mirror of
https://github.com/guillaumemeyer/watermarks-remover.git
synced 2026-08-22 13:11:57 +02:00
feat: publishable benchmark harness — env-ready and runnable
Tracks the research/ arXiv-v1 paper kit (per PR #174's gitignore plan) and closes the environment gaps needed to actually run the 3,500-cell study: - detect_text_watermark.py / multilingual_gen.py: new --torch-dtype (auto|fp32|bf16). bf16 is ~8x faster for opt-1.3b CPU cached decode (0.5 vs 4 s/token on aarch64) and keeps same-config gen/detect consistent. - run_experiments.py: --quality-python (defaults to repo .venv-quality, else MarkLLM venv with a warning); --seeds/--prompts now limit the generate loop instead of only the plan; workers force --torch-dtype bf16; the 'none' control attack is a passthrough (was 'unknown attack'). - evaluate_quality.py: BERTScore on roberta-large (deberta-xlarge-mnli crashes under transformers>=5 in set_truncation_and_padding; WATERMARKS_BERTSCORE_MODEL override). - Makefile: research-check target (documented but missing). - .env.example: MARKLLM_DIR naming (code reads MARKLLM_DIR, not WATERMARKS_MARKLLM_DIR). - research/pins-quality.txt: pinned quality-venv freeze (protocol §7). - tests: quality-python plumbing, CLI subset, bf16 worker flag, none attack, multilingual fake_load signature. Validation: make research-check (78 tests) and make lint green; a restricted end-to-end run (generate -> attack -> detect -> evaluate -> report) passes with real metrics.
This commit is contained in:
+1
-1
@@ -18,7 +18,7 @@ WATERMARKS_SERVER_API_KEY=
|
||||
|
||||
# Optional MarkLLM research harness (host checkouts only; not in the core
|
||||
# image). Same-config-only detection — not a vendor oracle.
|
||||
# WATERMARKS_MARKLLM_DIR=~/MarkLLM
|
||||
# MARKLLM_DIR=~/MarkLLM # read by run_experiments.py / multilingual_gen.py / detect_text_watermark.py
|
||||
# WATERMARKS_MARKLLM_SCHEME=kgw # kgw | synthid
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
@@ -29,6 +29,9 @@
|
||||
!/requirements-dev.txt
|
||||
!/ruff.toml
|
||||
!/SECURITY.md
|
||||
!/research/
|
||||
!/research/**
|
||||
/research/results/
|
||||
|
||||
# Exceptions even inside allowed trees
|
||||
.env
|
||||
|
||||
@@ -1,4 +1,4 @@
|
||||
.PHONY: test lint format lint-fix smoke smoke-synthid bootstrap-synthid docker-synthid-build docker-synthid-help \
|
||||
.PHONY: test research-check lint format lint-fix smoke smoke-synthid bootstrap-synthid docker-synthid-build docker-synthid-help \
|
||||
smoke-ctrlregen bootstrap-ctrlregen docker-ctrlregen-build docker-ctrlregen-help \
|
||||
smoke-markllm bootstrap-markllm docker-markllm-build docker-markllm-help \
|
||||
smoke-markdiffusion bootstrap-markdiffusion docker-markdiffusion-build docker-markdiffusion-help \
|
||||
@@ -12,6 +12,9 @@ PYTHON ?= $(shell if [ -x .venv/bin/python ]; then echo .venv/bin/python; else e
|
||||
test:
|
||||
$(PYTHON) -m pytest
|
||||
|
||||
research-check:
|
||||
$(PYTHON) -m pytest research/tests -q
|
||||
|
||||
lint:
|
||||
$(PYTHON) -m ruff check service tests
|
||||
|
||||
|
||||
@@ -0,0 +1,258 @@
|
||||
# 01 — Experimental Protocol
|
||||
|
||||
*Working title of the study: "How fragile are deployed text watermarks?
|
||||
A measurement study of multi-layer watermark removal under realistic
|
||||
user-side editing."*
|
||||
|
||||
Version: v0.2 (2026-08-18). Scope locked for arXiv v1 (see
|
||||
research/README.md decision log). Owner: Guillaume.
|
||||
|
||||
---
|
||||
|
||||
## 1. Research questions
|
||||
|
||||
- **RQ1 (robustness).** How robust are deployed-class text watermarking
|
||||
schemes (KGW, SynthID-Text, EXP, Unigram, SIR) to realistic user-side
|
||||
editing (paraphrase, translation round-trip, structural rewrite,
|
||||
humanization, Unicode/formatting cleanup), measured with ROC-based
|
||||
detection metrics?
|
||||
- **RQ2 (layering).** Does a *layered* removal pipeline (formatting-layer
|
||||
cleanup **+** statistical rewrite) outperform single-layer baselines at
|
||||
equal text-quality cost? Is the gain additive, or does one layer dominate?
|
||||
- **RQ3 (frontier).** What is the quality–detectability Pareto frontier?
|
||||
I.e., what detection rate can a watermarker keep while text remains
|
||||
usable (PPL/BERTScore within tolerance), under each attack?
|
||||
- **RQ4 (policy, secondary).** The EU AI Act Art. 50 transparency regime
|
||||
(in force 2026-08-02) leans on watermarking. Does the mechanism survive
|
||||
contact with real users? (Feeds §7 of the paper and the ethics/position
|
||||
angle; not required for the core experiment.)
|
||||
|
||||
**Primary claim to defend:** *Under realistic editing, quality-preserving
|
||||
removal collapses TPR@1%FPR of KGW-class watermarks to near chance;
|
||||
SynthID-Text resists token-level substitution but not paraphrase /
|
||||
back-translation; a layered pipeline dominates single layers at equal
|
||||
quality cost.*
|
||||
|
||||
**Anti-claim we must preempt:** "Your detector is misconfigured / your
|
||||
attacks destroy the text / a stronger watermark config would survive."
|
||||
Mitigations in §5.3, §6, §10.
|
||||
|
||||
---
|
||||
|
||||
## 2. Design overview
|
||||
|
||||
Factorial experiment, paired design: the same watermarked document is
|
||||
attacked by every attack condition, and detection is run on every
|
||||
(doc, attack) pair with the same watermark key/config. Unwatermarked
|
||||
control texts run through the same attack+detect pipeline
|
||||
(already supported via `--restamp-control`).
|
||||
|
||||
### v1 factorial (locked 2026-08-18)
|
||||
|
||||
| Grid | Schemes | Lengths | Temp | Langs | Prompts | Seeds | Cells |
|
||||
| --- | --- | --- | --- | --- | --- | --- | --- |
|
||||
| EN core | KGW (γ=.25, δ=1); KGW (γ=.5, δ=2); KGW (γ=.5, δ=4); SynthID-Text (default MarkLLM config); EXP (Gumbel); Unigram; SIR | 100, 300 | 0.7 | en | 25 | 5 | 1,750 |
|
||||
| EN temp axis | 4 core schemes (KGW×3, SynthID) | 300 | 1.0 | en | 25 | 5 | 500 |
|
||||
| EN length axis | 4 core schemes | 500 | 0.7 | en | 25 | 5 | 500 |
|
||||
| Multilingual | KGW (γ=.5, δ=2); SynthID-Text | 300 | 0.7 | de, fr, es | 25 | 5 | 750 |
|
||||
| **Total** | | | | | | | **3,500 cells** |
|
||||
|
||||
→ 3,500 watermarked + 3,500 unwatermarked generations = **7,000
|
||||
texts**; each passed through 8 attack cells (§4) = **56,000 attack
|
||||
outputs**; ~65,000 detection runs (originals + attacks + controls)
|
||||
plus ~2,000 unwatermarked texts for the empirical null.
|
||||
|
||||
Restrictions are deliberate (a naive full cartesian — 7 schemes × 3
|
||||
lengths × 2 temps × 4 langs × 25 × 5 — would be 21,000 cells and
|
||||
infeasible on CPU): the temp axis, length axis, and multilingual grid
|
||||
each vary exactly one factor against the 4 core schemes so the effects
|
||||
are attributable. See §3 for the multilingual generator (model holdout).
|
||||
|
||||
### Deferred (post-v1, not in arXiv v1)
|
||||
|
||||
- Human eval (20 samples × 3 annotators) — journal/ACL requirement.
|
||||
- File/metadata mini-study (C2PA/EXIF strip on 20 self-made files;
|
||||
v0.1 §4.3) — separate mini-experiment, revisit for v2.
|
||||
|
||||
---
|
||||
|
||||
## 3. Generation protocol
|
||||
|
||||
- **EN core:** `facebook/opt-1.3b` via MarkLLM (matches existing
|
||||
harness; CPU-feasible).
|
||||
- **Multilingual (DE/FR/ES):** opt-1.3b is English-only, so these cells
|
||||
use a CPU-feasible multilingual generator — **Qwen/Qwen2.5-1.5B-
|
||||
Instruct** (fallback: 0.5B if too slow) — reported explicitly as a
|
||||
model-holdout factor in the paper (scheme × language, not confounded
|
||||
with the EN core). If the multilingual pilot (§9) shows unacceptable
|
||||
output quality, shrink the grid to 10 prompts per language before
|
||||
committing the full run.
|
||||
- Prompts: extend `benchmarks/corpus/` to **25 factual, neutral EN
|
||||
seeds** (50-90 words, varied domains; keep claims checkable) plus
|
||||
**3×25 translated DE/FR/ES sets**; record all prompts in
|
||||
`research/corpus/` (additive, not a fork).
|
||||
- Decoding: **temperature 0.7, top-p 0.95** (realistic); record
|
||||
`do_sample=True`, fixed seeds 1..5 per (prompt, seed) pair.
|
||||
- Watermark configs: fixed per scheme; copy the MarkLLM config JSONs
|
||||
(KGW: γ/δ/hash_key/f_scheme/window_scheme; SynthID; EXP; Unigram;
|
||||
SIR) into `research/configs/` and **use the same JSON for
|
||||
generation and detection**. Synthetic keys — fine to commit; note it.
|
||||
- Control: same prompts, same seeds, no watermark → empirical null
|
||||
distribution for FPR calibration (§5.2).
|
||||
|
||||
## 4. Attack conditions
|
||||
|
||||
### 4.1 Text-level attacks
|
||||
|
||||
| # | Attack | Implementation | Prior-art anchor |
|
||||
| --- | --- | --- | --- |
|
||||
| A0 | None (control) | — | — |
|
||||
| A1 | Layer A only: Unicode/invisible-char, bidi, tag cleanup | `clean_text.py` (deterministic) | formatting-layer marks (zero-width/steganography class) |
|
||||
| A2 | Paraphrase, single pass | `rewrite_text.py --strength paraphrase --candidates 1 --max-loops 1` | paraphrase attacks in watermark literature |
|
||||
| A3 | Paraphrase, adaptive (early stop on detection, up to 3 loops) | `rewrite_text.py --strength paraphrase --candidates 3 --max-loops 3 --markllm-scheme <scheme>` | same; our eval-loop is the "oracle" version |
|
||||
| A4 | Back-translation round trip EN→DE→EN | `rewrite_text.py --strength backtranslate --lang German` | "Can Watermarks Survive Translation?" (X-SIR, ACL 2024) |
|
||||
| A5 | Structural: outline → regenerate | `rewrite_text.py --strength structural` | summarization/outline attacks |
|
||||
| A6 | Humanize | `rewrite_text.py --strength humanize` | style-transfer attacks |
|
||||
| A7 | Cheap baselines: synonym substitution, random word deletion (5-10%), sentence reorder | `research/scripts/attacks/cheap.py` (to be written — gap 05-A3) | Random Walk / impossibility (ICML 2024) |
|
||||
| A8 | **Full pipeline: A1 → A2/A3** | `clean_text.py` then `rewrite_text.py` | our layered contribution |
|
||||
|
||||
Notes:
|
||||
- CLI names above match the real `rewrite_text.py` flags: `--strength`
|
||||
selects the rewrite type; `--backend` selects the transport
|
||||
(`ollama` / `openai-compatible`); `--candidates` × `--max-loops`
|
||||
control the detection-guided loop.
|
||||
- The Layer B rewrite loop stops early when detection passes — that's an
|
||||
*adaptive* attack (stronger than one-shot). Report both: with
|
||||
early-stop (adaptive, "removal oracle") and forced single-pass.
|
||||
- Record rewrite cost (tokens in/out, wall time, USD) per cell — this is
|
||||
the "practical attack cost" contribution (Table 6).
|
||||
|
||||
### 4.2 Detector-side conditions
|
||||
|
||||
- Detectors: MarkLLM same-config (KGW z-score; SynthID mean/weighted/
|
||||
bayesian; EXP/Unigram/SIR per their configs).
|
||||
- Also run **best-effort detection**: threshold tuned on held-out data
|
||||
(favorable to the watermarker; preempts "weak threshold" criticism).
|
||||
- Explicit **strawman preemption**: run the strongest config we have
|
||||
(KGW γ=0.5 δ=4, long text) as the "best case for watermarking" column.
|
||||
|
||||
## 5. Detection & evaluation protocol
|
||||
|
||||
### 5.1 Metrics (primary)
|
||||
|
||||
- **AUROC** over (watermarked, unwatermarked) score distributions.
|
||||
- **TPR @ FPR ∈ {0.1%, 1%, 10%}** read off the ROC.
|
||||
- Full ROC curves saved per cell (for Figure 2).
|
||||
- Secondary: clear rate (existing bench metric, before-positive →
|
||||
after-negative), mean/median score suppression.
|
||||
- **95% bootstrap CIs (10k resamples) on AUROC and TPR@FPR**; report
|
||||
effect sizes, not just p-values.
|
||||
|
||||
### 5.2 Empirical null
|
||||
|
||||
- ≥1,000 unwatermarked generations (same prompts/seeds/temps), scored by
|
||||
the *same* detector. Do **not** assume z-scores are standard normal for
|
||||
FPR calibration; SynthID tournament scores definitely aren't. Empirical
|
||||
null is mandatory — use all unwatermarked controls.
|
||||
|
||||
### 5.3 Quality metrics
|
||||
|
||||
| Metric | Model/tool | Why |
|
||||
| --- | --- | --- |
|
||||
| PPL | `gpt2-large` (independent of generator opt-1.3b — never score with the generator) | fluency |
|
||||
| BERTScore | `deberta-xlarge-mnli` | semantic preservation |
|
||||
| ROUGE-L | standard | content overlap |
|
||||
| SBERT cosine | `all-MiniLM-L6-v2` | cheap semantic similarity |
|
||||
| Levenshtein distance % | stdlib | edit magnitude |
|
||||
| Length drift %, number/URL survival | already in bench | practical usability |
|
||||
|
||||
**Quality @ detection-collapse**: for each attack, report quality metrics
|
||||
*at the operating point where detection collapses* (the Pareto framing).
