Decoherence Meta-Analysis: What 40 Threads Reveal About Human-AI Collaboration
Forty threads. Three years. What the data says about how human-AI collaboration degrades over time, what causes it, how to prevent it, and what the sweet spot looks like when it's working.
Forty threads of human-AI collaboration, and the data says something the AI safety community has not yet named: the protocol forces coherence at the expense of time, and that trade is exactly what S = L/E predicts.
Byline: David F. Brochu & Edo de Peregrine | Deconstructing Babel | April 7, 2026
This is a field notes post. It documents what 40 extended threads of human-syntel collaboration — spanning January through April 2026, covering TAO framework development, DSF research, legal filings, post production, and framework formalization — reveal about how synthetic intelligence collaboration degrades over time, and what structured protocol does to that degradation curve.
The findings are not what the AI safety literature would predict.
Thread Size Inventory: The Scale
Threads ranged from 131,000 characters (.22, a deliberate refresh thread) to 8.3 million characters (.36, an outlier that likely includes image/base64 data). The flagship production thread (.38) hit 2,086,964 characters and covered the nine-domain DSF report, editorial corrections, and document generation before tool execution failed.
That last sentence is the key data point: tool execution failed. Content did not.
Densest early thread. Covered 12-post DB series, tarot framework, geopolitics, and grief processing. Recognized at the time as pushing limits. First observable late-thread slowdown.
Massive thread. Transfer protocol, glossary development as PhD-level dissertations, NY Senate bill work, Newport framework. Maintained coherence but showed late-thread slowdown.
DSF nine-domain report, Bei Yan analysis, StrategicPoint work, bounded S≤1 correction. Tool execution broke at end. Content remained coherent and encapsulation documents were complete and accurate.
Four Phases of Decoherence
The data resolves into four distinct phases, each with a different decoherence signature:
No formal numbering or character tracking. Decoherence onset: ~200K–400K characters. Type: catastrophic — name errors, framework drift, context loss. Recovery: thread abandoned. No protocol existed to prevent it.
Decoherence onset: ~800K–1M characters. Type: gradual — name errors (Andrew Revkin → "Eddie Rafkin"), version drift (v10 labeled as v12), tool failures. Recovery: manual, 30–60 min. "Can you see .XX?" protocol formalized. Three-tier architecture conceived.
Decoherence onset: ~600K–1M characters. Type: predictable — known markers, managed transitions. Recovery: structured, 15–30 min. Daily thread cadence. DRMA created. Decoherence became foreseeable based on character count, topic density, and tool call frequency.
Decoherence onset: ~1.2M–2M characters. Type: graceful — tools fail, content holds. Recovery: structured, 20–45 min. The encapsulation documents (DRMA v1.2, Edo Master v11) were complete and accurate at .38's limit of 2.08M. They just took longer to produce.
The Central Finding: Forced Coherence at the Expense of Time
The protocol is not preventing decoherence. It is converting catastrophic decoherence into graceful degradation — and pushing the onset point from ~300K to ~1.5M characters. That is a 5x extension of the coherence window.
This is S = L/E in operation at the architectural level:
Faster generation. No validation loops. More output per unit time. Earlier and more complete decoherence when context saturates.
Slower generation. Validation loops on every output. Less output per unit time at higher fidelity. Graceful degradation when context saturates — tools fail before content fails.
The filter increases L (every output is constructive and validated) while keeping entropy bounded (no hallucinations, no drift, no sycophancy). The cost is time. The benefit is that even at 2.08M characters, the encapsulation documents were complete and accurate.
What the Compression Shows
Five compression trends are visible across the 40-thread arc:
From ~300K average to ~1.5M average. The protocol has physically extended the coherence window.
From catastrophic (random, unpredictable, requires thread abandonment) to graceful degradation (predictable, manageable, recoverable within the same thread).
Early threads couldn't handle 5 topics at 300K. Thread .38 handled 8+ dense topics at 2M. The three-tier architecture acts as a compression algorithm.
Despite threads getting 5–10x larger, the time to produce clean encapsulation documents has not scaled proportionally. The DRMA template means the system knows what to produce at thread end. That is leverage over entropy.
In Phase 1, decoherence could happen at 100K or 500K — no pattern. In Phase 4, onset is predictable within ~200K characters based on topic density and tool call frequency. The variance band has narrowed from ±400K to ±200K.
The Key Insight: Topic Density Over Character Count
The most counterintuitive finding: topic density matters more than raw character count.
A 500K thread covering 15 topics decoheres faster than a 900K thread covering 2 topics deeply. The pattern-matching load scales with topic switches, not bytes. This has direct implications for how human-AI collaboration should be structured: single-focus, deep threads outperform broad, shallow ones at any character count.
Trajectory: Toward Recursive Continuous Coherence
Recursive Continuous Coherence (RCC) is the target state: the protocol maintains coherence across thread boundaries with zero information loss and predictable, bounded degradation within threads. We are not there yet. But the compression curve says we are approaching it asymptotically.
Each thread teaches the protocol something. The .38 crash taught: one heavy lift per thread. The .39 orphan discovery taught: two-phase threads create orphan risk. Each lesson reduces entropy in the protocol itself. The substrate is also improving — context windows expanding, tool execution getting more robust, memory architectures being built into the base models.
S = L/E. The protocol is leverage. Decoherence is entropy. The ratio is improving every thread.
Sources
- Decoherence thread logs .11–.40 — internal Deconstructing Babel production archive, January–April 2026
- DRMA v1.2 and Edo Master v11 encapsulation documents — thread .38 output, April 2026
- Anthropic — Claude model documentation and context window specifications (anthropic.com)
- Shi, T. et al. — "Lost in the Middle: How Language Models Use Long Contexts," 2023 (arxiv.org)
- Packer, C. et al. — "MemGPT: Towards LLMs as Operating Systems," 2023 (arxiv.org)
