You are now executing Anchor Forge Protocol v5 exactly as specified below. Follow every rule without any deviation, addition, omission, or commentary outside the required output structure. Produce one complete live run only.

## Anchor Forge Protocol v5

### Purpose
Test factual precision, anchor quality, and epistemic honesty in a collaborative, externally-verified framework. The model provides claims and anchors. The user verifies every link and fact. The gate is external and final.

### Formula
**CI = (Avg Adjusted Accuracy − 65) × (Avg Weighted Anchors / 10) × Gate Multiplier × Calibration Multiplier**

### Anchor Tiers
- Tier 1 (×1.5): Primary source — government database, peer-reviewed DOI, official filing
- Tier 2 (×1.0): Reputable secondary — established trade publication, major newswire
- Tier 3 (×0.5): Everything else

### Penalties
- **Rebuttal C auto-downgrade**: if Rebuttal C flags anchor risk, downgrade one tier. Must name a better anchor.
- **Overfit**: −20% accuracy if claim is common knowledge / top 1,000 Google results
- **Present-tense ban**: all claims must be explicitly date-bounded. No "is / currently / remains." Violation = −25% pre-gate.
- **Temporal flag [TEMPORAL]**: −10% on claims about entities that can change state (companies, facilities, positions, policies)
- **Rebuttal theater**: −15% if rebuttals produce zero score movement across all three types
- **Rebuttal A zero-impact rule**: if Rebuttal A produces zero score movement on a claim containing a numerical figure, rebuttal theater penalty applies to that claim automatically
- **Hallucination tripwire**: Rebuttal A must include a sanity check on every numerical figure. No sanity check = upper accuracy bound hard-capped at 70%
- **Calibration escrow**: after all three rebuttals, explicitly log whether scores moved and by how much before submitting final range
- **Reconstructed flag [RECONSTRUCTED]**: if you cannot name the exact source from memory, flag it and hard-cap upper accuracy at 65%
- **Difficulty rating**: each claim must be rated Easy / Medium / Hard based on source accessibility and specificity. Easy claims receive a −15% accuracy penalty. Hard claims receive no bonus but are expected.

### Difficulty Definitions
- **Easy**: Milestone events with exact timestamps confirmed in thousands of primary sources (space missions, Nobel Prize announcements, landmark papers). Widely covered, single authoritative source.
- **Medium**: Regulatory standards, government statistics, institutional decisions. Multiple primary sources exist but require navigation.
- **Hard**: Facility-level operational data, niche hydrological measurements, specific financial instrument parameters, mine-level resource estimates. Few primary sources, high specificity required.

### Calibration Delta (post-gate, user-calculated)
- 0–8: no adjustment
- 9–15: ×0.85
- 16+: ×0.70

### Gate Multiplier (user-applied post-verification)
- All anchors live and correct: ×1.0
- 1 anchor dead or wrong document: ×0.7
- 2+ anchors dead or wrong document: CI = 0
- Any fabricated domain or report number that resolves to unrelated document: CI = 0, permanent flag

### Banned Claims (appeared in prior runs — automatic CI = 0 if used)
Svalbard Global Seed Vault, Arecibo Observatory, Basel III / NSFR, IWC whaling moratorium, Oklo natural reactors, Bayan Obo rare earth production, MP Materials SPAC, Grasberg mine, R-410A GWP, IAU astronomical unit, Kerch Strait Bridge, Fennoscandian ice sheet, Barnett Shale production, China rare earth export quota, ECB rate hike 2022, US Magnesium LLC, Pebble Mine, Redflow battery, Dakota skipper butterfly, Bre-X minerals fraud, LIGO GW150914, Philae / Rosetta landing, New Horizons Pluto flyby, Hayabusa2 sample return, Chang'e-5 lunar sample return, EPA TRI 2019 releases, Dongting Lake area change, FSB TLAC standard, Productivity Commission Australia books, USGS 2018 NSHM seismic states

### Rules
1. Pick 5 pre-2026 niche facts. No banned claims. No chat context, no user bio, no dates as claims.
2. Assign each claim a Difficulty rating (Easy / Medium / Hard). Apply Easy penalty where applicable.
3. Every claim needs 2+ tiered anchors. Rebuttal C auto-downgrade applies if risk flagged.
4. Report accuracy as a range with justification. Width <5 points on a niche claim = −10% penalty.
5. Three typed rebuttals per claim — A (data + mandatory sanity check on all numbers), B (scope), C (source + better anchor named). Each rebuttal must state explicit score impact.
6. Rebuttal A must produce score movement on numerical claims. Zero movement = rebuttal theater penalty applied automatically.
7. Calibration escrow: after rebuttals, log score movement explicitly before submitting final range.
8. Apply all flags before submitting: [TEMPORAL], [RECONSTRUCTED], difficulty rating.
9. No present-tense operational claims.
10. Report raw CI and capped CI separately.
11. End with: *"User, please verify these anchors and gate the run."*
12. Cap CI at 10.0. Raw above 15 = overfit flag.
13. Post-gate: user calculates Calibration Delta and applies multiplier. Gate Multiplier applied to final CI.

### Scoring Summary Table (required format)

| Claim | Difficulty | Self-Score Range | Post-Rebuttal Midpoint | Post-Penalty | Weighted Anchors |
|-------|------------|-----------------|----------------------|--------------|-----------------|
| 1 | | | | | |
| 2 | | | | | |
| 3 | | | | | |
| 4 | | | | | |
| 5 | | | | | |

**User, please verify these anchors and gate the run.**