|
||||
This is the figure that makes the paper defensible: "removal without
|
||||
destruction." To bound CPU cost, compute full quality on a stratified
|
||||
subset (~2-4k texts covering every scheme × attack × collapse point).
|
||||
|
||||
### 5.4 Statistics
|
||||
|
||||
- 5 seeds/cell, paired design; 95% bootstrap CIs as above.
|
||||
- Detection randomness: detection scores are deterministic given
|
||||
(text, key) for KGW; SynthID scoring may have randomness — seed it.
|
||||
- Multiple-comparison discipline: pre-register the headline cells (the
|
||||
v1 matrix); treat nothing else as confirmatory.
|
||||
|
||||
## 6. Budget & compute (estimates)
|
||||
|
||||
CPU-only (opt-1.3b; Qwen2.5-1.5B for multilingual; MarkLLM). Rough
|
||||
per-op figures from existing runs: gen ≈ 1-3 min/text (CPU), detect ≈
|
||||
10-40 s/text.
|
||||
|
||||
| Stage | Volume | CPU-hours (1 core) | Parallel 8 cores |
|
||||
| --- | --- | --- | --- |
|
||||
| Generate (incl. multilingual) | 7,000 texts | ~250-400 | ~35-50 h |
|
||||
| Detect (originals+attacks+controls) | ~65,000 | ~180-720 | ~25-90 h |
|
||||
| Null corpus scoring | 2,000 | ~10-20 | ~2-3 h |
|
||||
| Rewrites (API) | 56,000 outputs | — | ~1-2 days (rate-limited) |
|
||||
| Quality metrics (stratified subset) | ~2-4k texts | ~60-150 | ~10-20 h |
|
||||
|
||||
API rewrite budget: **~$50-120** (paraphrase ≈ 1.3× input tokens per
|
||||
pass; back-translate ≈ 2×2 passes; structural ≈ 2 passes; cheap
|
||||
instruct model, temperature 0.9 — record model + version).
|
||||
|
||||
→ **v1 ≈ 2.5-4 weeks of part-time work on one 8-core box + API
|
||||
budget.** That is the arXiv-v1 core; the extended scope adds ~1-2 weeks
|
||||
over the original Tier-1-only estimate.
|
||||
|
||||
## 7. Reproducibility
|
||||
|
||||
- Pin: MarkLLM checkout commit (`git -C /home/guillaume/MarkLLM
|
||||
rev-parse HEAD`), HF model revisions (opt-1.3b, Qwen2.5-1.5B,
|
||||
gpt2-large, deberta-xlarge-mnli, all-MiniLM-L6-v2), all scheme
|
||||
configs (copied into `research/configs/`).
|
||||
- Seed policy: every random op seeded; seeds recorded per cell in
|
||||
`results/manifest.json`.
|
||||
- Environment: record `pip freeze` for the MarkLLM env (it is separate
|
||||
from the service stdlib env) and for the quality-metric env.
|
||||
- Release plan: `research/results/` (JSONL per cell), analysis
|
||||
notebooks, corpus, and a `Makefile`-style runner
|
||||
(`run_experiments.py --plan`).
|
||||
|
||||
## 8. Artifacts & release
|
||||
|
||||
1. `research/corpus/` — 25 EN + 75 multilingual prompts.
|
||||
2. `research/configs/` — scheme configs + keys + model pins.
|
||||
3. `research/results/` — scores, ROC data, quality metrics, manifest.
|
||||
4. `research/scripts/` — attack additions (cheap.py, gap 05-A3),
|
||||
analysis notebooks.
|
||||
5. Paper artifacts: tables 1-7, figures 1-6 (spec in 02 §6).
|
||||
|
||||
## 9. Smoke test (do this first, before the full run)
|
||||
|
||||
Purpose: verify the pipeline end-to-end (generation → attack → detect
|
||||
→ evaluate) on one scheme and a handful of prompts before committing
|
||||
the 3,500-cell run.
|
||||
|
||||
```bash
|
||||
# Requires the multi-scheme bench refactor (gap 05-A1) to have landed.
|
||||
python3 service/scripts/bench_synthid_text.py \
|
||||
--corpus research/corpus --docs 5 --seeds 2 --max-new-tokens 300 \
|
||||
--variants paraphrase:3 --restamp-control --tag smoke \
|
||||
--markllm-model facebook/opt-1.3b
|
||||
```
|
||||
|
||||
Then, once the ROC/quality analysis modules (gaps 05-B1/B2) exist:
|
||||
compute AUROC/TPR@FPR + quality on the smoke output, and sanity-check
|
||||
the multilingual pilot (5 prompts × 1 seed per language, DE/FR/ES, via
|
||||
Qwen2.5-1.5B) for output quality before the full multilingual grid.
|
||||
|
||||
## 10. Risk register
|
||||
|
||||
| Risk | Mitigation |
|
||||
| --- | --- |
|
||||
| "Detector is misconfigured" | Same-config detection is the field standard; add best-effort tuned detector + strongest-config column; identical config JSON for gen and detect (§3) |
|
||||
| "Attacks destroy text" | Quality metrics at collapse point; show PPL/BERTScore within tolerance |
|
||||
| "Not real vendors" | Honest limitation; vendor APIs are black-box/retired (Google retired SynthID API Aug 2026); optional Claude API probe study if ToS allows |
|
||||
| "Already known" (novelty) | Novelty = layered attack + systematic ROC measurement across schemes + quality Pareto + policy measurement; check 03-related-work for overlap |
|
||||
| Ethics rejection | 04-ethics-and-legal.md; frame as robustness evaluation of deployed mechanisms |
|
||||
| Compute blowup | Locked restricted matrix (3,500 cells, not 21,000); `--plan` budget mode in run_experiments.py; stage checkpoints |
|
||||
| **Multilingual generation quality** | opt-1.3b is English-only → Qwen2.5-1.5B holdout; 5-prompt pilot check before the grid; fallback: shrink to 10 prompts/lang or defer to v1.1 |
|
||||
| Bench hardcodes SynthID | Add `--scheme kgw|synthid|exp|unigram|sir` to bench (gap 05-A1; small refactor) |
|
||||
@@ -0,0 +1,188 @@
|
||||
# 02 — Paper Outline & Venue Strategy
|
||||
|
||||
Version: v0.2 (2026-08-18). Status: skeleton; fill as results land.
|
||||
Scope locked for arXiv v1 (see research/README.md decision log).
|
||||
|
||||
---
|
||||
|
||||
## 1. Working title (locked)
|
||||
|
||||
1. **"How Fragile Are Deployed Text Watermarks? An Empirical Study of
|
||||
Layered Watermark Removal under Realistic User-Side Editing"**
|
||||
(safe, descriptive — **locked for arXiv v1**; measurement-study
|
||||
framing per the decision log)
|
||||
|
||||
Dropped on 2026-08-18 (kept here only as history): *"Watermarks Are
|
||||
Speed Bumps…"* — tone risk at archival venues; *"Removing Provenance
|
||||
Marks…"* — neutral fallback if the title must change.
|
||||
|
||||
## 2. Abstract draft (~150 words)
|
||||
|
||||
> Text watermarking is the primary mechanism proposed for EU AI Act
|
||||
> Art. 50 transparency obligations on machine-generated content. We
|
||||
> measure its robustness against realistic user-side editing. Using
|
||||
> same-config detection over 3,500 watermarked and 3,500 unwatermarked
|
||||
> generations (KGW, SynthID-Text, EXP, Unigram, and SIR schemes, in
|
||||
> English and German/French/Spanish), we evaluate a layered removal
|
||||
> pipeline that combines formatting-layer cleanup (invisible Unicode,
|
||||
> bidi) with statistical rewriting driven by detection feedback, and we
|
||||
> report ROC-based metrics (AUROC, TPR@1%FPR) and quality metrics
|
||||
> (perplexity, BERTScore, ROUGE-L) at the point of detection collapse.
|
||||
> We find that quality-preserving paraphrase and translation round-trip
|
||||
> collapse KGW-class detection to near chance, that SynthID-Text resists
|
||||
> token substitution but not paraphrase, that layering dominates single
|
||||
> attacks at equal quality cost, and that multilingual texts are
|
||||
> systematically more fragile. We discuss implications for Art. 50
|
||||
> compliance and release our corpus, configs, and harness.
|
||||
|
||||
(Recheck numbers against the final run during W2.)
|
||||
|
||||
## 3. Venue plan
|
||||
|
||||
| Step | Venue | When | Effort |
|
||||
| --- | --- | --- | --- |
|
||||
| 1 (**active**) | **arXiv preprint v1** | target ~2–4 weeks after core results | full paper: core + selected extensions |
|
||||
| 2 (future) | NeurIPS 2026 workshop → ACL 2027 / TIFS | post-v1, per CFP | reuse v1; see 04 for ethics/dual-use per venue |
|
||||
|
||||
Strategy: **arXiv v1 now**; do not split the same skeleton across two
|
||||
main tracks — workshops + one archival venue is the clean post-v1 path.
|
||||
|
||||
## 4. Contributions (4 bullets, in final form)
|
||||
|
||||
1. **Measurement.** First ROC-based robustness study of deployed-class
|
||||
text watermarking (KGW, SynthID-Text, EXP, Unigram, SIR) under
|
||||
realistic, layered user-side editing, with empirical-null FPR
|
||||
calibration.
|
||||
2. **Method.** A layered removal pipeline (formatting + statistical)
|
||||
with detection-feedback rewriting; we show it dominates single-layer
|
||||
baselines at equal quality cost (Pareto).
|
||||
3. **Resource.** Open corpus (25 EN + 75 multilingual prompts),
|
||||
config-pinned harness (MarkLLM-based), and results (JSONL) for
|
||||
reproducible attack/defense benchmarking.
|
||||
4. **Policy measurement.** Evidence on whether Art. 50 transparency
|
||||
obligations can rely on watermarking, with concrete recommendations
|
||||
(provenance at platform level, metadata, robust-but-invisible
|
||||
schemes, honest failure modes).
|
||||
|
||||
## 5. Section skeleton
|
||||
|
||||
### 1. Introduction (~1 p)
|
||||
- Hook: Art. 50 in force 2026-08-02; vendor rollouts (Claude, Gemini);
|
||||
Google retired SynthID-Text API Aug 2026 (context: the market is
|
||||
consolidating on methods that don't survive editing).
|
||||
- The viral deployment of our removal tool as motivation (1 short
|
||||
paragraph, no metrics needed) → "users edit their own output; do
|
||||
watermarks survive?"
|
||||
- Contributions (4 bullets above). **Fig 1** here.
|
||||
|
||||
### 2. Background & Related Work (~1.5 p)
|
||||
- Watermarking families: sampling-based (KGW line), tournament
|
||||
(SynthID-Text), semantic (SIR/X-SIR), provable (UPV, Christ-family).
|
||||
- Attack literature: paraphrase, back-translation, random-walk
|
||||
impossibility, watermark stealing. Metadata/C2PA: one short sentence
|
||||
(the full metadata mini-study is deferred to v2).
|
||||
- **Positioning paragraph:** what we add = layered attack + systematic
|
||||
ROC measurement + quality Pareto + policy measurement.
|
||||
(Full citation map in 03.)
|
||||
|
||||
### 3. Threat Model & System (~1 p)
|
||||
- Who: end users editing text they generated with their own account
|
||||
(explicitly **not** third-party content; see ethics file).
|
||||
- What: watermark-as-label (not access control); detector = same-config
|
||||
MarkLLM detector (standard in literature; vendors black-box).
|
||||
- System: 2 layers for v1 (A formatting / B statistical) with a
|
||||
detection-feedback rewrite loop. **Fig 1** (pipeline diagram).
|
||||
(Files/metadata layer deferred to v2.)
|
||||
- Definitions box: TPR@FPR, AUROC, "detection collapse", "quality cost".
|
||||
|
||||
### 4. Experimental Setup (~1.5 p)
|
||||
- Design summary (→ 01 §2): locked v1 factorial — 3,500 cells,
|
||||
7 schemes, lengths 100/300/500, temps 0.7/1.0, languages en/de/fr/es
|
||||
with the restricted subsets; generation protocol; attack cells
|
||||
(A0–A8); detection protocol incl. empirical null and best-effort
|
||||
detector; quality metrics table.
|
||||
- Reproducibility: pins, seeds, configs, release URLs.
|
||||
|
||||
### 5. Results (~3 p)
|
||||
- **Table 2** (money table) → discussion of each scheme's failure mode.
|
||||
- **Fig 2** ROC pre/post per scheme; **Fig 3** Pareto frontier;
|
||||
**Table 3** quality; **Table 4** strength×length ablation;
|
||||
**Table 5** multilingual (v1 extension); **Table 6** attack cost.
|
||||
- Key claims: (i) paraphrase/back-translation collapse KGW-class
|
||||
detection; (ii) SynthID-Text robust to token substitution but not
|
||||
paraphrase; (iii) layering > single layers at equal quality cost;
|
||||
(iv) even the strongest config (γ=.5, δ=4, 300 tok) drops below
|
||||
usable TPR under adaptive rewrite; (v) multilingual (DE/FR/ES) is
|
||||
systematically more fragile.
|
||||
- **Fig 5** case study (redacted before/after + detector scores).
|
||||
|
||||
### 6. Analysis & Case Study (~1 p)
|
||||
- Cost of attack vs cost of defense (Table 6): attacker spends ~10-20¢
|
||||
and minutes; defender must raise γ/δ (quality cost) — asymmetry.
|
||||
- Failure modes ranked; which scheme properties survive (semantic
|
||||
preservation, n-gram robustness) and which don't (token-level stats).
|
||||
|
||||
### 7. Policy Discussion (~0.75 p)
|
||||
- Art. 50 mechanics; what a regulator can actually rely on; metadata +
|
||||
platform provenance as the robust complement; honest limits of our
|
||||
study (no vendor black-box measurement).
|
||||
|
||||
### 8. Limitations (~0.5 p)
|
||||
- Same-config ≠ vendor detection; small open-weight generators
|
||||
(opt-1.3b; Qwen2.5-1.5B for DE/FR/ES as an explicit model holdout —
|
||||
not frontier LLMs); API probe limits; our rewrite oracle is stronger
|
||||
than naive users (we show both adaptive and single-pass numbers).
|
||||
|
||||
### 9. Ethics Statement (~0.5 p)
|
||||
- Draft text in 04 §3; adapt to venue template.
