chunks | 320 tok | 37.3 ms | 564 MB |
+| 1,000 | 812 chunks | 335 tok | 89.7 ms | 594 MB |
+| 2,000 | 1,612 chunks | 332 tok | 156.8 ms | 667 MB |
 
-## Limits, honestly
+**The context is flat and the cost is not.** The assembled context moves
+by 10 tokens across a 7.6x range of store - the curation's promise holds - and nothing is OOM-killed over 2,000 turns. But the curation's
+per-turn cost is **linear in the store**: about 7x for 7.6x the chunks,
+because every chunk is scored every turn.
+
+So the memory's trade is explicit, and the published "≈18 ms added per
+turn" is the *small-store* number: the prompt stays bounded (the
+expensive part for the model) while the curation gets more expensive as
+the conversation grows. The two costs trade against each other, and the
+crossover - the point where curating costs more than the tokens it saves
+- is a measurement the system does not yet make.
+
+## 9. Limits and what is not claimed
+
+### 9.1 Limits
 
 - **Precision is the open question.** The context that reaches the model is
   fatter than it needs to be. The recall evidence says the facts the model
@@ -426,7 +479,7 @@ and the cost, not the better answers.
   circularity but does not remove it: the auditor is the same model family
   that produced the answer.
 - **Synthetic-corpus bias.** The long-horizon corpus is generated. The
-  measurement in §7 is on real traffic; the P1-P12 verdicts mostly are not.
+  measurement in §8 is on real traffic; the P1-P12 verdicts mostly are not.
 - **Cross-conversation contamination and hedge-replies** are harness-level
   failure modes that look like architectural failures if you do not
   isolate conversations per store and filter refusals out of the store.
@@ -447,7 +500,26 @@ and the cost, not the better answers.
   keeps what the VRAM cache cannot hold in ordinary RAM, and the storage
   read on the second request falls from 1.4 GB to 146 MB.
 
-## Reproduce
+### 9.2 What is not claimed
+
+- **Not that retrieval precision is solved.** The 10.7% sentence-proxy
+  precision is the encoder ceiling, not a tuning miss.
+- **Not that curated context always wins on answer quality.** On the 35B-A3B
+  the paired answers tie (88.4% at least as good; 13 strict wins to 17); what
+  survives is the bounded context, fidelity and cost.
+- **Not that P6 or P10 work.** Both failed; the mechanisms and the
+  falsification records are kept.
+- **Not that §8 is a generic ranking result.** It is what was measured on
+  this harness's own traffic; the canonical ranking ceiling still applies.
+- **Not that the auditor is independent.** P7 is a single-rater, same-family
+  agreement measurement, not ground truth.
+- **Not that P12 is measured.** It is a registered draft.
+- **Not that the results generalise.** The P1-P12 verdicts are mostly
+  synthetic corpora; §8 is the real-traffic exception.
+
+---
+
+## 10. Reproduce
 
 ```bash
 # the offline suite (deterministic diagnostics, no model needed)
@@ -466,7 +538,7 @@ open http://127.0.0.1:8765/cascade
 The repository ships the run bundles, the event logs and the label files, so
 a verdict can be traced back to the bytes that produced it.
 
-### The artifacts
+### 10.1 The artifacts
 
 | artifact | what it is |
 |---|---|
@@ -479,7 +551,7 @@ a verdict can be traced back to the bytes that produced it.
 The releases are private while the results are being refreshed; they open
 when the refreshed numbers are in.
 
-## License and citation
+## 11. License and community invitation
 
 MIT. See `LICENSE` and `CITATION.cff` in this repository.
 
@@ -489,7 +561,7 @@ MIT. See `LICENSE` and `CITATION.cff` in this repository.
             for Long-Horizon LLM Conversations},
   author = {sky-is-green},
   year   = {2026},
-  url    = {https://github.com/sky-is-green/hive/tree/main/splinter}
+  url    = {https://github.com/sky-is-green/hive}
 }
 ```