Hindsight epistemic classifier

HINDSIGHT 4-Network Epistemic Classifier

Overview

Production memory systems benefit from distinguishing what the agent observed from what it believes. HINDSIGHT (Latimer et al., 2025, cited in Ch4) organizes memory into four networks:

• World network — objective facts about external reality. Verifiable by external sources (the production region is us-east-1, the CEO of ACME is X, the API endpoint returned 503).
• Experience network — the agent's own first-person actions. "I called the deploy API at 22:30." "I retrieved 5 documents." First-person, timestamped, agent-as-actor.
• Opinion network — subjective beliefs with confidence scores. "I believe the root cause is X with 0.7 confidence." Inference, not observation.
• Observation network — synthesized entity summaries. "Sarah is the on-call lead this week" derived from the union of {Sarah's calendar, on-call rotation doc, prior incidents}. Distillation, not evidence.

Per the HINDSIGHT paper as quoted in Ch4: "developers and users can see what the agent knows versus what it believes." This skill is the classification layer that makes the distinction queryable.

When to Use

• Audit-grade agents — when a user asks "how do you know that," the response must trace evidence to network
• Regulated environments — opinion must be flagged as opinion, not stated as fact
• Multi-agent systems — Agent A's opinion should not become Agent B's fact via uncritical knowledge sharing
• Debugging hallucinations — if the agent stated X confidently, the network classification tells you whether X is evidence-grounded (World/Experience) or inference (Opinion/Observation)

Phrases: "where did the agent get this", "is this fact or inference", "justify the answer", "trust calibration", "HINDSIGHT", "epistemic status".

When NOT to Use

• One-shot agents that need no justification trail
• Storage-only systems (the classification is for retrieval-time reasoning, not just persistence)
• Pure-retrieval agents that never synthesize — the Observation network is empty, the Opinion network is empty; just use World + Experience

Process

Step	Input	Action	Output	Verification
1	Raw fact text + optional metadata	lib.classify(text, metadata)	EpistemicClass(network, confidence)	network ∈ {World, Experience, Opinion, Observation}; confidence in [0, 1]
2	List of facts	lib.classify_batch(facts)	List of EpistemicClass	counts per network match expected distribution for the source
3	Memory containing classified facts	lib.justify(memory, query)	Provenance chain: which World+Experience facts ground a given Opinion/Observation	every Opinion or Observation traces back to World or Experience
4	Classified memory	lib.network_audit(memory)	Health report — % per network, orphan Opinions (no provenance chain), Experience without timestamps	flags experience_without_timestamp_count > 0

Classification Heuristics

Network	Linguistic signals	Metadata signals
World	declarative third-person, no agent pronoun, verb tense past or present, no confidence hedging	source != agent, external_ref present
Experience	first-person agent pronouns ("I called", "I retrieved"), action verbs, timestamp	source == agent, action_type present
Opinion	hedging ("I believe", "likely", "probably", "appears to"), explicit confidence	confidence < 1.0, inferred_from present
Observation	synthesis language ("based on X and Y", "in summary"), multiple inputs	derived_from has multiple refs

The default classifier is heuristic — production should swap in an LLM classifier with a typed-output schema at this seam.

Rationalizations

Agent rationalization	Documented rebuttal
"All facts are facts — networks are overengineering."	Then when the user asks "how do you know," the agent cannot distinguish "the API returned 503" (World, evidence) from "I believe the database was overloaded" (Opinion, inference). The trust calibration collapses.
"Just track a confidence score; networks are redundant."	Confidence is a property of Opinion. World facts don't have confidence — they have provenance. Confidence-only loses the World/Experience/Opinion distinction.
"Observation can be folded into World."	Observation is synthesized from World; it inherits the synthesis-step's failure modes. If you fold it into World, your "World" silently includes summarizations that may contradict raw evidence. The chapter's worked example: Sarah's role is World; "Sarah is the most reliable on-call" is Observation.
"I'll skip Experience because the agent always logs its actions."	Logging is not classification. Experience is the network that answers "what did I do" vs "what happened in the world." If your logs are merged into World, you've conflated agent-action with external-event.

Red Flags

• Opinion network is empty. Either the agent never reasons (unusual) or opinions are being misclassified as World (more likely; check hedging-language detection).
• Experience network has no timestamps. Replay / forensic reconstruction is broken — fix at lib boundary.
• Observation network references nothing. The provenance chain is broken; observations should reference the World/Experience facts they synthesize from.
• World network has confidence < 1.0 entries. World is supposed to be evidence; if it has confidence, it's Opinion misclassified.

Non-Negotiable Verification

1. Run the benchmark battery. python cli.py benchmark must report:
   • all 4 networks distinguishable from typical training-data examples
   • classify_batch produces deterministic output
   • justify returns a non-empty provenance chain for Opinion/Observation
2. Verify CLI help. Exits 0, prints SKILL.md description.

Security Posture

• Prompt injection. Fact text is untrusted input classified against fixed signals - never executed. The attack shape is epistemic masquerade: adversarial phrasing that dresses an opinion as an objective World fact so downstream agents over-trust it. The provenance chain from justify is the cross-check; use it.
• Data exfiltration. No network calls, no file writes. Facts and their provenance chains may reference sensitive sources; they surface only in the stdout report the caller owns.
• Privilege escalation. No shell invocation, no eval, no dynamic import. Network labels are trust metadata: silently promoting Opinion/Observation to World is the escalation to guard against - the label is advisory and grants no evidentiary standing by itself.

Source Attribution

Distilled from Agentic GraphRAG (O'Reilly, by Anthony Alcaraz and Sam Julien) Ch4 — Epistemic Classification subsection. Primary research: HINDSIGHT (Latimer et al., 2025).