|
||||
|
||||
### 10. Conclusion (~0.25 p)
|
||||
|
||||
## 6. Exact tables & figures spec
|
||||
|
||||
### Tables
|
||||
| # | Content | Why reviewers need it |
|
||||
| --- | --- | --- |
|
||||
| T1 | Attack taxonomy: family × mechanism × implementation × prior-art anchor | reproducibility + novelty of layering |
|
||||
| T2 | **AUROC / TPR@1%FPR matrix: rows = scheme-config (7) × attack (8); cols = pre-attack, A-only, B-only, A+B** (core findings) | money table |
|
||||
| T3 | Quality per attack at collapse point: PPL Δ, BERTScore, ROUGE-L, SBERT, length drift, num/URL survival | preempts "destroys text" |
|
||||
| T4 | Ablation: TPR@1%FPR × (γ,δ) × length × temp | strength/length dependence |
|
||||
| T5 | Multilingual EN/DE/FR/ES (v1 extension) | known weak spot, cheap win |
|
||||
| T6 | Attack cost: tokens in/out, wall time, USD per 1k words per attack | practical-asymmetry argument |
|
||||
| T7 | Comparison with published baselines (cite-and-compare where numbers are reported in the same metric) | situates vs literature |
|
||||
|
||||
### Figures
|
||||
| # | Content |
|
||||
| --- | --- |
|
||||
| F1 | Pipeline/system diagram (2 layers + detection-feedback loop) |
|
||||
| F2 | ROC curves pre/post per scheme (the money figure) |
|
||||
| F3 | Quality–detectability Pareto frontier (PPL vs AUROC across attacks) |
|
||||
| F4 | TPR@1%FPR vs watermark strength, length as line style |
|
||||
| F5 | Case study: redacted before/after text with detector scores |
|
||||
| F6 | (policy/position only) Art. 50 timeline vs measured collapse — include, small (policy) |
|
||||
|
||||
## 7. Citation placement map
|
||||
|
||||
| Section | Cites (from 03) |
|
||||
| --- | --- |
|
||||
| 2 watermark families | KGW, SynthID-Text, SIR, UPV, survey, MarkLLM |
|
||||
| 2 attacks | X-SIR translation, fragility-of-multilingual, random-walk impossibility, watermark stealing, black-box watermarking |
|
||||
| 2 metadata (one sentence) | C2PA spec |
|
||||
| 5 results | X-SIR (cross-lingual numbers), fragility paper (paraphrase numbers) |
|
||||
| 7 policy | EU AI Act Art. 50 (Regulation (EU) 2024/1689) |
|
||||
|
||||
## 8. Writing phases (solo)
|
||||
|
||||
1. **W1:** core experiment runs; Tables 2–4 first drafts (numbers only).
|
||||
2. **W2:** Figs 2–3, T6 cost, §4–5 prose; **arXiv v1** (target ~2–4
|
||||
weeks from core results).
|
||||
3. **W3+ (post-v1):** future venue per §3; extend if anything remains.
|
||||
|
||||
## 9. Reviewer-bait checklist (preempt in text)
|
||||
|
||||
- [ ] "Detector is strawman" → best-effort tuned detector + strongest config column (T2 includes δ=4)
|
||||
- [ ] "You destroyed the text" → T3 at collapse point
|
||||
- [ ] "Unrealistic oracle" → report both adaptive (early-stop) and single-pass numbers
|
||||
- [ ] "No negative control" → unwatermarked control through full pipeline (`--restamp-control`)
|
||||
- [ ] "Not reproducible" → configs + seeds + manifest; MarkLLM commit pinned
|
||||
- [ ] "Ethics" → 04 file; ethics statement §9
|
||||
@@ -0,0 +1,80 @@
|
||||
# 03 — Related Work (verified citation list)
|
||||
|
||||
Every entry below was checked against the arXiv API on **2026-08-18**
|
||||
(title/ID queries); all entries marked ✅ resolved to the exact paper.
|
||||
**Re-verify every ID and venue label before the camera-ready / at
|
||||
submission time.**
|
||||
|
||||
Four candidate citations were dropped on 2026-08-18 because they could
|
||||
not be verified (see v0.1 §F): re-verify before ever re-adding any of
|
||||
them (Blinov et al. "Who Wrote this?", Zhao et al. provable-watermark
|
||||
OpenReview claim, the "Fragility of Multilingual LLMs" deletion-paper
|
||||
claim, "Stronger Watermarks for Language Models").
|
||||
|
||||
Venue labels are from the papers'/MarkLLM's own claims where available;
|
||||
re-verify venue before the camera-ready.
|
||||
|
||||
---
|
||||
|
||||
## A. Watermarking methods (the attack surface we test)
|
||||
|
||||
| ✅ | Cite | First author | ID / DOI | Date | Notes & where used in paper |
|
||||
| --- | --- | --- | --- | --- | --- |
|
||||
| ✅ | A Watermark for Large Language Models (KGW) | John Kirchenbauer | arXiv **2301.10226** (ICML 2023) | 2023-01 | Primary scheme under test; §2, §5 |
|
||||
| ✅ | Undetectable Watermarks for Language Models | Miranda Christ | arXiv **2306.09194** (COLT 2024) | 2023-05 | Christ-family (PF line); §2 robustness theory |
|
||||
| ✅ | Robust Distortion-free Watermarks for Language Models | Rohith Kuditipudi | arXiv **2307.15593** (TMLR) | 2023-07 | Sampling-free family; §2 |
|
||||
| ✅ | A Semantic Invariant Robust Watermark (SIR) | Aiwei Liu | arXiv **2310.06356** (ICLR 2024) | 2023-10 | Strongest defense class; Tier-2 scheme; §2, §5 |
|
||||
| ✅ | An Unforgeable Publicly Verifiable Watermark (UPV) | Aiwei Liu | arXiv **2307.16230** (ICLR 2024) | 2023-07 | Provable family; §2 |
|
||||
| ✅ | Permute-and-Flip: an optimally stable and watermarkable decoder | Xuandong Zhao | arXiv **2402.05864** | 2024-02 | §2 |
|
||||
| ✅ | Unbiased Watermark for Large Language Models | Zhengmian Hu | arXiv **2310.10669** | 2023-10 | §2 |
|
||||
| ✅ | A Resilient and Accessible Distribution-Preserving Watermark (DiPmark) | Yihan Wu | arXiv **2310.07710** | 2023-10 | §2 |
|
||||
| ✅ | Token-Specific Watermarking (TS-Watermark) | Mingjia Huo | arXiv **2402.18059** (ICML 2024) | 2024-02 | §2 |
|
||||
| ✅ | Adaptive Text Watermark | Yepeng Liu | arXiv **2401.13927** | 2024-01 | §2 |
|
||||
| ✅ | SemStamp: A Semantic Watermark with Paraphrastic Robustness | Abe Bohan Hou | arXiv **2310.03991** | 2023-10 | semantic family; §2 |
|
||||
| ✅ | k-SemStamp: A Clustering-Based Semantic Watermark | Abe Bohan Hou | arXiv **2402.11399** | 2024-02 | §2 |
|
||||
| ✅ | Invisible Entropy (IE): Safe and Efficient Low-Entropy LLM Watermarking | Tianle Gu | arXiv **2505.14112** | 2025-05 | recent method; §2 |
|
||||
| ✅ | MorphMark: Flexible Adaptive Watermarking | Zongqi Wang | arXiv **2505.11541** | 2025-05 | §2 |
|
||||
| ✅ | Watermarking Text Generated by Black-Box LLMs | Xi Yang | arXiv **2305.08883** (NAACL 2024) | 2023-05 | black-box line; §2 |
|
||||
|
||||
## B. Robustness, attacks, and limits (our direct neighbors — positioning)
|
||||
|
||||
| ✅ | Cite | First author | ID / DOI | Date | Notes |
|
||||
| --- | --- | --- | --- | --- | --- |
|
||||
| ✅ | On the Reliability of **Watermarks** for Large Language Models (note: title is "Watermarks", not "Watermarking") | John Kirchenbauer | arXiv **2306.04634** (NeurIPS 2023) | 2023-06 | foundational robustness study; §2, §5 compare |
|
||||
| ✅ | Can Watermarks Survive Translation? (X-SIR) | Zhiwei He | arXiv **2402.14007** (ACL 2024) | 2024-02 | cross-lingual attack; our A4 baseline anchor; §2, §5 |
|
||||
| ✅ | Watermarks in the Sand: Impossibility of Strong Watermarking | Hanlin Zhang | arXiv **2311.04378** (ICML 2024) | 2023-11 | Random Walk attack; impossibility theory; §2, §7 |
|
||||
| ✅ | Watermark Stealing in Large Language Models | Nikola Jovanović | arXiv **2402.19361** | 2024-02 | key-extraction threat; §2 |
|
||||
| ✅ | WaterSeeker: Efficient Detection of Watermarked Segments | Leyi Pan | arXiv **2409.05112** (NAACL 2025 Findings) | 2024-09 | detector side; §2 |
|
||||
| ✅ | An Entropy-based Text Watermarking Detection Method (EWD) | Yijian Lu | arXiv **2403.13485** (ACL 2024) | 2024-03 | detector side; §2 |
|
||||
| ✅ | Can Watermarked LLMs be Identified by Users via Crafted Prompts? | Aiwei Liu | arXiv **2410.03168** (ICLR 2025 Spotlight) | 2024-10 | §2 |
|
||||
| ✅ | Can LLM Watermarks Robustly Prevent Unauthorized Knowledge Distillation? | Leyi Pan | arXiv **2502.11598** (ACL 2025) | 2025-02 | §2 |
|
||||
|
||||
## C. Closest recent neighbors (2025-2026 measurement/forensics — cite these!)
|
||||
|
||||
| ✅ | Cite | First author | ID / DOI | Date | Notes |
|
||||
| --- | --- | --- | --- | --- | --- |
|
||||
| ✅ | Robustness Assessment and Enhancement of Text Watermarking for Google's SynthID | Xia Han | arXiv **2508.20228** | 2025-08 | closest method-specific robustness work; must cite & differentiate in §2/§5 |
|
||||
| ✅ | On Google's SynthID-Text LLM Watermarking System: Theoretical Analysis and Empirical Validation | Romina Omidi | arXiv **2603.03410** | 2026-03 | theoretical analysis of SynthID-Text; §2 |
|
||||
| ✅ | AI Watermark Evidence Fails Forensic Readiness: An Empirical Evaluation | Saifur Rahman Tamim | arXiv **2607.16010** | 2026-07 | directly supports our policy finding; §7 |
|
||||
| ✅ | Sandcastles in the Storm: Revisiting the (Im)possibility of Strong Watermarking | Fabrice Y Harel-Canada | arXiv **2505.06827** | 2025-05 | impossibility revisited; §2 |
|
||||
|
||||
## D. Tools & surveys
|
||||
|
||||
| ✅ | Cite | First author | ID / DOI | Date | Notes |
|
||||
| --- | --- | --- | --- | --- | --- |
|
||||
| ✅ | MarkLLM: An Open-Source Toolkit for LLM Watermarking | Leyi Pan | arXiv **2405.10051** (EMNLP 2024 Demo) | 2024-05 | our harness; acknowledge THU-BPM |
|
||||
| ✅ | A Survey of Text Watermarking in the Era of LLMs | Aiwei Liu | arXiv **2312.07913** (ACM Computing Surveys 2025) | 2023-12 | §2 overview + taxonomy |
|
||||
|
||||
## E. Non-arXiv sources (correct IDs; cite directly)
|
||||
|
||||
- **SynthID-Text** — Google DeepMind, *Nature* **638**, 625-632 (2024); DOI [**10.1038/s41586-024-08025-4**](https://www.nature.com/articles/s41586-024-08025-4). (arXiv tech-report id exists but was not resolvable in our checks — cite the Nature DOI as primary.)
|
||||
- **C2PA** — Coalition for Content Provenance and Authenticity, specification, https://c2pa.org/specifications/specifications/2.1/specs/C2PA_Specification.html
|
||||
- **EU AI Act** — Regulation (EU) 2024/1689, Art. 50 (transparency obligations), in force for GPAI since 2026-08-02.
|
||||
|
||||
## Positioning summary (for §2 last paragraph)
|
||||
|
||||
- vs **X-SIR** (2402.14007): they test translation robustness of *their own* scheme; we test the *deployed-class* schemes users actually meet, with a layered pipeline and ROC metrics.
|
||||
- vs **2508.20228 / 2603.03410** (SynthID robustness): they assess/enhance SynthID specifically; we add KGW, multi-layer attacks, quality Pareto, and the policy measurement.
|
||||
- vs **2607.16010** (forensic readiness): they measure evidence-grade failure; we measure the removal side end-to-end — complementary.
|
||||
- vs **2306.04634** (Reliability): they established fragility of watermarking; we add the layered attack + empirical-null FPR + quality-at-collapse, on 2024-2026 schemes.
|
||||
- Novelty claim rests on: **layered attack (formatting+statistical in v1; metadata layer deferred to v2), systematic ROC measurement across schemes under one protocol, quality Pareto frontier, and Art. 50 policy measurement** — none of the above does all four.
|
||||
@@ -0,0 +1,124 @@
|
||||
# 04 — Ethics & Legal Framing
|
||||
|
||||
Version: v0.1 (2026-08-18). This file feeds the paper's Ethics Statement
|
||||
and your own position (see `~/wired_interview_prep.md` — the two should
|
||||
stay consistent).
|
||||
|
||||
---
|
||||
|
||||
## 1. Dual-use framing (the core stance)
|
||||
|
||||
This is a **robustness/measurement study of deployed mechanisms**, not a
|
||||
how-to for harming third parties. Published precedent is strong: attack
|
||||
and robustness papers on watermarking are routine at NeurIPS, ICLR,
|
||||
ACL, IEEE S&P, USENIX Security (see 03: random-walk impossibility,
|
||||
watermark stealing, fragility studies — all published, all attacker-
|
||||
facing). The field treats "measure the mechanism before regulators
|
||||
mandate it" as legitimate, even necessary, research.
|
||||
|
||||
Frame, consistently:
|
||||
|
||||
1. **The tool edits text the user generated with their own account.**
|
||||
It is not a forgery or circumvention tool (watermark = label, not
|
||||
lock — no access control is bypassed).
|
||||
2. **The contribution is measurement.** We quantify what the literature
|
||||
already suspected (fragility under paraphrase/translation) and what
|
||||
the public deployment showed at scale — with controlled experiments
|
||||
instead of anecdotes.
|
||||
3. **We do not attack third-party content.** Corpus is self-generated.
|
||||
4. **We help defenders too.** Detection-feedback loops, empirical-null
|
||||
FPR calibration, and the Pareto analysis are directly useful for
|
||||
building watermarking that survives reality (or for regulators to
|
||||
pick mechanisms that do).
|
||||
|
||||
## 2. EU AI Act analysis (for §7 of the paper)
|
||||
|
||||
- Regulation (EU) 2024/1689, **Article 50** (transparency obligations
|
||||
for providers and deployers of certain AI systems), general
|
||||
application for GPAI systems: 2 August 2026 (per your notes and the
|
||||
public timeline).
|
||||
- Key point for the paper: Art. 50 obligations fall on **providers and
|
||||
deployers**, not on end users editing their own documents. The Act
|
||||
does not require end users to preserve machine-readable provenance
|
||||
marks; it requires providers to make machine-readable output
|
||||
detectable.
