Merge pull request #29 from MachineWisdomAI/docs/sweep-remaining-factual-errors

docs: sweep remaining factual errors across all docs and protocol modules

Author: timeleft--

AGENTS_SETUP_INSTRUCTIONS.md

@@ -22,7 +22,7 @@ git clone https://github.com/MachineWisdomAI/fava-trails.git && cd fava-trails &
 
 ### LLM Configuration (for Trust Gate)
 
-The Trust Gate reviews thoughts before promotion using an LLM. By default, FAVA Trails uses [OpenRouter](https://openrouter.ai/) for unified access to 100+ models.
+The Trust Gate reviews thoughts before promotion using an LLM. By default, FAVA Trails uses [OpenRouter](https://openrouter.ai/) for unified access to 300–500+ models.
 
 **OpenRouter (default, recommended):**
 
@@ -338,7 +338,7 @@ FAVA Trails ships with protocol hook modules that can be enabled via `module:` e
 | Protocol | Install | Description |
 |----------|---------|-------------|
 | **SECOM** | `pip install fava-trails[secom]` | Extractive compression at promote time via LLMLingua-2 ([docs](../src/fava_trails/protocols/secom/README.md)) |
-| **ACE** | included | Playbook-driven reranking and anti-pattern detection (Stanford/SambaNova ACE) |
+| **ACE** | included | Playbook-driven reranking and anti-pattern detection (Stanford, UC Berkeley, and SambaNova ACE) |
 | **RLM** | included | MapReduce orchestration hooks for batch workflows (MIT RLM) |
 
 **Quickest way to add a protocol** — use the CLI setup command:

README.md

@@ -223,7 +223,7 @@ FAVA Trails supports optional **lifecycle protocols** — hook modules that run
 
 ### SECOM — Compression at Promote Time
 
-Extractive token-level compression via [LLMLingua-2](https://github.com/microsoft/LLMLingua), based on [Microsoft's ICLR 2025 SECOM paper](https://arxiv.org/abs/2502.05589). Thoughts are compressed once at promote time (WORM pattern), reducing storage and boosting recall density. Zero hallucination — only original tokens survive.
+Extractive token-level compression via [LLMLingua-2](https://github.com/microsoft/LLMLingua), based on the [SECOM paper](https://arxiv.org/abs/2502.05589) (Tsinghua University and Microsoft, ICLR 2025). Thoughts are compressed once at promote time (WORM pattern), reducing storage and boosting recall density. Purely extractive — only original tokens survive, no paraphrasing or rewriting.
 
 ```bash
 pip install fava-trails[secom]
@@ -267,9 +267,9 @@ fava-trails secom setup --write
 fava-trails secom warmup
 ```
 
