EACKS — Evidence-Aware Conservative Knowledge Synthesis

English | 简体中文

EACKS stops AI agents from prematurely harmonizing heterogeneous knowledge sources into a single, falsely unified story.

Ask an AI to synthesize two books on the same topic, and it will usually hand back one clean narrative — conflicts smoothed away, conditional claims made unconditional, dependent sources presented as independent. The result looks coherent. That is the problem.

EACKS is a knowledge-synthesis protocol that treats unification as something to be earned, not assumed: it preserves differences, keeps conditions explicit, makes conflicts visible, and refuses to produce a unified result when the sources do not warrant one.

What goes wrong without it

The failure	Without EACKS	With EACKS
Semantic conflation — two systems use the same word for different concepts	Merged into one concept	Eight-type Correspondence judgment (S2); default-to-differentiate
Inferential conflation — two systems reach a similar conclusion by different reasoning	Treated as equivalent	Anti-Compression + Type-Specific Validation (S4)
Conditional collapse — a claim holds only under condition X, another only under Y	Combined into one unconditional claim	Conditional Divergence checked first; conditions preserved (S4)
Evidential conflation — sources that quote each other counted as independent confirmation	Reported as "3 sources support this"	Evidence Type ≠ Evidence Quality; provenance chain enforced (S4)
Uncertainty hidden or faked — the model cannot decide, so it pretends it did	A confident answer with no basis	Three legal outcomes: Accepted / Provisional / Abstained
Sources erased by canonicalization — original concepts disappear into a merged vocabulary	Source concepts gone, no way to verify	Source concepts are never physically deleted; Canonical is an organizational layer only (S5)
Synthesis distortion invisible to the reader	A coherent story with no checkable relation to its sources	S7 reverse reconstruction (S → K̂) + ten-item Semantic Loss Taxonomy; hard vetoes beat soft scores
High-order claims with no chain back to sources	An abstraction that cannot be audited	Mandatory trace: High-Order Claim → Structural Relation → Validated Relation → Claim → Evidence → Source

The core term: Spurious Unification

Spurious unification — the erroneous representation of heterogeneous knowledge systems as more semantically, inferentially, conditionally, or evidentially unified than their source structures warrant.

The failure mechanism that produces it is premature harmonization: the model, aiming for a fluent, coherent, answerable whole, dissolves real differences between systems too early — before checking whether the sources actually support the merge.

Source A: "Freedom means X" ──┐
                             ├──→  One Unified Theory   ←  spurious unification
Source B: "Freedom means Y" ──┘

Actually:
Source A: "Freedom means X" ──┐
                             ├──→  Correspondence + unresolved divergence,
Source B: "Freedom means Y" ──┘     conditional relations, preserved conflicts

EACKS exists to make the bottom path the default.

Who this is for

You are	Your situation	What EACKS does
Knowledge engineer	Building RAG or knowledge-graph pipelines over many sources	Typed object model (Concept / Claim / Evidence / Relation / EvidenceLink) whose provenance survives integration
AI agent developer	Multi-source synthesis in agent workflows	One guard skill per pipeline block (S0–S8): spurious unification, false causality, and silent compression are rejected before they enter the graph
Researcher	Synthesizing literature with conflicting findings	An EFSM workflow with explicit verification gates, a reconstruction gate, and structured human arbitration (S8)
LLM output quality engineer	Governing model-generated synthesis	Every derived claim carries explicit provenance; every higher-order structure must survive reverse reconstruction

Why this protocol

1. Conservativity is the goal, not a constraint. K ⪯ S: the higher-order system is a conservative extension of the source knowledge. New semantic commitments require explicit provenance; derived claims must never masquerade as source claims; evidence strength must never be upgraded.
2. Workflow and truth are separated. An EFSM (Layer 4) orchestrates what to do next — fallback, escalation, abstention — and never decides what is true. The typed knowledge model, the inference & validation engine, and the source & provenance layer each stay in their own layer.
3. Guard behavior is validated, not aspirational. Each of the 8 stage guards was exercised against implanted violations (missing conditions, hidden type upgrades, forced unification) before shipping; the machine-executable workflow (skills/eacks-execution/templates/eacks.workflow.yaml, 10 nodes / 30 edges) is the EFSM made runnable.
4. Shippable pieces. The S7 round-trip engine is pure Python stdlib (zero dependencies) and self-tests in one command. The S2/S5 onto_merger adapter is optional — the lightweight L1 channel runs with zero dependencies, the L2 channel only needs onto_merger + networkit.

Quick start

git clone https://github.com/gootf/eacks.git

# copy the skills into your agent's skills directory
# (e.g. ~/.hermes/skills/ — each directory under skills/ is a self-contained skill)

# self-check the round-trip engine (pure stdlib, no install):
python skills/eacks-execution/scripts/eacks_roundtrip.py
# expect: 正常重建: 1.0 PASS / 方向反转: 0.5 HIGH_RISK / ... / ROUGE-L: 1.0 0.0

Then start with eacks-execution — it defines the 9 block contracts, the transition rules, and the shared object schema, and routes to the stage guards. A synthesis run flows S0 → S1 → S2 → S3 → S4 → S5 → S6 → S7 → Outcome (Accepted / Provisional / Abstained), with fallback, escalation, and abstention edges between stages.

All dependencies ship inside this repo under skills/ — clone, copy, run:

Skill	Purpose
log-decisions	Decision journal for S8 arbitration write-back (built-in)
grounded-citations	Web-source ledger for S0 (MIT)
hypothesis-generation	Candidate discipline for S3/S4 (MIT, K-Dense)
scientific-critical-thinking	Evidence evaluation for S4 (MIT, K-Dense)
ocr-and-documents	Document reading for S0 (MIT)
corpus-knowledge-engineering	Ingestion/decomposition/integration discipline (author-built, standalone repo)

Optional tool: onto_merger (AstraZeneca, Apache-2.0) + networkit for the L2 adapter channel.

What it deliberately does NOT do

• No fact-checking. EACKS preserves provenance; it does not arbitrate which source is right.
• No promise of a more unified result. Unification happens only where evidence, definitions, conditions, and logical structure jointly allow it — sometimes the honest output is less unified than the input.
• No fully-automated pipeline. S8 is a structured, write-back human arbitration channel (human_decided decisions become traceable knowledge assets); some decisions are deliberately kept out of the algorithm's hands.
• No LLM-free reconstruction. S7's narrative channel needs an LLM to restate K̂; when unavailable, it degrades to mechanical structure checks only.

Structure

skills/                    the protocol itself — 9 self-contained agent skills:
  eacks-execution/              Layer-4 orchestrator: 9 block contracts, transition
                                rules, shared object schema, tool wiring
                                (SKILL.md + references/ + templates/ + scripts/)
  eacks-s1-decomposition … eacks-s8-arbitration-template
                                8 stage guards, one per processing block (S1–S8)
docs/protocol.md           the protocol explained: EFSM, block contracts, object
                           model, transition rules (for readers, not just agents)
LICENSE                    MIT
README.md / README.zh-CN.md

License

MIT — see LICENSE.