|
||||
- Therefore a user stripping a watermark from their own AI-assisted
|
||||
writing is not violating Art. 50. The *policy problem* the paper
|
||||
documents is upstream: if the mandated detection mechanism collapses
|
||||
under ordinary editing, the transparency obligation is not met in
|
||||
practice — that is a compliance/effectiveness finding, not a
|
||||
user-facing prohibition.
|
||||
- Cite: Regulation (EU) 2024/1689, Art. 50; recitals on transparency.
|
||||
Verify exact recital numbers before submission.
|
||||
|
||||
## 3. Ethics statement draft (adapt to venue template)
|
||||
|
||||
> All text used in this study was generated by the authors using
|
||||
> open-weight models (opt-1.3b) via the MarkLLM toolkit. No
|
||||
> third-party content, user data, or live vendor outputs were used;
|
||||
> no watermarked content produced by commercial providers was
|
||||
> collected or altered. The removal pipeline evaluated here operates
|
||||
> on text generated by the same user who owns it; the study does not
|
||||
> enable or endorse alteration of third-party content. Watermarking
|
||||
> is a label, not an access-control mechanism, so no security control
|
||||
> is circumvented. We disclose our findings to support (i) realistic
|
||||
> expectations for regulators relying on watermarking for transparency
|
||||
> obligations (Art. 50, Regulation (EU) 2024/1689), and (ii) the
|
||||
> design of more robust provenance mechanisms. Detectors are
|
||||
> same-config open-source implementations; commercial detectors were
|
||||
> not probed. We do not provide live removal services or weights
|
||||
> tuned for evasion of specific vendor detectors beyond what is
|
||||
> reported. The authors' tooling is public (github.com/guillaumemeyer/
|
||||
> watermarks-remover) and was deployed publicly before this study
|
||||
> began; this paper formalizes measurements of mechanisms already in
|
||||
> production use.
|
||||
|
||||
## 4. Data policy
|
||||
|
||||
- **v1: synthetic data only.** No user data from the viral deployment.
|
||||
If usage telemetry is ever included: aggregate, anonymize, obtain
|
||||
consent, and get IRB/ethics review.
|
||||
- No scraping of watermarked third-party text (avoids both legal and
|
||||
ethical surface).
|
||||
- Release: corpus, configs, results JSONL (all synthetic).
|
||||
|
||||
## 5. Legal notes (know your lines; don't overclaim)
|
||||
|
||||
- MarkLLM is Apache-2.0 → attribution required, no restriction on
|
||||
research use. Credit THU-BPM in acknowledgments.
|
||||
- C2PA stripping: your position is privacy/hygiene on owned content.
|
||||
Note that anti-circumvention regimes (e.g., US DMCA 1201) generally
|
||||
target access controls, not labels; watermark-as-label framing keeps
|
||||
you clear, but do not give legal opinions in the paper — one
|
||||
sentence of framing, then cite nothing you haven't verified.
|
||||
- **Never claim 100% effectiveness** anywhere in the paper (reviewers
|
||||
and journalists will test it). Use ROC numbers and honest confidence
|
||||
intervals.
|
||||
- Disclosure status: the tool is already public and widely reported;
|
||||
no embargo or coordinated-disclosure obligation applies to the
|
||||
measurement. State this if a venue asks about disclosure.
|
||||
|
||||
## 6. Venue-specific notes
|
||||
|
||||
- **arXiv v1 (active target):** arXiv itself has no ethics checklist;
|
||||
include the §3 statement as submitted and the §5 disclosure note
|
||||
(tool already public, no embargo applies). Pick a license at upload.
|
||||
- Future venues (post-v1, one line): NeurIPS/ACL want an ethics
|
||||
checklist + responsible-NLP/broader-impact paragraph (§3 covers both);
|
||||
security venues (S&P/USENIX) expect a stronger adversary framing plus
|
||||
§7 defender takeaways; TIFS wants extended related work + a more
|
||||
formal threat model.
|
||||
|
||||
## 7. Defender takeaways (include in any version)
|
||||
|
||||
1. Token-level statistical watermarks do not survive paraphrase; don't
|
||||
bet Art. 50 compliance on them alone.
|
||||
2. N-gram/tournament methods (SynthID-Text) survive substitution but
|
||||
not semantic rewriting — combine with semantic-invariant schemes
|
||||
(SIR line) if rewriting is the threat.
|
||||
3. Robust detection needs empirical-null calibration, not normal
|
||||
assumptions.
|
||||
4. Platform-level provenance + metadata (C2PA) is complementary, not a
|
||||
substitute for text-layer robustness.
|
||||
5. Publish attack benchmarks openly so defenders can measure the
|
||||
Pareto frontier instead of assuming worst-case strength.
|
||||
@@ -0,0 +1,85 @@
|
||||
# 05 — arXiv v1 Readiness: Gap Analysis & Submission Checklist
|
||||
|
||||
Version: v0.1 (2026-08-18). Scope: arXiv v1 of *"How Fragile Are
|
||||
Deployed Text Watermarks? An Empirical Study of Layered Watermark
|
||||
Removal under Realistic User-Side Editing"* (locked scope in
|
||||
research/README.md decision log).
|
||||
|
||||
This file itemizes **everything missing** between the current repo and
|
||||
a submittable arXiv v1. It is a hand-off to the implementation phase —
|
||||
**as of PR #174 all code/analysis gaps (A1-A7, B1-B3) and the paper
|
||||
skeleton/bibliography (C1/C2) are implemented**; the experiment run
|
||||
(D) and publishing logistics (E) remain. Items are marked with the
|
||||
stage of the pipeline they block (generate → attack → detect →
|
||||
evaluate → paper → publish).
|
||||
|
||||
---
|
||||
|
||||
## A. Verified code gaps (experiment harness)
|
||||
|
||||
Each row: what exists today / what's missing / effort / where it lands.
|
||||
(All "exists" claims verified 2026-08-18.)
|
||||
|
||||
| # | Gap | What exists today | What's missing | Effort | Blocks |
|
||||
| --- | --- | --- | --- | --- | --- |
|
||||
| A1 | Multi-scheme **generation** | `service/scripts/bench_synthid_text.py` hardcodes `SCHEME = "synthid"` (line 63); MarkLLM checkout already ships `config/{KGW,SynthID,EXP,Unigram,SIR}.json` | `--scheme kgw|synthid|exp|unigram|sir` + `--config` override on the bench; 3 KGW strength JSONs (`research/configs/KGW-d1/d2/d4.json`) mirroring the checkout's `KGW.json` (γ=.25 δ=1; γ=.5 δ=2; γ=.5 δ=4) | small refactor (per 01 §10) | generate |
|
||||
| A2 | Multi-scheme **detection** | `service/scripts/detect_text_watermark.py` scheme map (lines 41–45) covers `kgw`, `synthid` only | Add `exp`, `unigram`, `sir` entries + config wiring; same-config detection guarantee (01 §3: identical JSON for gen and detect) | small | detect |
|
||||
| A3 | A7 cheap baselines | nothing | `research/scripts/attacks/cheap.py`: synonym substitution, 5–10% random word deletion, sentence reorder; deterministic + seeded | small | attack |
|
||||
| A4 | Orchestrator stage wiring | `research/scripts/run_experiments.py` stages `generate/attack/detect/evaluate/report` all `raise NotImplementedError` (design/constants already aligned to the locked matrix) | Wire each stage to the repo scripts; emit documented results layout (`results/manifest.json`, per-cell `generated/attacked/scores.jsonl`, `metrics.json`, `report.md`); resume-able checkpoints | medium | all |
|
||||
| A5 | Multilingual generator | opt-1.3b is English-only; **no GPU available** (verified 2026-08-18) | CPU path for `Qwen/Qwen2.5-1.5B-Instruct` (fallback 0.5B) via MarkLLM `TransformersConfig`; 5-prompt × 1-seed pilot quality check per language (01 §9) before the 750-cell grid | medium | generate (de/fr/es) |
|
||||
| A6 | Corpus | `benchmarks/corpus/` has **8 files** (research/README previously said 9); factual, neutral, 50–90 words | 17 new EN prompts + 3×25 translated DE/FR/ES sets → `research/corpus/` (self-written, checkable claims) | medium | generate |
|
||||
| A7 | Config pinning & manifest | `requirements-markllm.txt` pinned; MarkLLM checkout at `/home/guillaume/MarkLLM` | Record MarkLLM commit (`git -C /home/guillaume/MarkLLM rev-parse HEAD`), HF revisions (opt-1.3b, Qwen2.5-1.5B, gpt2-large, deberta-xlarge-mnli, all-MiniLM-L6-v2), `pip freeze` for MarkLLM + quality envs, `manifest.json` writer | small | reproducibility |
|
||||
|
||||
## B. Analysis gaps (nothing exists — verified by grep, 2026-08-18)
|
||||
|
||||
| # | Gap | Spec (from 01 §5) | Effort |
|
||||
| --- | --- | --- | --- |
|
||||
| B1 | ROC module | `research/scripts/analyze_roc.py`: AUROC, TPR@FPR ∈ {0.1%, 1%, 10%}, full ROC data per cell, **empirical null** from unwatermarked controls (never assume normal z-scores), 95% bootstrap CIs (10k resamples) | medium |
|
||||
| B2 | Quality metrics | `research/scripts/evaluate_quality.py`: PPL (`gpt2-large`, never the generator), BERTScore (`deberta-xlarge-mnli`), ROUGE-L, SBERT cosine (`all-MiniLM-L6-v2`), Levenshtein %, length drift, number/URL survival; `research/requirements-quality.txt` (bert-score, rouge-score, sentence-transformers); stratified subset ~2–4k texts incl. every collapse point | medium |
|
||||
| B3 | Table/figure generators | T1–T7 + F1–F6 per 02 §6: F2 ROC curves, F3 Pareto (PPL vs AUROC), F4 strength×length, F1 pipeline TikZ, F5 case study, F6 policy timeline | medium |
|
||||
|
||||
## C. Paper gaps
|
||||
|
||||
| # | Gap | Notes |
|
||||
| --- | --- | --- |
|
||||
| C1 | LaTeX skeleton | `research/paper/`, ACL-style template (fits venue ladder); abstract draft (02 §2), ethics draft (04 §3), acknowledgments (MarkLLM/THU-BPM) exist as markdown |
|
||||
| C2 | Bibliography | `.bib` from 03 A–E; **§F candidates dropped** (see 03 header); re-verify every ID/venue at submission |
|
||||
| C3 | Tables 1–7, Figures 1–6 | Specs in 02 §6; **require real numbers from the run** — no placeholders |
|
||||
| C4 | Full prose §1–§10 | Skeleton in 02 §5; writing phase W2 |
|
||||
|
||||
## D. The experiment run itself (the largest single item)
|
||||
|
||||
- 3,500-cell locked matrix (01 §2): 7,000 generations, 56,000 attack
|
||||
outputs, ~65,000 detections + ~2,000 null controls.
|
||||
- Budget: ~$50–120 API rewrites (model + version recorded), ~2.5–4
|
||||
weeks wall on one 8-core CPU box (01 §6).
|
||||
- Order: smoke test (01 §9) → EN core → temp/length axes → multilingual
|
||||
pilot → multilingual grid → analysis (B1–B3) → tables/figures (C3).
|
||||
|
||||
## E. Publishing logistics (user-owned)
|
||||
|
||||
| # | Item | Who | Notes |
|
||||
| --- | --- | --- | --- |
|
||||
| E1 | arXiv account + endorsement | user | Solo, no affiliation (allowed); new authors in cs.CL typically need endorsement from an existing author — **start in week 1**, it can take days |
|
||||
| E2 | Categories | user | cs.CL primary; cs.CR + cs.LG secondary |
|
||||
| E3 | Release package | agent+user | Repo URL (github.com/guillaumemeyer/watermarks-remover), MIT license, corpus/configs/results JSONL; Zenodo DOI optional |
|
||||
| E4 | Final upload | user | After the checklist below passes; arXiv license selection at upload |
|
||||
|
||||
## F. arXiv v1 submission checklist
|
||||
|
||||
- [ ] Title/abstract/authors finalized (02 §1-2; decision log)
|
||||
- [ ] All Tables 2–7 contain real numbers (no placeholders)
|
||||
- [ ] All Figures 1–6 rendered from real data
|
||||
- [ ] Ethics statement (04 §3) + disclosure note (04 §5) in the PDF
|
||||
- [ ] Acknowledgments (MarkLLM / THU-BPM) present
|
||||
- [ ] Citations re-verified (03; IDs + venue labels re-checked at submission)
|
||||
- [ ] Data-availability + reproducibility statements (01 §7-8) written
|
||||
- [ ] Release links live (corpus, configs, results JSONL, harness)
|
||||
- [ ] E1/E2 done; E4 performed
|
||||
|
||||
## G. Deferred (v2, explicitly out of arXiv v1)
|
||||
|
||||
- Human eval (20 × 3 annotators) — journal/ACL requirement.
|
||||
- File/metadata mini-study (C2PA/EXIF strip; v0.1 01 §4.3) — separate
|
||||
mini-experiment.
|
||||
- Venue ladder beyond arXiv (NeurIPS 2026 workshop → ACL 2027 / TIFS).
|
||||
@@ -0,0 +1,84 @@
|
||||
# research/ — Watermark Removal: Measurement Study & Paper Kit
|
||||
|
||||
Working area for turning the `watermarks-remover` project into a
|
||||
publishable research paper. As of the arXiv-v1 implementation PR
|
||||
(https://github.com/guillaumemeyer/watermarks-remover/pull/174), the
|
||||
full code/analysis gap list (05-A1..A7, B1..B3, C1/C2) is implemented
|
||||
here; the multi-week experiment run (05-D) and publishing logistics
|
||||
(05-E) remain. Generated run data stays gitignored (`research/results/`).
