docs: Design Philosophy — "imperfection by design"

[duanyiqun]

Summary

Adds a dedicated Design Philosophy writeup explaining why CogniFold deliberately models memory as situated, lossy, and biased rather than chasing a perfect, omniscient, unbiased ground-truth recall store. The core thesis: bias is the mechanism that makes proactivity possible, not a defect — so we optimize for useful proactive structure, not maximal ground-truth fidelity. The flaw is the point.

What's in it

• docs/PHILOSOPHY.md (new, standalone) — the full argument:
  • The thesis + why "a faithful memory" (a lossless index) is the wrong target.
  • Four cognitive realities modeled on purpose, each as phenomenon → agent analogue → CogniFold's stance → concrete mechanism:
    1. Situated cognition → intent-conditioned retrieval, context bands, intent-routed edge weights.
    2. Confirmation bias / reasoning inertia → accumulation/compression/decay/completion as graph rewrites.
    3. Locality of working memory → bounded hierarchical context window, not full-graph dumps.
    4. Metacognitive blind spots (unknown unknowns) → intent crystallization from topology → proactive surfacing.
  • A mapping table (commitment → mechanism), an evaluation section, and a "what we explicitly reject" section.
• README — a condensed Design Philosophy section near the top (after Highlights), plus nav + TOC links to the full doc.

Why this matters for our numbers

The evaluation section makes explicit why we report the proactive-substrate stack, not per-benchmark tuned ceilings: older benchmarks (ToMi especially) are trivially inflatable with a task-specialized reader, but that path rewards auto-loop hallucination — the model confabulating to satisfy a metric instead of reading memory, which is exactly the metacognitive-blind-spot failure the doc describes. This is the conceptual backing for the lowered numbers in the separate benchmark-sync PR (#29).

Notes

• @wsl2000 (Suli) — requested as reviewer. Please review; I won't self-merge this.
• This is intentionally a separate PR from the benchmark-number sync (docs: sync BENCHMARK.md + README to paper (arXiv 2605.13438v3) #29) so the philosophy can be reviewed on its own.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HCToJStvPhZbWGY2ZcBfy7

---

Generated by Claude Code

[wsuli615]

Clear, well-argued philosophy doc whose cited mechanisms actually check out against the code.