Rewind Agents: Guardrails and Fix Patterns

Use this page when your orchestration uses Rewind-style agents that capture local context across apps, then plan and act. If you see privacy leaks, wrong app selection, citation mismatches, or answers that flip between runs, follow these checks and jump to the exact WFGY fix pages.

Acceptance targets

ΔS(question, retrieved) ≤ 0.45
Coverage ≥ 0.70 to the intended section or record
λ stays convergent across 3 paraphrases and 2 seeds
E_resonance stays flat on long windows

Open these first

Visual map and recovery
RAG Architecture & Recovery
End to end retrieval knobs
Retrieval Playbook
Why this snippet
Retrieval Traceability
Ordering control
Rerankers
Embedding vs meaning
Embedding ≠ Semantic
Hallucination and chunk edges
Hallucination
Long chains and entropy
Context Drift · Entropy Collapse
Structural collapse and recovery
Logic Collapse
Prompt injection and schema locks
Prompt Injection
Multi agent conflicts
Multi-Agent Problems
Bootstrap and deployment ordering
Bootstrap Ordering · Deployment Deadlock · Pre-Deploy Collapse
Snippet and citation schema
Data Contracts

Typical Rewind breakpoints and the right fix

Context capture is noisy or oversized and raises ΔS
Tighten capture filters and re-score with deterministic reranking.
Open: Retrieval Playbook · Rerankers
Private strings leak from raw screen or clipboard into prompts
Add a redaction prefilter and a contract gate before the LLM step.
Open: Data Contracts · Prompt Injection
High similarity yet wrong meaning after capture
Mixed embedding functions or metric mismatch between capture and store.
Open: Embedding ≠ Semantic
Wrong app gets chosen in cross app routing
Split the query into intent vs retrieval and lock a two stage rerank.
Open: Query Parsing Split · Rerankers
Citations do not line up because DOM based capture differs from visible text
Require cite then explain with snippet_id, section_id, offsets.
Open: Retrieval Traceability · Data Contracts
Agent handoff loops or shared memory overwrite between apps
Split namespaces per app and stamp mem_rev and mem_hash.
Open: Multi-Agent Problems · role drift · memory desync
Cold boot errors when capture begins before indexes are ready
Guard with warm up checks and backoff.
Open: Bootstrap Ordering · Pre-Deploy Collapse

Fix in 60 seconds

Measure ΔS
Compute ΔS(question, retrieved) and ΔS(retrieved, expected anchor).
Stable < 0.40, transitional 0.40 to 0.60, risk ≥ 0.60.
Probe λ_observe
Do a k sweep and reorder headers. If λ flips on paraphrases, lock the schema and clamp variance with BBAM.
Apply the module

Retrieval drift → BBMC plus Data Contracts
Reasoning collapse → BBCR bridge plus BBAM, verify with Logic Collapse
Hallucination re entry after correction → Pattern: Hallucination Re-entry

Verify
Coverage ≥ 0.70. ΔS ≤ 0.45. Three paraphrases and two seeds with λ convergent.

Minimal Rewind pattern with WFGY checks

# Pseudocode. Show only the control points that matter.

CAPTURE_FIELDS = ["app", "window_title", "text", "dom_path", "timestamp"]
SNIPPET_FIELDS = ["snippet_id", "section_id", "source_url", "offsets", "tokens"]

def capture_context(apps, budget_chars=8000):
    # per-app capture with privacy filters and dedupe
    raw = []
    for app in apps:
        raw.extend(capture_from(app, fields=CAPTURE_FIELDS))
    return redact_and_truncate(raw, budget=budget_chars)

def build_candidates(raw):
    # convert capture into retrievable snippets with a unified analyzer and metric
    return chunk_and_embed(raw, fields=SNIPPET_FIELDS)

def route_intent(question, candidates):
    # two stage: intent selection then deterministic rerank
    intent = detect_intent(question, candidates)
    ordered = rerank(intent, candidates)
    return ordered[:10]

def assemble_prompt(snippets, question):
    # schema-locked prompt with cite then explain
    return prompt.format(context=snippets, question=question)

def wfgy_gate(q, context, answer):
    m = metrics_and_trace(q, context, answer)
    if m["ΔS"] >= 0.60 or m["λ_state"] == "divergent":
        raise RuntimeError("WFGY gate: high ΔS or divergent λ")
    return m

def run_rewind_agent(question):
    raw = capture_context(apps=["browser","docs","mail"])
    candidates = build_candidates(raw)
    topk = route_intent(question, candidates)
    msg = assemble_prompt(topk, question)
    result = agent.invoke(msg)  # the agent must respect strict JSON for tools
    metrics = wfgy_gate(question, topk, result)
    return {"answer": result, "metrics": metrics}

What this enforces

Capture is filtered and budgeted before retrieval. Privacy redaction happens first.
Retrieval uses a unified analyzer and metric. Deterministic reranking controls ordering.
Prompt is schema locked with cite first, then answer.
A post generation WFGY gate can halt the run when ΔS is high or λ flips.
Traces record snippet to citation mapping for audits.

Specs and recipes RAG Architecture & Recovery · Retrieval Playbook · Retrieval Traceability · Data Contracts

Rewind-specific gotchas

Capture order changes across windows and breaks reproducibility Stamp capture_rev and sort by app priority before rerank.
Clipboard or screenshot text bypasses redaction rules Force the same redaction pass for every capture source.
PDF or canvas based apps produce different text than visible content Add a DOM or accessible text fallback and record the path in source_url.
Multi account confusion in Gmail, Drive, Notion Add account id to the namespace and to dedupe_key.
Live side effects before citation checks Require successful WFGY gate and idempotency check before any writes.

When to escalate

ΔS remains ≥ 0.60 after capture filters and retrieval fixes Rebuild the index using the checklists and verify with a small gold set. Retrieval Playbook
Identical input yields different answers across sessions Check version skew, capture order, and session state. Pre-Deploy Collapse

🔗 Quick-Start Downloads (60 sec)

Tool	Link	3-Step Setup
WFGY 1.0 PDF	Engine Paper	1️⃣ Download · 2️⃣ Upload to your LLM · 3️⃣ Ask “Answer using WFGY + <your question>”
TXT OS (plain-text OS)	TXTOS.txt	1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly

🧭 Explore More

Module	Description	Link
WFGY Core	WFGY 2.0 engine is live: full symbolic reasoning architecture and math stack	View →
Problem Map 1.0	Initial 16-mode diagnostic and symbolic fix framework	View →
Problem Map 2.0	RAG-focused failure tree, modular fixes, and pipelines	View →
Semantic Clinic Index	Expanded failure catalog: prompt injection, memory bugs, logic drift	View →
Semantic Blueprint	Layer-based symbolic reasoning & semantic modulations	View →
Benchmark vs GPT-5	Stress test GPT-5 with full WFGY reasoning suite	View →
🧙‍♂️ Starter Village 🏡	New here? Lost in symbols? Click here and let the wizard guide you through	Start →