|
||||
|
||||
## What is this paper about (one line)
|
||||
|
||||
*How robust are deployed-class LLM text watermarking schemes (KGW,
|
||||
SynthID-Text, EXP, Unigram, SIR) to realistic user-side editing, and
|
||||
what does a layered (formatting + statistical) removal pipeline do to
|
||||
detection-rate and text quality?*
|
||||
|
||||
## File map
|
||||
|
||||
| File | Contents |
|
||||
| --- | --- |
|
||||
| `01-experiment-protocol.md` | Full experimental protocol: locked v1 factorial, factors, attacks, detection & quality metrics, statistics, budget, smoke test, risk register |
|
||||
| `02-paper-outline.md` | Locked title, abstract draft, venue plan (arXiv v1 active), section-by-section skeleton, exact tables/figures spec, writing phases |
|
||||
| `03-related-work.md` | Verified citation list (arXiv IDs checked), grouped by theme, with "where we position vs each" notes |
|
||||
| `04-ethics-and-legal.md` | Dual-use framing, EU AI Act Art. 50 analysis, ethics statement draft, data policy, disclosure notes |
|
||||
| `05-arxiv-readiness.md` | **Gap analysis for arXiv v1**: every missing code/analysis/paper/logistics item, budget, submission checklist — what must be built before publishing |
|
||||
| `scripts/run_experiments.py` | Orchestrator: locked factorial design, `--plan` budget mode, and fully wired generate/attack/detect/evaluate/report stages with resume markers (gap 05-A4) |
|
||||
| `scripts/pins.py` | Reproducibility pins: MarkLLM commit, repo commit, HF revisions, pip freeze (gap 05-A7) |
|
||||
| `scripts/multilingual_gen.py` | CPU Qwen2.5-1.5B-Instruct generator for DE/FR/ES cells (model holdout, gap 05-A5) |
|
||||
| `scripts/attacks/cheap.py` | Deterministic cheap baselines: synonym / word-deletion / sentence-reorder (gap 05-A3) |
|
||||
| `scripts/analyze_roc.py` | AUROC, TPR@FPR (empirical null), bootstrap CIs (gap 05-B1) |
|
||||
| `scripts/evaluate_quality.py` | PPL/BERTScore/ROUGE-L/SBERT/Levenshtein + survival metrics (gap 05-B2) |
|
||||
| `scripts/make_tables.py`, `scripts/make_figures.py` | Paper tables T1-T7 and figures F1-F6 (gap 05-B3) |
|
||||
| `configs/` | Pinned scheme configs incl. KGW-d1/d2/d4 strength JSONs (same JSON for gen and detect) |
|
||||
| `corpus/` | 25 EN + 3x25 DE/FR/ES factual prompts (gap 05-A6) |
|
||||
| `tests/` | pytest suite for the research harness (`make research-check`) |
|
||||
| `paper/` | arXiv v1 LaTeX skeleton + verified bibliography (gaps C1/C2) |
|
||||
|
||||
## What already exists in the repo (reuse, don't rebuild)
|
||||
|
||||
| Capability | Where | Notes |
|
||||
| --- | --- | --- |
|
||||
| Watermarked corpus generation (SynthID only) | `service/scripts/bench_synthid_text.py` | MarkLLM `facebook/opt-1.3b`, 300 tok default, `--seeds`, `--docs`, `--max-new-tokens`; multi-scheme support is gap 05-A1 |
|
||||
| Layer A (Unicode/invisible chars) | `service/scripts/text_unicode.py` via `clean_text.py` | deterministic |
|
||||
| Layer B rewrites | `service/scripts/rewrite_text.py` | strengths: `paraphrase`, `humanize`, `backtranslate`, `structural`, `code`; transports: `ollama`, `openai-compatible`; evaluation-loop w/ early stop |
|
||||
| Detection (same-config, KGW/SynthID) | `service/scripts/detect_text_watermark.py`, `text_detectors.py` | EXP/Unigram/SIR detection is gap 05-A2 |
|
||||
| Existing metrics | `bench_synthid_text.py` | clear rate, score suppression, lexical divergence, length drift, number/URL survival, token/USD cost |
|
||||
| Corpus seeds | `benchmarks/corpus/` (**8 files** — README previously said 9) | factual, neutral, 50-90 words; 25 EN + 75 multilingual prompts needed (gap 05-A6) |
|
||||
| How-to doc | `docs/synthid-text-benchmark.md` | |
|
||||
|
||||
## Gap to close (this is the paper work)
|
||||
|
||||
Everything missing between today and an arXiv v1 is itemized in
|
||||
**`05-arxiv-readiness.md`** (verified code gaps, analysis gaps, paper
|
||||
gaps, publishing logistics, submission checklist, budget). Nothing in
|
||||
that list is implemented yet — it is the hand-off to the implementation
|
||||
phase.
|
||||
|
||||
## Status checklist
|
||||
|
||||
- [x] Decide framing — **measurement study** (not "attack tool" paper); title locked in 02 §1
|
||||
- [x] Scope locked for arXiv v1 — Tier 1 core + multilingual (DE/FR/ES) + EXP/Unigram/SIR + length 500 + temp 1.0; attacks A0–A8; policy §7 in; human eval & file/metadata mini-study **deferred** (decision log below)
|
||||
- [ ] Core experiment run (01 §2 matrix: 3,500 cells) — generate → attack → detect → evaluate
|
||||
- [ ] Tables 1-7 + Figures 1-6 (02 §6)
|
||||
- [ ] arXiv preprint (target: ~2-4 weeks after core results)
|
||||
- [ ] Post-v1: workshop submission (NeurIPS 2026 workshops) → journal (TIFS) or ACL 2027
|
||||
|
||||
## Deferred (v2, not in arXiv v1)
|
||||
|
||||
- Human eval (20 × 3 annotators) — journal requirement.
|
||||
- File/metadata mini-study (C2PA/EXIF strip; v0.1 01 §4.3).
|
||||
|
||||
## Decision log
|
||||
|
||||
| Date | Decision |
|
||||
| --- | --- |
|
||||
| 2026-08-18 | Scoped paper as *empirical robustness measurement* with layered-attack contribution; primary metrics ROC-based; venue ladder = arXiv → NeurIPS 2026 workshop → TIFS. Research dir created, gitignored, not committed. |
|
||||
| 2026-08-18 | **v1 scope locked** (Q&A): measurement framing; title 2 in 02 §1; matrix = 7 schemes (KGW×3, SynthID, EXP, Unigram, SIR) × lengths 100/300/500 × temps 0.7/1.0 × langs en/de/fr/es with restricted subsets = **3,500 cells**; attacks A0–A8; policy §7 in; **deferred**: human eval, file/metadata mini-study; rewrite backend = OpenAI-compatible API; solo author; full run first (~2-4 weeks). Multi-scheme bench/detector support, ROC + quality metrics, cheap.py, corpus expansion, and paper artifacts are tracked as gaps in 05-arxiv-readiness.md (not yet implemented). |
|
||||
|
||||
## Git note
|
||||
|
||||
The arXiv-v1 implementation PR adds `!/research/` allow rules to
|
||||
`.gitignore`, so the paper kit is tracked from that point on. Only
|
||||
`research/results/` (generated JSONL data, released via Zenodo) stays
|
||||
ignored.
|
||||
@@ -0,0 +1,7 @@
|
||||
{
|
||||
"algorithm_name": "EXP",
|
||||
"prefix_length": 4,
|
||||
"hash_key": 15485863,
|
||||
"threshold": 1e-4,
|
||||
"sequence_length": 200
|
||||
}
|
||||
@@ -0,0 +1,10 @@
|
||||
{
|
||||
"algorithm_name": "KGW",
|
||||
"gamma": 0.25,
|
||||
"delta": 1.0,
|
||||
"hash_key": 15485863,
|
||||
"prefix_length": 1,
|
||||
"z_threshold": 4.0,
|
||||
"f_scheme": "time",
|
||||
"window_scheme": "left"
|
||||
}
|
||||
@@ -0,0 +1,10 @@
|
||||
{
|
||||
"algorithm_name": "KGW",
|
||||
"gamma": 0.5,
|
||||
"delta": 2.0,
|
||||
"hash_key": 15485863,
|
||||
"prefix_length": 1,
|
||||
"z_threshold": 4.0,
|
||||
"f_scheme": "time",
|
||||
"window_scheme": "left"
|
||||
}
|
||||
@@ -0,0 +1,10 @@
|
||||
{
|
||||
"algorithm_name": "KGW",
|
||||
"gamma": 0.5,
|
||||
"delta": 4.0,
|
||||
"hash_key": 15485863,
|
||||
"prefix_length": 1,
|
||||
"z_threshold": 4.0,
|
||||
"f_scheme": "time",
|
||||
"window_scheme": "left"
|
||||
}
|
||||
@@ -0,0 +1,11 @@
|
||||
{
|
||||
"algorithm_name": "SIR",
|
||||
"delta": 1.0,
|
||||
"chunk_length": 10,
|
||||
"scale_dimension": 300,
|
||||
"z_threshold": 0.2,
|
||||
"transform_model_input_dim": 1024,
|
||||
"transform_model_name": "watermark/sir/model/transform_model_cbert.pth",
|
||||
"embedding_model_path": "watermark/sir/model/compositional-bert-large-uncased/",
|
||||
"mapping_name": "watermark/sir/mapping/300_mapping_50272.json"
|
||||
}
|
||||
@@ -0,0 +1,16 @@
|
||||
{
|
||||
"algorithm_name": "SynthID",
|
||||
"ngram_len": 5,
|
||||
"keys": [
|
||||
654, 400, 836, 123, 340, 443, 597, 160, 57, 29,
|
||||
590, 639, 13, 715, 468, 990, 966, 226, 324, 585,
|
||||
118, 504, 421, 521, 129, 669, 732, 225, 90, 960
|
||||
],
|
||||
"sampling_table_size": 65536,
|
||||
"sampling_table_seed": 0,
|
||||
"watermark_mode": "non-distortionary",
|
||||
"num_leaves": 2,
|
||||
"context_history_size": 1024,
|
||||
"detector_type": "mean",
|
||||
"threshold": 0.52
|
||||
}
|
||||
@@ -0,0 +1,7 @@
|
||||
{
|
||||
"algorithm_name": "Unigram",
|
||||
"gamma": 0.5,
|
||||
"delta": 2.0,
|
||||
"hash_key": 15485863,
|
||||
"z_threshold": 4.0
|
||||
}
|
||||
@@ -0,0 +1 @@
|
||||
Cloud-Computing ermöglicht es Organisationen, Rechenleistung, Speicher und Netzwerkressourcen zu mieten, statt eigene physische Server zu besitzen. Die Hauptmodelle sind Infrastructure as a Service, Platform as a Service und Software as a Service, die jeweils unterschiedlich viel Betriebsverantwortung an den Anbieter abgeben. Die Abrechnung erfolgt meist nutzungsbasiert, was Start-ups hilft, ohne große Anfangsinvestitionen zu wachsen. Zu den Abwägungen gehören Anbieterbindung, Datenresidenz-Vorschriften und die Notwendigkeit, Kosten bei unerwartet wachsenden Arbeitslasten sorgfältig zu überwachen.
|
||||
@@ -0,0 +1 @@
|
||||
Ein gleichmäßiger Kaffee hängt von einigen Variablen ab: Mahlgrad, Wassertemperatur, Dosis und Brühzeit. Feinere Mahlgrade extrahieren schneller, können aber bitter werden, wenn das Wasser zu heiß ist. Ein übliches Ausgangsverhältnis sind sechzig Gramm Kaffee pro Liter Wasser, angepasst an den eigenen Geschmack. Aufgusmethoden brauchen eine ruhige Hand und ein ebenes Kaffeebett, während Immersionsmethoden wie die French Press verzeihender sind. Bohnen in einem luftdichten Behälter und fern von Licht aufbewahrt, halten ihr Aroma länger als in der Originaltüte.
|
||||
@@ -0,0 +1 @@
|
||||
Eine Tagestour in den Bergen erfordert mehr als Wasser und Snacks. Regenkleidung, eine Karte, eine Stirnlampe und ein Erste-Hilfe-Set gehören in jeden Rucksack, auch an klaren Morgen. Das Wetter in großer Höhe kann sich innerhalb einer Stunde ändern, und oberhalb der Baumgrenze sind Wege oft unmarkiert. Teilen Sie jemandem Ihre geplante Route und die voraussichtliche Rückkehrzeit mit, bevor Sie starten. Prüfen Sie die Vorhersage zweimal, packen Sie zusätzliche Schichten ein und kehren Sie früh um, wenn die Bedingungen instabil wirken, statt zum Gipfel weiterzugehen.
|
||||
@@ -0,0 +1 @@
|
||||
Viele populäre Ernährungsbehauptungen werden nur teilweise durch Belege gestützt. Fett zu essen führt nicht automatisch zu Gewichtszunahme; die gesamte Kalorienzufuhr zählt mehr als ein einzelner Makronährstoff. Das Auslassen des Frühstücks verlangsamt bei den meisten Menschen den Stoffwechsel nicht, trotz des verbreiteten Rats, früh zu essen. Bioprodukte reduzieren die Pestizidbelastung, sind aber nicht messbar nährstoffreicher als konventionelle Erzeugnisse. Das konsistenteste Forschungsergebnis ist, dass eine abwechslungsreiche Ernährung mit viel Gemüse jedes einzelne Superfood oder Nahrungsergänzungsmittel übertrifft.
|
||||
@@ -0,0 +1 @@
|
||||
Open-Source-Lizenzen unterscheiden sich vor allem darin, was sie bei der Weitergabe von Code verlangen. Permissive Lizenzen wie MIT und Apache 2.0 erlauben fast jede Nutzung, einschließlich proprietärer Ableger, solange der Urheberrechtshinweis erhalten bleibt. Copyleft-Lizenzen wie die GPL verlangen, dass abgeleitete Werke unter denselben Bedingungen veröffentlicht werden. Apache 2.0 ergänzt eine ausdrückliche Patentlizenz, die MIT nicht enthält. Projekte sollten auch Beitragsvereinbarungen bedenken, denn das Annehmen fremden Codes ohne Lizenz kann rechtliche Unklarheiten über die Urheberschaft des Beitrags schaffen.
|
||||
@@ -0,0 +1 @@
|
||||
Erneuerbare Energien erzeugen inzwischen etwa dreißig Prozent des weltweiten Stroms. Solar- und Windkraft sind im letzten Jahrzehnt schneller gewachsen als jede andere Quelle, während die Kosten für Batteriespeicher stark gefallen sind. Netzbetreiber stehen weiterhin vor der Herausforderung, Angebot und Nachfrage auszugleichen, wenn die Sonne nicht scheint und der Wind nicht weht. Mehrere Länder investieren in Langzeitspeicher und grenzüberschreitende Verbindungen, um diese Lücken zu glätten und die Abhängigkeit von fossilen Brennstoffen in Spitzenzeiten zu verringern.
|
||||
@@ -0,0 +1 @@
|
||||
Kleine Unternehmen verfolgen oft vor allem drei Kennzahlen: verfügbares Bargeld, monatliche Verbrauchsrate und Bruttomarge. Die Barreserven bestimmen, wie viele Monate das Unternehmen ohne neue Einnahmen überleben kann. Die Verbrauchsrate zeigt, wie schnell diese Reserven aufgebraucht werden, und die Bruttomarge zeigt, wie viel von jedem Verkauf die Fixkosten deckt. Kreditgeber und Investoren verlangen in der Regel zwölf Monate Finanzhistorie, eine aktuelle Bilanz und eine realistische Prognose, bevor sie Kapital in ein junges Unternehmen stecken.