-### ACE — Agentic Context Engine
+### ACE — Agentic Context Engineering
 
-Playbook-driven reranking and anti-pattern detection, based on [Stanford/SambaNova ACE (arXiv:2510.04618)](https://arxiv.org/abs/2510.04618). Applies multiplicative scoring using rules stored in the `preferences/` namespace.
+Playbook-driven reranking and anti-pattern detection, based on [ACE (arXiv:2510.04618)](https://arxiv.org/abs/2510.04618) (Stanford, UC Berkeley, and SambaNova, ICLR 2026). Applies multiplicative scoring using rules stored in the `preferences/` namespace.
 
 ```bash
 pip install fava-trails  # included in base install

docs/fava_trails_faq.md

@@ -37,11 +37,11 @@ This is not post-hoc rollback — it is **containment at the source**. The hallu
 
 ### How does FAVA Trails relate to Context Engineering protocols like SECOM or ACE?
 
-Academic protocols like Microsoft's SECOM (Segmentation and Compression, ICLR 2025) or Stanford's ACE (Agentic Context Engine) define *what* to do with context — compress memories, curate retrieval, manage information density — but leave the production substrate unspecified. Where do compressed memories live? How do you version them? What happens when compression fails mid-operation?
+Academic protocols like SECOM (Segmentation and Compression; Tsinghua University and Microsoft, ICLR 2025) or ACE (Agentic Context Engineering; Stanford, UC Berkeley, and SambaNova, ICLR 2026) define *what* to do with context — compress memories, curate retrieval, manage information density — but leave the production substrate unspecified. Where do compressed memories live? How do you version them? What happens when compression fails mid-operation?
 
 FAVA Trails provides the versioned substrate and the **Event-Action Pipeline** (lifecycle hooks) to run these protocols safely. Hooks fire at key lifecycle points (`before_propose`, `before_save`, `on_recall`, etc.) and return typed actions (`Mutate`, `Advise`, `RecallSelect`) that the pipeline executes atomically.
 
-For example, the built-in [SECOM protocol](../src/fava_trails/protocols/secom/README.md) uses a Write-Once, Read-Many (WORM) optimization: the `before_propose` hook compresses content inline via LLMLingua-2 (extractive token-level compression, zero hallucination) before the thought is committed to its permanent namespace. This avoids read-path latency entirely, amortizes compression cost from O(recalls × thoughts) to O(promotes), and preserves the original verbose draft in the Jujutsu commit history. If compression fails, `fail_mode: open` lets the thought through unchanged — the operation never blocks.
+For example, the built-in [SECOM protocol](../src/fava_trails/protocols/secom/README.md) uses a Write-Once, Read-Many (WORM) optimization: the `before_propose` hook compresses content inline via LLMLingua-2 (extractive token-level compression — only original tokens survive) before the thought is committed to its permanent namespace. This avoids read-path latency entirely, amortizes compression cost from O(recalls × thoughts) to O(promotes), and preserves the original verbose draft in the Jujutsu commit history. If compression fails, `fail_mode: open` lets the thought through unchanged — the operation never blocks.
 
 Install with `pip install fava-trails[secom]` and add a `hooks:` entry to your data repo's `config.yaml`. See the [Protocols section](../README.md#protocols) for quick start.
 

src/fava_trails/protocols/__init__.py

@@ -3,15 +3,16 @@
 Each protocol is a standalone hook module implementing a specific
 context engineering pattern from the research literature:
 
-- ace: Agentic Context Engine (Curator Pattern) -- playbook-driven
+- ace: Agentic Context Engineering (Curator Pattern) -- playbook-driven
   recall reranking and quality enforcement
 - secom: SECOM Compression (WORM Pattern) -- extractive compression
   at promote time for information density
 - rlm: RLM MapReduce (Orchestration Pattern) -- parallel mapper
   validation, progress tracking, and deterministic reducer retrieval
 
-Protocols are independent. Users pick ONE via their config.yaml hooks
-section. They are not designed to run simultaneously.
+Protocols are composable. Each ships with a staggered default order
+(ace=10, rlm=15, secom=20) so they execute in a defined sequence when
+combined. Enable any combination via config.yaml hooks section.
 
 Usage::
 

src/fava_trails/protocols/ace/README.md

@@ -1,6 +1,6 @@
 # ACE Playbook Hooks (Curator Pattern)
 
-FAVA Trails' reference implementation of the [ACE (Agentic Context Engine)](https://arxiv.org/abs/2510.04618) Curator pattern (Stanford/SambaNova, ICLR 2026). ACE treats agent context as an evolving **playbook** of structured rules that grow and refine through feedback loops.
+FAVA Trails' reference implementation of the [ACE (Agentic Context Engineering)](https://arxiv.org/abs/2510.04618) Curator pattern (Stanford, UC Berkeley, and SambaNova, ICLR 2026). ACE treats agent context as an evolving **playbook** of structured rules that grow and refine through feedback loops.
 
 ## Architecture: FAVA Trails as the Curator
 
@@ -188,6 +188,6 @@ propose_truth(trail_name="my-trail", thought_id="<ULID>")
 
 ## Literature
 
-- Stanford/SambaNova [arXiv:2510.04618](https://arxiv.org/abs/2510.04618) (ICLR 2026)
+- Stanford, UC Berkeley, and SambaNova [arXiv:2510.04618](https://arxiv.org/abs/2510.04618) (ICLR 2026)
 - ACL 2025 Reflective Memory Management
-- Reference implementations: [ace-agent/ace](https://github.com/ace-agent/ace) (Apache-2.0), [kayba-ai/agentic-context-engine](https://github.com/kayba-ai/agentic-context-engine) (MIT)
+- Official implementation: [ace-agent/ace](https://github.com/ace-agent/ace) (Apache-2.0)

src/fava_trails/protocols/ace/__init__.py

@@ -1,8 +1,8 @@
 """ACE Playbook Hooks (Curator Pattern).
 
-Implements FAVA Trails' adaptation of the ACE (Agentic Context Engine) Curator
+Implements FAVA Trails' adaptation of the ACE (Agentic Context Engineering) Curator
 pattern, based on:
-  Stanford/SambaNova arXiv:2510.04618 (ICLR 2026)
+  Stanford, UC Berkeley, and SambaNova arXiv:2510.04618 (ICLR 2026)
   ACL 2025 Reflective Memory Management
 
 Seven lifecycle hooks provide:

src/fava_trails/protocols/secom/README.md

@@ -2,7 +2,7 @@
 
 Extractive token-level compression at promote time for information density, based on:
 
-> Microsoft ICLR 2025 "On Memory Construction and Retrieval for Personalized Conversational Agents" (arXiv:2502.05589)
+> Tsinghua University and Microsoft, ICLR 2025 "On Memory Construction and Retrieval for Personalized Conversational Agents" (arXiv:2502.05589)
 > Reference implementation: [microsoft/SeCom](https://github.com/microsoft/SeCom)
 
 ## WORM Architecture (Write-Once-Read-Many)
@@ -108,7 +108,7 @@ Unknown engine types fail loudly at configure time. See the [LLMLingua docs](htt
 
 Uses **extractive token-level compression**. For each token, the model predicts keep/discard. Key properties:
 
-- **Zero hallucination**: Only original tokens survive. No paraphrasing, no rewriting.
+- **Purely extractive**: Only original tokens survive. No paraphrasing, no rewriting, no new tokens generated.
 - **Preserves named entities and identifiers**: Token-level decisions maintain factual anchors.
 - **Optimal rate**: tau = 0.5-0.7 (retain 50-70% of tokens). Below 0.5, critical information loss.
 

src/fava_trails/protocols/secom/__init__.py

@@ -1,8 +1,8 @@
 """SECOM Compression Hooks (WORM Pattern).
 
 Implements the SECOM (SEgmentation + COMpression) pattern from:
-  Microsoft ICLR 2025 "On Memory Construction and Retrieval for
-  Personalized Conversational Agents" (arXiv:2502.05589)
+  Tsinghua University and Microsoft, ICLR 2025 "On Memory Construction and
+  Retrieval for Personalized Conversational Agents" (arXiv:2502.05589)
 
 Three lifecycle hooks:
   - before_propose: Inline extractive compression via Mutate(ThoughtPatch)