Reverse compression is when LLM output expands without adding information
Type: kb/types/note.md · Tags: learning-theory, discovery
Compression removes redundancy to preserve information. Reverse-compression is the inverse: expanding a compact signal into verbose prose without adding extractable structure. This is the common failure mode of vibe-noting — a human offers a one-sentence insight, the agent builds a whole article grounded in its training knowledge, and the result reads like depth but teaches the reader nothing beyond the seed.
Epiplexity motivates one part of the measure
What counts as "adding information" depends on who you measure against. Three candidate tests, each more precise than the last:
- Does the body contain claims not in the title? Insufficient — an LLM can generate novel-looking claims by connecting the seed to common knowledge.
- Does the body contain information not derivable from the title combined with the LLM's training knowledge? Well-defined but wrong — it measures novelty relative to the model, not usefulness to the reader.
- Does the body make structure accessible to the reader that wasn't before? The right question — usefulness is relative to the audience, not the source.
Epiplexity formalizes one component of this test: structural information extractable from data by a computationally bounded model under a runtime constraint. It is task-agnostic and does not by itself formalize relevance, surprise, or usefulness relative to an individual reader. Test 3 is this note's operational extension: compare what structure an intended reader can access with and without the artifact, given that reader's prior knowledge and goals. A connection to common knowledge can pass that test when it makes a surprising or previously unavailable pattern accessible. When the connection is already obvious to the audience, the elaboration adds tokens without changing what that reader can extract. A reverse-compressed article is one where the relevant structure was already accessible to its intended readers without the article.
How a linked KB resists reverse-compression
In a linked KB, each link can carry the reader to a node with independently extractable structure — a practitioner report with quantified results, a formal framework, or a prior argument with its own evidence. The network is where accessible epistemic complexity accumulates, not the prose of any single note.
But this resistance requires that links are load-bearing — the linked notes must actually contribute to the argument. A note full of "see also" links to tangentially related material is still reverse-compressed; the links are decorative, not structural. The test: remove the links — does the argument collapse, or does it read identically?
The same failure appears in code generation
SuperARC gives a mechanically classified instance outside KB writing. In its sequence-reproduction code tasks, the report says most programs classified as correct directly printed target sequences. The framework labels direct prints Type 3 and weights non-trivial solutions more heavily, so a correct output need not encode a generative rule. The report also says print-statement solutions dominated correct outputs across programming languages and that temperature variations produced nearly identical no-compression percentages. Those results make the finding robust to the reported language and temperature variations; because the capture omits the named languages, settings, counts, and uncertainty, they do not rule out sampling effects beyond those runs. Calling this reverse-compression is local analysis: the emitted code reproduces the requested output while adding no reusable generative rule for the capability the benchmark scores.
Toward a validation gate
A reverse-compression check is semantic, not structural — it can't be grepped. One heuristic worth testing manually before mechanizing into /validate:
For the intended reader (an agent or human with access to this KB and general training knowledge), does the note's body — including the nodes its links reach — make structure accessible that wasn't before? If the answer is no, the note is reverse-compressed regardless of its length or link count.
Relevant Notes:
- vibe-noting — context: the inflation failure mode this note names and analyzes
- information value is observer-relative — grounds: separates compute-bounded structural extraction from this note's reader-relative relevance test
- Epiplexity paper — source: the formal measure of extractable structure for bounded observers
- SuperARC AIT benchmark — evidenced-by: reports direct-print programs classified as correct and their dominance under the paper's language and temperature variations; classifying this as reverse-compression is local analysis
- linking-theory — enables: the load-bearing vs decorative distinction is a core question for linking theory
- skills derive from methodology — contrasts: skill derivation is the productive inverse — compressing while preserving; reverse-compression is the failure mode — expanding while adding nothing