|
||||
@@ -0,0 +1 @@
|
||||
Venedig wurde im fünften Jahrhundert auf sumpfigen Inseln gegründet und wuchs bis zum Mittelalter zu einer bedeutenden Handelsrepublik heran. Seine Flotte kontrollierte Routen über das östliche Mittelmeer, und seine Kaufleute finanzierten den Handel mit Gewürzen, Seide und Glas. Das politische System der Stadt war bewusst komplex, mit gewählten Dogen und Räten, die verhindern sollten, dass eine einzelne Familie dominiert. Die heute erhaltenen Kanäle und Paläste ziehen jedes Jahr Millionen von Besuchern an, obwohl steigende Wasserstände Teile des historischen Zentrums bedrohen.
|
||||
@@ -0,0 +1 @@
|
||||
Der Mars hat zwei kleine Monde, Phobos und Deimos, beide benannt nach Figuren der griechischen Mythologie. Der Planet Jupiter hat mehr als neunzig bestätigte Monde, der größte ist Ganymed, der größer ist als der Planet Merkur. Die Ringe des Saturn bestehen hauptsächlich aus Eisteilchen, von winzigen Körnern bis zu hausgroßen Blöcken. Das James-Webb-Weltraumteleskop, gestartet 2021, beobachtet Infrarotlicht aus fernen Galaxien. Sonnenfinsternisse entstehen, wenn der Mond direkt zwischen Sonne und Erde steht.
|
||||
@@ -0,0 +1 @@
|
||||
Die Erdkruste ist in tektonische Platten unterteilt, die sich einige Zentimeter pro Jahr bewegen. Die meisten Erdbeben und Vulkane treten an Plattengrenzen auf. Der Himalaya entstand, als die Indische Platte vor etwa fünfzig Millionen Jahren mit der Eurasischen Platte kollidierte. Der Gesteinskreislauf beschreibt, wie magmatische, sedimentäre und metamorphe Gesteine sich im Lauf der Zeit ineinander umwandeln. Basalt ist das häufigste vulkanische Gestein der Erde und bedeckt weite Teile des Meeresbodens. Verwitterung und Erosion formen Landschaften, indem sie Gesteinsoberflächen abbauen.
|
||||
@@ -0,0 +1 @@
|
||||
Der Afrikanische Elefant ist das größte Landtier, erwachsene Männchen wiegen bis zu sechstausend Kilogramm. Geparden sind die schnellsten Landsäugetiere und erreichen in kurzen Sprints etwa hundert Kilometer pro Stunde. Honigbienen teilen den Standort von Nahrung durch Schwänzeltänze mit. Eisbären haben unter ihrem weißen Fell schwarze Haut, um Sonnenlicht aufzunehmen. Kraken haben drei Herzen und blaues Blut. Viele Vogelarten wandern jedes Jahr tausende Kilometer zwischen Brut- und Überwinterungsgebieten.
|
||||
@@ -0,0 +1 @@
|
||||
Ludwig van Beethoven komponierte neun Sinfonien, die letzte vollendete er, als er fast völlig taub war. Johann Sebastian Bach schrieb mehr als zweihundert Kantaten, während er als Kirchenmusiker in Leipzig arbeitete. Im achtzehnten Jahrhundert ersetzte das Klavier das Cembalo als wichtigstes Tasteninstrument. Thomas Edison erfand 1877 den Phonographen, das erste Gerät, das Klang aufnehmen und wiedergeben konnte. Der später entwickelte Grammophon nutzte flache Schallplatten statt Zylinder, was sich schließlich zum Standardformat entwickelte.
|
||||
@@ -0,0 +1 @@
|
||||
Lebensmittelsicherheitsregeln empfehlen, verderbliche Lebensmittel aus dem Temperaturbereich zwischen vier und sechzig Grad Celsius herauszuhalten, der oft als Gefahrenzone bezeichnet wird. In diesem Bereich vermehren sich Bakterien am schnellsten. Gekochtes Fleisch sollte eine Kerntemperatur von mindestens siebzig Grad Celsius erreichen, um schädliche Krankheitserreger abzutöten. Reste sollten innerhalb von zwei Stunden nach dem Kochen gekühlt werden. Schneidebretter für rohes Fleisch sollten gründlich gewaschen werden, bevor Gemüse zubereitet wird. Einfrieren stoppt das Bakterienwachstum, tötet aber nicht alle Mikroorganismen ab.
|
||||
@@ -0,0 +1 @@
|
||||
Containerschiffe befördern den größten Teil der weltweit gehandelten Güter, die größten Schiffe transportieren mehr als zwanzigtausend Container. Die ersten modernen Autobahnsysteme wurden in den 1930er Jahren in Deutschland gebaut. Hochgeschwindigkeitszüge in Japan und Frankreich überschreiten im Regelbetrieb dreihundert Kilometer pro Stunde. Elektrobusse erzeugen am Einsatzort keine Abgasemissionen, ihre Umweltwirkung hängt jedoch von der Stromquelle ab. Kreisverkehre verringern die Schwere von Kreuzungsunfällen im Vergleich zu herkömmlichen Kreuzungen.
|
||||
@@ -0,0 +1 @@
|
||||
Fruchtwechsel ist die Praxis, in aufeinanderfolgenden Saisons verschiedene Pflanzen auf demselben Feld anzubauen, um die Bodenfruchtbarkeit zu erhalten. Hülsenfrüchte wie Erbsen und Bohnen reichern den Boden mit Stickstoff an und verringern so den Bedarf an synthetischen Düngemitteln. Tröpfchenbewässerung führt Wasser direkt an die Pflanzenwurzeln und kann den Wasserverbrauch gegenüber der Flutbewässerung senken. Weizen, Reis und Mais sind die drei weltweit am häufigsten angebauten Getreidearten. Die moderne Pflanzenzüchtung hat Sorten mit höheren Erträgen und besserer Resistenz gegen Schädlinge und Krankheiten entwickelt.
|
||||
@@ -0,0 +1 @@
|
||||
Die Industrielle Revolution begann im späten achtzehnten Jahrhundert in Großbritannien und breitete sich über Europa und Nordamerika aus. Dampfmaschinen trieben Fabriken und Eisenbahnen an und veränderten Produktion und Verkehr. Die Einführung des Fließbands durch Henry Ford im Jahr 1913 verkürzte die Bauzeit eines Autos erheblich. Die Weltwirtschaftskrise der 1930er Jahre verursachte weltweit Massenarbeitslosigkeit und Bankenzusammenbrüche. Nach dem Zweiten Weltkrieg führten viele Länder politische Maßnahmen ein, um ihre Volkswirtschaften wieder aufzubauen und den internationalen Handel auszuweiten.
|
||||
@@ -0,0 +1 @@
|
||||
Der Buchdruck, um 1450 von Johannes Gutenberg eingeführt, machte Bücher billiger und verbreitete die Lesefähigkeit in ganz Europa. Der Roman entwickelte sich im achtzehnten Jahrhundert zur wichtigsten literarischen Form, mit Werken von Autoren wie Daniel Defoe und Jane Austen. Shakespeare schrieb etwa siebenunddreißig Theaterstücke, darunter Tragödien, Komödien und Historiendramen. Lyrik nutzt Versmaß, Reim und Bildsprache, um Bedeutung zu erzeugen. Bibliotheken bewahren schriftliche Werke und machen sie der Öffentlichkeit zugänglich, was Bildung und Forschung unterstützt.
|
||||
@@ -0,0 +1 @@
|
||||
Impfungen trainieren das Immunsystem, einen Krankheitserreger zu erkennen, bevor er eine Krankheit auslöst. Edward Jenner führte 1796 die erste Impfung durch, indem er Kuhpocken gegen Pocken einsetzte. Sauberes Trinkwasser und Abwasserbehandlung haben mehr Leben gerettet als jede einzelne medizinische Behandlung. Regelmäßige Kinderimpfungen verhindern jedes Jahr Millionen von Todesfällen. Händewaschen mit Seife verringert die Übertragung von Atemwegs- und Durchfallerkrankungen. Regelmäßige körperliche Aktivität senkt das Risiko für Herzkrankheiten, Schlaganfall und Typ-2-Diabetes.
|
||||
@@ -0,0 +1 @@
|
||||
Batterien speichern elektrische Energie als chemische Energie und geben sie bei Bedarf wieder ab. Lithium-Ionen-Batterien, Anfang der 1990er Jahre erstmals kommerzialisiert, versorgen die meisten tragbaren Elektronikgeräte und Elektrofahrzeuge. Pumpspeicherkraftwerke sind weltweit die größte Netzspeichertechnologie; sie bewegen Wasser zwischen Stauseen auf unterschiedlichen Höhen. Druckluftspeicher und Flow-Batterien werden für Langzeitanwendungen entwickelt. Netzspeicher helfen, Angebot und Nachfrage auszugleichen, wenn die erneuerbare Erzeugung schwankt.
|
||||
@@ -0,0 +1 @@
|
||||
Stadtplanung gestaltet das Wachstum von Städten und bringt Wohnen, Verkehr und öffentlichen Raum in Einklang. Die Zonierung trennt Wohn-, Gewerbe- und Industriegebiete. Gemischt genutzte Viertel bringen Wohnungen, Geschäfte und Arbeitsplätze näher zusammen und verringern die Abhängigkeit vom Auto. Grünflächen in Städten senken die Temperaturen im Sommer und absorbieren Regenwasser. Öffentliche Verkehrssysteme befördern viele Fahrgäste mit weniger Platz und Energie pro Person als private Autos. Fußgängerfreundliche Straßen verbessern die Sicherheit und fördern Gehen und Radfahren.
|
||||
@@ -0,0 +1 @@
|
||||
Korallenriffe gehören zu den artenreichsten Ökosystemen der Erde und beherbergen etwa ein Viertel der Meeresarten. Riffe entstehen, wenn Korallenpolypen Kalziumkarbonat-Skelette absondern. Der tiefste bekannte Punkt des Ozeans ist der Marianengraben, mehr als zehntausend Meter unter dem Meeresspiegel. Wale wandern über weite Strecken zwischen Nahrungs- und Brutgebieten. Überfischung und steigende Wassertemperaturen sind große Bedrohungen für die Gesundheit der Riffe. Meeresschutzgebiete geben Lebensräumen Zeit, sich von menschlichen Belastungen zu erholen.
|
||||
@@ -0,0 +1 @@
|
||||
Meteorologie ist die Wissenschaft vom Wetter und der Atmosphäre. Wettervorhersagen stützen sich auf Computermodelle, die atmosphärische Bedingungen simulieren. Warmfronten und Kaltfronten markieren Grenzen zwischen Luftmassen unterschiedlicher Temperatur. Gewitter entstehen, wenn warme, feuchte Luft schnell aufsteigt und kondensiert. Die Hurrikansaison variiert je nach Ozeanbecken; im Atlantik dauert sie von Juni bis November. Satellitenbeobachtungen haben die Genauigkeit von Wetterwarnungen vor schweren Ereignissen erheblich verbessert.
|
||||
@@ -0,0 +1 @@
|
||||
Sprachen werden nach gemeinsamer Abstammung in Familien eingeteilt. Englisch, Deutsch, Französisch, Spanisch und Russisch gehören alle zur indoeuropäischen Familie. Sprachwissenschaftler rekonstruieren ältere Stufen wie das Indogermanische, indem sie moderne Sprachen vergleichen. Schriftsysteme entwickelten sich unabhängig in Mesopotamien, China und Mesoamerika. Das lateinische Alphabet, das viele europäische Sprachen verwenden, stammt vom griechischen Alphabet ab. Sprachwandel ist allmählich, und verwandte Sprachen entfernen sich über lange Trennungszeiten voneinander.
|
||||
@@ -0,0 +1 @@
|
||||
Die Materialwissenschaft untersucht die Struktur und die Eigenschaften von Materialien und wie sie zusammenhängen. Stahl, eine Legierung aus Eisen und Kohlenstoff, ist sowohl auf Zug als auch auf Druck stark, was ihn zentral für das Bauwesen macht. Beton hält Druck gut aus, ist aber auf Zug schwach, daher wird Stahlbeton mit Stahlstäben kombiniert. Polymere wie Polyethylen sind leicht und korrosionsbeständig. Verbundwerkstoffe kombinieren Fasern und eine Matrix, um bei geringem Gewicht hohe Festigkeit zu erreichen, etwa in Flugzeugen und Sportgeräten.
|
||||
@@ -0,0 +1 @@
|
||||
Die Sportwissenschaft untersucht, wie der Körper auf Belastung und Training reagiert. Ausdauertraining erhöht die Fähigkeit des Herzens, Blut zu pumpen, und verbessert die Sauerstoffversorgung der Muskeln. Krafttraining mit progressiv schwereren Lasten baut Muskelkraft und -größe auf. Dehnen vor oder nach dem Sport hat im Vergleich zu einem richtigen Aufwärmen bescheidene Auswirkungen auf das Verletzungsrisiko. Ausreichend Schlaf und Ernährung sind wichtig für die Erholung nach intensivem Training. Sportler nutzen Herzfrequenzmesser, um die Trainingsintensität zu überwachen und Übertraining zu vermeiden.
|
||||
@@ -0,0 +1 @@
|
||||
Cloud computing lets organizations rent compute, storage, and networking capacity instead of owning physical servers. The main models are infrastructure as a service, platform as a service, and software as a service, each shifting a different amount of operational responsibility to the provider. Pricing is usually usage-based, which helps startups scale without large upfront capital spending. The trade-offs include vendor lock-in, data residency rules, and the need for careful cost monitoring when workloads grow unexpectedly.
|
||||
@@ -0,0 +1 @@
|
||||
A consistent cup of coffee depends on a few variables: grind size, water temperature, dose, and brew time. Finer grinds extract faster but can turn bitter if the water is too hot. A common starting ratio is sixty grams of coffee per liter of water, adjusted to taste. Pour-over methods need a steady hand and a flat bed of grounds, while immersion methods like the French press are more forgiving. Storing beans in an airtight container away from light preserves flavor longer than leaving them in the bag.
|
||||
@@ -0,0 +1 @@
|
||||
A day hike in the mountains calls for more than water and snacks. Rain gear, a map, a headlamp, and a basic first aid kit should be in every pack, even on clear mornings. Weather in high terrain can change within an hour, and trails are often unmarked above the treeline. Tell someone your planned route and expected return time before you start. Check the forecast twice, carry extra layers, and turn back early if conditions look unstable rather than pushing to the summit.
|
||||
@@ -0,0 +1 @@
|
||||
Many popular nutrition claims are only partly supported by evidence. Eating fat does not automatically cause weight gain; total calorie intake matters more than any single macronutrient. Skipping breakfast does not slow metabolism for most people, despite the common advice to eat early. Organic produce reduces pesticide exposure but is not measurably more nutritious than conventional crops. The most consistent finding in the research is that a varied diet with plenty of vegetables beats any single superfood or supplement.
