Files
ietf-draft-analyzer/data/reports/survey-findings.md

72 lines
3.4 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Verified Findings Ledger — IETF AI/Agent Landscape Survey
Status of every headline claim, its evidence, and its reliability.
Corpus: **clean IETF** = `source='ietf'` AND not false-positive = **524 Internet-Drafts**.
Snapshot: data/drafts.db as of 2026-05-23. Verification: deterministic recompute
(Phase 0) + two-model Batch re-rating with Cohen's/weighted κ (Phase 1).
Legend: ✅ verified deterministic · 🟡 LLM-derived, reliable enough to report ·
🔴 LLM-derived, NOT reliable — do not use as a finding.
## Corpus & scope
| Claim | Value | Status |
|---|---|---|
| Clean IETF corpus | 524 (597 raw 73 false-positive) | ✅ |
| All dates ISO-parseable within IETF | 597/597 | ✅ (date problems were ISO/ETSI/ITU, excluded) |
| Temporal span | 2024-01 … 2026-05 | ✅ |
## Temporal trend (replaces stale "36×")
| Claim | Value | Status |
|---|---|---|
| Monthly avg Jun24May25 | 3.7 | ✅ |
| Monthly avg since Jun25 | 38.8 (~10×) | ✅ |
| Peak month | 2026-03 = 106 | ✅ |
| Peak vs typical 2024 month (~3) | ~35× | ✅ (state as peak-vs-baseline, NOT "average growth") |
| Tail (2026-04/05) | provisional (fetch lag) | ✅ caveat required |
## Author / WG structure
| Claim | Value | Status |
|---|---|---|
| Distinct authors | 619 | ✅ |
| Author concentration | top-10 = 10.9% of drafts (LOW) | ✅ |
| Individual (no-WG) submissions | 456/524 (87%) | ✅ — key "pre-standardization" signal |
| Distinct WGs (group_uri) | 28 | ✅ (acronyms unresolved; optional Datatracker backfill) |
## Embedding redundancy (768-d nomic, 100% coverage, 137k pairs)
| Claim | Value | Status |
|---|---|---|
| Cosine mean / median | 0.711 / ~0.711 | ✅ |
| p90 / p99 / max | 0.790 / 0.850 / 1.000 | ✅ |
| Near-duplicate pairs (>0.9) | 125 | ✅ |
| Drafts with ≥1 near-dup | 170 (32.4%) | ✅ — redundancy finding |
| Cosine≈1.0 pairs | individual↔WG-adopted same doc | ✅ legit, not error |
## Category landscape (primary category)
| Claim | Value | Status |
|---|---|---|
| Distribution sums to 524 | yes; 0 uncategorized | ✅ |
| Multi-category drafts | 92.9% carry >1 | ✅ |
| Top categories | Identity/auth 141, A2A 108, NetOps 64 | 🟡 |
| Sparsest (neutral) | Human-agent 5, Other 9, Model-serving 16 | 🟡 |
| **Inter-rater κ (primary category)** | **Sonnet↔Haiku 0.652, Sonnet↔Prod 0.645 (substantial)** | 🟡 reliable enough |
| Confusion = semantic neighbours | A2A↔NetOps, A2A↔Discovery, Identity↔Other | 🟡 report as boundary fuzziness |
## LLM ordinal quality scores — DO NOT report as findings
| Dimension | weighted κ (Sonnet↔Haiku) | Status |
|---|---|---|
| relevance | 0.728 (but 0.234 vs prod) | 🔴 inconsistent |
| maturity | 0.592 | 🟡/🔴 borderline |
| momentum | 0.457 | 🔴 |
| novelty | 0.206 | 🔴 |
| **overlap** | **0.127 (slight = noise)** | 🔴 — invalidates any "overlap N/5" claim |
**Headline methodological finding:** LLM categorical assignment of standards
documents is substantially reproducible (κ≈0.65); LLM *ordinal quality* ratings
(novelty, overlap) are not (κ≈0.130.21). A landscape survey may report the
category distribution; it must NOT rest on the quality scores.
## Artifacts (reproducible)
- scripts/survey-phase0.py → data/reports/survey-phase0.md
- scripts/rerate-intercoder.py → data/rerate/{sonnet,haiku}.jsonl (Batch API, $2.41)
- scripts/survey-kappa.py → data/reports/survey-kappa.md