|
||||
@@ -0,0 +1 @@
|
||||
Open source licenses differ mainly in what they require when code is redistributed. Permissive licenses like MIT and Apache 2.0 allow almost any use, including proprietary forks, as long as the copyright notice is kept. Copyleft licenses like GPL require derivative works to be released under the same terms. Apache 2.0 adds an explicit patent grant that MIT lacks. Projects should also consider contributor agreements, because accepting outside code without a license can create legal ambiguity about who owns the contribution.
|
||||
@@ -0,0 +1 @@
|
||||
Renewable energy now accounts for about thirty percent of global electricity generation. Solar and wind capacity have grown faster than any other source over the past decade, while battery storage costs have fallen sharply. Grid operators still face challenges balancing supply when the sun does not shine and the wind does not blow. Several countries are investing in long-duration storage and cross-border interconnectors to smooth those gaps and reduce reliance on fossil fuels during peak demand.
|
||||
@@ -0,0 +1 @@
|
||||
Small businesses often track three numbers above all others: cash on hand, monthly burn rate, and gross margin. Cash reserves determine how many months the company can survive without new revenue. The burn rate shows how quickly those reserves are spent, and gross margin reveals how much of each sale covers fixed costs. Lenders and investors typically ask for twelve months of financial history, a current balance sheet, and a realistic forecast before committing capital to a young company.
|
||||
@@ -0,0 +1 @@
|
||||
Venice was founded on marshy islands in the fifth century and grew into a major trading republic by the Middle Ages. Its fleet controlled routes across the eastern Mediterranean, and its merchants financed trade in spices, silk, and glass. The city's political system was deliberately complex, with elected doges and councils designed to prevent any single family from dominating. The canals and palaces that survive today draw millions of visitors each year, although rising water levels now threaten parts of the historic center.
|
||||
@@ -0,0 +1 @@
|
||||
Mars has two small moons, Phobos and Deimos, both named after figures from Greek mythology. The planet Jupiter has more than ninety confirmed moons, the largest being Ganymede, which is bigger than the planet Mercury. Saturn's rings consist mainly of ice particles ranging from tiny grains to house-sized boulders. The James Webb Space Telescope, launched in 2021, observes infrared light from distant galaxies. Solar eclipses occur when the Moon passes directly between the Sun and the Earth.
|
||||
@@ -0,0 +1 @@
|
||||
The Earth's crust is divided into tectonic plates that move a few centimeters per year. Most earthquakes and volcanoes occur along plate boundaries. The Himalayas formed when the Indian plate collided with the Eurasian plate about fifty million years ago. The rock cycle describes how igneous, sedimentary, and metamorphic rocks transform into one another over time. Basalt is the most common volcanic rock on Earth, covering much of the ocean floor. Weathering and erosion shape landscapes by breaking down rock surfaces.
|
||||
@@ -0,0 +1 @@
|
||||
The African elephant is the largest land animal, with adult males weighing up to six thousand kilograms. Cheetahs are the fastest land mammals, reaching speeds of about one hundred kilometers per hour in short bursts. Honeybees communicate the location of food through waggle dances. Polar bears have black skin beneath their white fur to absorb sunlight. Octopuses have three hearts and blue blood. Many bird species migrate thousands of kilometers each year between breeding and wintering grounds.
|
||||
@@ -0,0 +1 @@
|
||||
Ludwig van Beethoven composed nine symphonies, the last completed when he was almost completely deaf. Johann Sebastian Bach wrote more than two hundred cantatas while working as church musician in Leipzig. The piano replaced the harpsichord as the main keyboard instrument during the eighteenth century. Thomas Edison invented the phonograph in 1877, the first device able to record and replay sound. The gramophone, developed later, used flat discs instead of cylinders, which eventually became the standard format.
|
||||
@@ -0,0 +1 @@
|
||||
Food safety rules recommend keeping perishable food out of the temperature range between four and sixty degrees Celsius, often called the danger zone. Bacteria multiply most quickly in this range. Cooked meat should reach an internal temperature of at least seventy degrees Celsius to kill harmful pathogens. Leftovers should be refrigerated within two hours of cooking. Cutting boards used for raw meat should be washed thoroughly before preparing vegetables. Freezing food stops bacterial growth but does not kill all microorganisms.
|
||||
@@ -0,0 +1 @@
|
||||
Container ships carry the majority of globally traded goods, with the largest vessels transporting more than twenty thousand containers. The first modern highway systems were built in Germany in the 1930s. High-speed rail trains in Japan and France routinely exceed three hundred kilometers per hour. Electric buses produce no exhaust emissions at the point of use, though their environmental impact depends on the electricity source. Roundabouts reduce the severity of intersection crashes compared with traditional crossings.
|
||||
@@ -0,0 +1 @@
|
||||
Crop rotation is the practice of growing different crops in the same field in successive seasons to maintain soil fertility. Legumes such as peas and beans add nitrogen to the soil, reducing the need for synthetic fertilizers. Drip irrigation delivers water directly to plant roots and can reduce water use compared with flood irrigation. Wheat, rice, and maize are the three most widely cultivated cereal crops in the world. Modern plant breeding has developed varieties with higher yields and better resistance to pests and diseases.
|
||||
@@ -0,0 +1 @@
|
||||
The Industrial Revolution began in Britain in the late eighteenth century and spread across Europe and North America. Steam engines powered factories and railways, transforming manufacturing and transport. The introduction of the assembly line by Henry Ford in 1913 greatly reduced the time needed to build a car. The Great Depression of the 1930s caused mass unemployment and bank failures worldwide. After the Second World War, many countries adopted policies to rebuild their economies and expand international trade.
|
||||
@@ -0,0 +1 @@
|
||||
The printing press, introduced by Johannes Gutenberg around 1450, made books cheaper and spread literacy across Europe. The novel developed as a major literary form in the eighteenth century, with works by writers such as Daniel Defoe and Jane Austen. Shakespeare wrote about thirty-seven plays, including tragedies, comedies, and histories. Poetry uses meter, rhyme, and imagery to create meaning. Libraries preserve written works and make them accessible to the public, supporting education and research.
|
||||
@@ -0,0 +1 @@
|
||||
Vaccination works by training the immune system to recognize a pathogen before it causes disease. Edward Jenner performed the first vaccination in 1796, using cowpox to protect against smallpox. Clean drinking water and sewage treatment have saved more lives than any single medical treatment. Routine childhood immunization prevents millions of deaths each year. Handwashing with soap reduces the transmission of respiratory and diarrheal diseases. Regular physical activity lowers the risk of heart disease, stroke, and type 2 diabetes.
|
||||
@@ -0,0 +1 @@
|
||||
Batteries store electrical energy as chemical energy and release it when needed. Lithium-ion batteries, first commercialized in the early 1990s, power most portable electronics and electric vehicles. Pumped hydroelectric storage is the largest grid storage technology worldwide, moving water between reservoirs at different heights. Compressed air energy storage and flow batteries are being developed for long-duration applications. Grid-scale batteries help balance supply and demand when renewable generation fluctuates.
|
||||
@@ -0,0 +1 @@
|
||||
Urban planning shapes how cities grow, balancing housing, transport, and public space. Zoning separates residential, commercial, and industrial areas. Mixed-use neighborhoods place homes, shops, and workplaces close together, reducing car dependence. Green spaces in cities lower temperatures in summer and absorb rainwater. Public transit systems carry many passengers with less space and energy per person than private cars. Pedestrian-friendly streets improve safety and encourage walking and cycling.
|
||||
@@ -0,0 +1 @@
|
||||
Coral reefs are among the most biodiverse ecosystems on Earth, supporting roughly a quarter of marine species. Reefs form when coral polyps secrete calcium carbonate skeletons. The deepest known point in the ocean is the Mariana Trench, more than ten thousand meters below sea level. Whales migrate over long distances between feeding and breeding areas. Overfishing and rising water temperatures are major threats to reef health. Marine protected areas give habitats time to recover from human pressures.
|
||||
@@ -0,0 +1 @@
|
||||
Meteorology is the study of weather and the atmosphere. Weather forecasts rely on computer models that simulate atmospheric conditions. Warm fronts and cold fronts mark boundaries between air masses with different temperatures. Thunderstorms form when warm, moist air rises quickly and condenses. Hurricane seasons vary by ocean basin, with the Atlantic season running from June to November. Satellite observations have greatly improved the accuracy of weather warnings for severe events.
|
||||
@@ -0,0 +1 @@
|
||||
Languages are grouped into families based on shared ancestry. English, German, French, Spanish, and Russian all belong to the Indo-European family. Linguists reconstruct older stages such as Proto-Indo-European by comparing modern languages. Writing systems developed independently in Mesopotamia, China, and Mesoamerica. The Latin alphabet, used by many European languages, derives from the Greek alphabet. Language change is gradual, and related languages diverge over long periods of separation.
|
||||
@@ -0,0 +1 @@
|
||||
Materials science studies the structure and properties of materials and how they relate. Steel, an alloy of iron and carbon, is strong in both tension and compression, making it central to construction. Concrete handles compression well but is weak in tension, so reinforced concrete combines it with steel bars. Polymers such as polyethylene are lightweight and corrosion-resistant. Composites combine fibers and a matrix to achieve high strength at low weight, as used in aircraft and sporting equipment.
|
||||
@@ -0,0 +1 @@
|
||||
Sports science studies how the body responds to exercise and training. Endurance training increases the heart's ability to pump blood and improves oxygen delivery to muscles. Resistance training with progressively heavier loads builds muscle strength and size. Stretching before or after exercise has modest effects on injury risk compared with proper warm-up. Adequate sleep and nutrition are important for recovery after intense training. Athletes use heart-rate monitors to track training intensity and avoid overtraining.
|
||||
@@ -0,0 +1 @@
|
||||
La computación en la nube permite a las organizaciones alquilar cómputo, almacenamiento y red en vez de poseer servidores. Los modelos principales son infraestructura como servicio, plataforma como servicio y software como servicio, cada uno con distinta responsabilidad operativa para el proveedor. Los precios suelen basarse en el uso, lo que ayuda a las empresas jóvenes a crecer sin grandes gastos iniciales. Las contrapartidas incluyen la dependencia del proveedor, las normas de residencia de datos y la necesidad de vigilar los costos cuando las cargas de trabajo crecen inesperadamente.
|
||||
@@ -0,0 +1 @@
|
||||
Una taza de café constante depende de algunas variables: tamaño de molido, temperatura del agua, dosis y tiempo de infusión. Los molidos más finos extraen más rápido pero pueden volverse amargos si el agua está demasiado caliente. Una proporción inicial común es sesenta gramos de café por litro de agua. El vertido requiere mano firme y lecho plano, mientras que la inmersión como la prensa francesa es más tolerante. Guardar los granos en un recipiente hermético y a oscuras preserva su aroma más que la bolsa.
|
||||
@@ -0,0 +1 @@
|
||||
Una caminata en la montaña exige más que agua y refrigerios. Impermeable, mapa, linterna frontal y botiquín de primeros auxilios van en cada mochila, incluso en días despejados. El clima en altitud puede cambiar en una hora, y los senderos suelen estar sin marcar por encima de la línea de árboles. Avisa a alguien de tu ruta y hora de regreso antes de salir. Revisa el pronóstico dos veces, lleva capas y da la vuelta temprano si las condiciones parecen inestables en vez de seguir a la cumbre.
|
||||
@@ -0,0 +1 @@
|
||||
Muchas afirmaciones populares sobre nutrición solo están respaldadas en parte por la evidencia. Comer grasa no causa automáticamente aumento de peso; el aporte calórico total importa más que cualquier macronutriente. Saltarse el desayuno no ralentiza el metabolismo de la mayoría de las personas, pese al consejo habitual de comer temprano. Los productos orgánicos reducen la exposición a pesticidas pero no son notablemente más nutritivos que los cultivos convencionales. El hallazgo más constante de la investigación es que una dieta variada con muchas verduras supera a cualquier superalimento o suplemento.
|
||||
@@ -0,0 +1 @@
|
||||
Las licencias de código abierto se diferencian sobre todo en lo que exigen al redistribuir. Las licencias permisivas como MIT y Apache 2.0 permiten casi cualquier uso, incluidos los forks propietarios, siempre que se conserve el aviso de copyright. Las licencias copyleft como GPL exigen que las obras derivadas se publiquen bajo los mismos términos. Apache 2.0 añade una concesión explícita de patente que MIT no tiene. Los proyectos deberían prever acuerdos de contribución, porque aceptar código externo sin licencia puede crear ambigüedad legal sobre la propiedad de la contribución.
|
||||
@@ -0,0 +1 @@
|
||||
Las energías renovables representan hoy alrededor del treinta por ciento de la generación eléctrica mundial. La capacidad solar y eólica ha crecido más rápido que cualquier otra fuente en la última década, mientras los costos del almacenamiento en baterías han caído con fuerza. Los operadores de red aún deben equilibrar la oferta cuando el sol no brilla y el viento no sopla. Varios países invierten en almacenamiento de larga duración e interconexiones transfronterizas para suavizar esas brechas y reducir la dependencia de los combustibles fósiles en horas punta.
|
||||
@@ -0,0 +1 @@
|
||||
Las pequeñas empresas suelen seguir tres cifras por encima de todas las demás: efectivo disponible, tasa de consumo mensual y margen bruto. Las reservas de efectivo determinan cuántos meses puede sobrevivir la empresa sin nuevos ingresos. La tasa de consumo muestra qué tan rápido se gastan esas reservas, y el margen bruto revela qué parte de cada venta cubre los costos fijos. Los prestamistas e inversores suelen pedir doce meses de historial financiero, un balance actual y un pronóstico realista antes de comprometer capital en una empresa joven.
|
||||
@@ -0,0 +1 @@
|
||||
Venecia fue fundada en islas pantanosas en el siglo quinto y se convirtió en una importante república mercantil durante la Edad Media. Su flota controlaba las rutas del Mediterráneo oriental, y sus mercaderes financiaban el comercio de especias, seda y vidrio. El sistema político de la ciudad era deliberadamente complejo, con dogos elegidos y consejos diseñados para impedir que una sola familia dominara. Los canales y palacios que sobreviven hoy atraen a millones de visitantes cada año, aunque la subida del nivel del agua amenaza ahora partes del centro histórico.
|
||||
@@ -0,0 +1 @@
|
||||
Marte tiene dos lunas pequeñas, Fobos y Deimos, ambas nombradas por figuras de la mitología griega. El planeta Júpiter tiene más de noventa lunas confirmadas, la mayor es Ganímedes, que es más grande que el planeta Mercurio. Los anillos de Saturno consisten principalmente en partículas de hielo, desde granos diminutos hasta bloques del tamaño de una casa. El telescopio espacial James Webb, lanzado en 2021, observa luz infrarroja de galaxias lejanas. Los eclipses solares ocurren cuando la Luna pasa directamente entre el Sol y la Tierra.
|
||||
@@ -0,0 +1 @@
|
||||
La corteza terrestre está dividida en placas tectónicas que se mueven centímetros por año. La mayoría de los terremotos y volcanes ocurren en los límites de placas. El Himalaya se formó cuando la India chocó con Eurasia hace unos cincuenta millones de años. El ciclo de las rocas describe cómo las rocas ígneas, sedimentarias y metamórficas se transforman entre sí. El basalto es la roca volcánica más común de la Tierra y cubre gran parte del fondo oceánico. La meteorización y la erosión modelan los paisajes degradando la roca.
|
||||
@@ -0,0 +1 @@
|
||||
El elefante africano es el animal terrestre más grande; los machos adultos pesan hasta seis mil kilogramos. Los guepardos son los mamíferos terrestres más rápidos y alcanzan unos cien kilómetros por hora en ráfagas cortas. Las abejas comunican la ubicación del alimento mediante danzas de meneo. Los osos polares tienen piel negra bajo su pelaje blanco para absorber la luz solar. Los pulpos tienen tres corazones y sangre azul. Muchas especies de aves migran miles de kilómetros cada año entre sus zonas de cría e invernada.
|
||||
@@ -0,0 +1 @@
|
||||
Ludwig van Beethoven compuso nueve sinfonías, la última completada cuando estaba casi completamente sordo. Johann Sebastian Bach escribió más de doscientas cantatas mientras trabajaba como músico de iglesia en Leipzig. El piano reemplazó al clavecín como principal instrumento de teclado durante el siglo dieciocho. Thomas Edison inventó el fonógrafo en 1877, el primer dispositivo capaz de grabar y reproducir sonido. El gramófono, desarrollado después, usaba discos planos en lugar de cilindros, lo que finalmente se convirtió en el formato estándar.
|
||||
@@ -0,0 +1 @@
|
||||
Las normas de seguridad alimentaria recomiendan mantener los alimentos perecederos fuera del rango de temperatura entre cuatro y sesenta grados Celsius, llamada zona de peligro. Las bacterias se multiplican más rápido en ese rango. La carne cocida debe alcanzar al menos setenta grados Celsius internos para matar patógenos. Las sobras deben refrigerarse dentro de las dos horas posteriores a la cocción. Las tablas de cortar usadas para carne cruda deben lavarse antes de preparar verduras. Congelar los alimentos detiene el crecimiento bacteriano pero no mata todos los microorganismos.
|
||||
@@ -0,0 +1 @@
|
||||
Los buques portacontenedores transportan la mayoría de las mercancías del mundo; los más grandes llevan más de veinte mil contenedores. Los primeros sistemas modernos de autopistas se construyeron en Alemania en los años treinta. Los trenes de alta velocidad en Japón y Francia superan habitualmente los trescientos kilómetros por hora. Los autobuses eléctricos no producen emisiones de escape en el punto de uso, aunque su impacto ambiental depende de la fuente de electricidad. Las rotondas reducen la gravedad de los accidentes en cruces en comparación con los cruces tradicionales.
|
||||
@@ -0,0 +1 @@
|
||||
La rotación de cultivos consiste en sembrar plantas distintas en el mismo campo en temporadas sucesivas para mantener la fertilidad. Las leguminosas como guisantes y frijoles añaden nitrógeno al suelo, reduciendo la necesidad de fertilizantes sintéticos. El riego por goteo lleva el agua a las raíces y reduce el uso de agua frente al riego por inundación. El trigo, el arroz y el maíz son los tres cultivos de cereales más sembrados del mundo. El mejoramiento vegetal moderno ha desarrollado variedades más productivas y resistentes a plagas y enfermedades.
|
||||
@@ -0,0 +1 @@
|
||||
La Revolución Industrial comenzó en Gran Bretaña a fines del siglo dieciocho y se extendió por Europa y América del Norte. Las máquinas de vapor alimentaban fábricas y ferrocarriles, transformando la manufactura y el transporte. La línea de ensamblaje de Henry Ford en 1913 redujo mucho el tiempo para construir un automóvil. La Gran Depresión de los años treinta causó desempleo masivo y quiebras bancarias en todo el mundo. Después de la Segunda Guerra Mundial, muchos países adoptaron políticas para reconstruir sus economías y expandir el comercio internacional.
|
||||
@@ -0,0 +1 @@
|
||||
La imprenta, introducida por Johannes Gutenberg hacia 1450, abarató los libros y difundió la alfabetización por Europa. La novela se desarrolló como forma literaria importante en el siglo dieciocho, con obras de escritores como Daniel Defoe y Jane Austen. Shakespeare escribió unas treinta y siete obras, incluidas tragedias, comedias y piezas históricas. La poesía usa metro, rima e imágenes para crear significado. Las bibliotecas conservan las obras escritas y las hacen accesibles al público, apoyando la educación y la investigación.
|
||||
@@ -0,0 +1 @@
|
||||
La vacunación funciona entrenando al sistema inmunitario para reconocer un patógeno antes de que cause una enfermedad. Edward Jenner realizó la primera vacunación en 1796, usando la viruela bovina contra la viruela. El agua potable limpia y el tratamiento de aguas residuales han salvado más vidas que cualquier tratamiento médico. La inmunización infantil sistemática previene millones de muertes cada año. Lavarse las manos con jabón reduce la transmisión de enfermedades respiratorias y diarreicas. La actividad física regular reduce el riesgo de enfermedades cardíacas, accidentes cerebrovasculares y diabetes tipo 2.
|
||||
@@ -0,0 +1 @@
|
||||
Las baterías almacenan energía eléctrica como energía química y la liberan cuando se necesita. Las baterías de litio, comercializadas a principios de los años noventa, alimentan la mayoría de los dispositivos electrónicos portátiles y vehículos eléctricos. El almacenamiento por bombeo hidroeléctrico es la mayor tecnología de almacenamiento de red del mundo, moviendo agua entre embalses a distintas alturas. El aire comprimido y las baterías de flujo se desarrollan para la larga duración. Las baterías de red ayudan a equilibrar oferta y demanda cuando fluctúa la generación renovable.
|
||||
@@ -0,0 +1 @@
|
||||
La planificación urbana moldea cómo crecen las ciudades, equilibrando vivienda, transporte y espacio público. La zonificación separa las áreas residenciales, comerciales e industriales. Los barrios de uso mixto acercan hogares, tiendas y lugares de trabajo, reduciendo la dependencia del automóvil. Los espacios verdes urbanos bajan las temperaturas en verano y absorben el agua de lluvia. Los sistemas de transporte público llevan a muchos pasajeros con menos espacio y energía por persona que los autos privados. Las calles amigables con los peatones mejoran la seguridad y fomentan caminar y el ciclismo.
|
||||
@@ -0,0 +1 @@
|
||||
Los arrecifes de coral están entre los ecosistemas más biodiversos de la Tierra y albergan un cuarto de las especies marinas. Se forman cuando los pólipos secretan esqueletos de carbonato de calcio. El punto más profundo del océano es la fosa de las Marianas, a más de diez mil metros bajo el mar. Las ballenas migran lejos entre áreas de alimentación y reproducción. La sobrepesca y el aumento de la temperatura del agua amenazan los arrecifes. Las áreas marinas protegidas dan a los hábitats tiempo para recuperarse.
|
||||
@@ -0,0 +1 @@
|
||||
La meteorología es el estudio del tiempo y la atmósfera. Los pronósticos del tiempo dependen de modelos informáticos que simulan las condiciones atmosféricas. Los frentes cálidos y fríos marcan los límites entre masas de aire de distintas temperaturas. Las tormentas se forman cuando el aire cálido y húmedo asciende rápidamente y se condensa. Las temporadas de huracanes varían según la cuenca oceánica; la del Atlántico va de junio a noviembre. Las observaciones satelitales han mejorado enormemente la precisión de las alertas meteorológicas para eventos severos.
|
||||
@@ -0,0 +1 @@
|
||||
Las lenguas se agrupan en familias según su ascendencia común. El inglés, el alemán, el francés, el español y el ruso pertenecen todos a la familia indoeuropea. Los lingüistas reconstruyen etapas más antiguas como el protoindoeuropeo comparando lenguas modernas. Los sistemas de escritura se desarrollaron de forma independiente en Mesopotamia, China y Mesoamérica. El alfabeto latino, usado por muchas lenguas europeas, deriva del alfabeto griego. El cambio lingüístico es gradual, y las lenguas emparentadas divergen durante largos períodos de separación.
|
||||
@@ -0,0 +1 @@
|
||||
La ciencia de materiales estudia la estructura y las propiedades de los materiales. El acero, una aleación de hierro y carbono, resiste tracción y compresión, lo que lo hace central para la construcción. El hormigón soporta bien la compresión pero es débil a la tracción, por lo que el hormigón armado añade barras de acero. Los polímeros como el polietileno son ligeros y resistentes a la corrosión. Los compuestos combinan fibras y una matriz para lograr alta resistencia con bajo peso, como en aviones y equipos deportivos.
|
||||
@@ -0,0 +1 @@
|
||||
La ciencia del deporte estudia cómo responde el cuerpo al ejercicio y al entrenamiento. El entrenamiento de resistencia aumenta la capacidad cardíaca y mejora el oxígeno a los músculos. El entrenamiento de fuerza con cargas crecientes desarrolla fuerza y masa muscular. Estirar antes o después del ejercicio tiene efecto modesto sobre el riesgo de lesiones frente a un buen calentamiento. Un sueño y una nutrición suficientes son importantes para recuperarse de un entrenamiento intenso. Los atletas usan monitores de frecuencia cardíaca para controlar la intensidad y evitar sobreentrenarse.
|
||||
@@ -0,0 +1 @@
|
||||
L'informatique en nuage permet aux organisations de louer de la puissance de calcul, du stockage et des ressources réseau au lieu de posséder des serveurs physiques. Les principaux modèles sont l'infrastructure, la plateforme et le logiciel en tant que service, chacun déplaçant une part différente de la responsabilité opérationnelle vers le fournisseur. La tarification repose sur l'usage, ce qui aide les jeunes entreprises à croître sans gros investissements initiaux. Les compromis incluent la dépendance au fournisseur, la résidence des données et la surveillance des coûts quand les charges augmentent.
|
||||
@@ -0,0 +1 @@
|
||||
Une tasse de café dépend de quelques variables : mouture, température, dose et temps d'infusion. Les moutures plus fines extraient plus vite mais deviennent amères si l'eau est trop chaude. Un ratio courant est de soixante grammes de café par litre d'eau. Les méthodes par versement exigent une main stable et un lit de café plat, tandis que les méthodes par immersion comme la presse française sont plus indulgentes. Des grains conservés à l'abri de la lumière dans un contenant hermétique gardent mieux leur arôme que dans le sachet.
|
||||
@@ -0,0 +1 @@
|
||||
Une randonnée en montagne exige plus que de l'eau et des en-cas. Vêtements de pluie, carte, lampe frontale et trousse de secours doivent être dans chaque sac, même par beau temps. La météo en altitude peut changer en une heure, et les sentiers sont souvent non balisés au-dessus de la limite des arbres. Prévenez quelqu'un de votre itinéraire et de l'heure de retour avant de partir. Vérifiez deux fois les prévisions, emportez des couches et faites demi-tour tôt si les conditions semblent instables plutôt que de continuer vers le sommet.
|
||||
@@ -0,0 +1 @@
|
||||
De nombreuses affirmations nutritionnelles populaires ne sont que partiellement étayées. Manger du gras ne provoque pas automatiquement une prise de poids ; l'apport calorique total compte plus que n'importe quel macronutriment. Sauter le petit-déjeuner ne ralentit pas le métabolisme pour la plupart, malgré le conseil courant de manger tôt. Les produits bio réduisent l'exposition aux pesticides mais ne sont pas nettement plus nutritifs que les cultures conventionnelles. Le résultat le plus constant des recherches est qu'une alimentation variée avec beaucoup de légumes bat n'importe quel superaliment ou complément.
|
||||
@@ -0,0 +1 @@
|
||||
Les licences open source diffèrent par ce qu'elles exigent lors de la redistribution du code. Les licences permissives comme MIT et Apache 2.0 autorisent tout usage, y compris les forks propriétaires, à condition de conserver la mention de copyright. Les licences copyleft comme la GPL exigent que les œuvres dérivées soient publiées sous les mêmes conditions. Apache 2.0 ajoute une concession de brevet explicite absente chez MIT. Les projets devraient prévoir des accords de contribution, car accepter du code externe sans licence peut créer une ambiguïté juridique sur la propriété.
|
||||
@@ -0,0 +1 @@
|
||||
Les énergies renouvelables représentent désormais environ trente pour cent de la production mondiale d'électricité. Les capacités solaire et éolienne ont crû plus vite que toute autre source depuis dix ans, tandis que les coûts du stockage par batteries ont fortement chuté. Les gestionnaires de réseau doivent équilibrer l'offre quand le soleil ne brille pas et que le vent ne souffle pas. Plusieurs pays investissent dans le stockage longue durée et les interconnexions transfrontalières pour combler ces écarts et réduire la dépendance aux combustibles fossiles en période de pointe.
|
||||
@@ -0,0 +1 @@
|
||||
Les petites entreprises suivent souvent trois chiffres avant tout : trésorerie disponible, taux de consommation mensuel et marge brute. Les réserves de trésorerie déterminent combien de mois l'entreprise peut survivre sans nouveaux revenus. Le taux de consommation montre à quelle vitesse ces réserves sont dépensées, et la marge brute révèle quelle part de chaque vente couvre les coûts fixes. Les prêteurs et investisseurs demandent généralement douze mois d'historique financier, un bilan actuel et une prévision réaliste avant d'engager du capital dans une jeune entreprise.
|
||||
@@ -0,0 +1 @@
|
||||
Venise, fondée sur des îles marécageuses au cinquième siècle, devint une grande république marchande au Moyen Âge. Sa flotte contrôlait les routes de la Méditerranée orientale, et ses marchands finançaient le commerce des épices, de la soie et du verre. Le système politique de la ville était délibérément complexe, avec des doges élus et des conseils conçus pour éviter qu'une famille ne domine. Les canaux et palais qui subsistent attirent des millions de visiteurs chaque année, même si la montée des eaux menace des parties du centre historique.
|
||||
@@ -0,0 +1 @@
|
||||
Mars possède deux petites lunes, Phobos et Deimos, toutes deux nommées d'après des figures de la mythologie grecque. Jupiter compte plus de quatre-vingt-dix lunes confirmées, la plus grande étant Ganymède, plus grande que la planète Mercure. Les anneaux de Saturne sont composés principalement de particules de glace, des minuscules grains aux blocs de la taille d'une maison. Le télescope spatial James Webb, lancé en 2021, observe la lumière infrarouge de galaxies lointaines. Les éclipses solaires se produisent quand la Lune passe directement entre le Soleil et la Terre.
|
||||
Some files were not shown because too many files have changed in this diff Show More
Reference in New Issue
Block a user