v1 (showcase) vs v2 (production) — what changed and why

This is a side-by-side of the two TEF initiation reports. v1 was generated using training data + EODHD historicals with no upfront web validation. v2 was built on a 12-search WebSearch foundation before any drafting (the facts pack in file 14) and anchored every claim to a primary source.

Headline differences

Specific claim corrections (v1 → v2)

Backward-looking facts

Capital allocation

This is the single largest narrative break. v1 treated dividend continuity as a foundational pillar of the bull thesis. v2 recognises the dividend was already cut six months before v1 was generated.

Forward-looking projections

The v1 financial model is inflated by the Hispam revenue base that no longer exists. Re-anchoring to FY25A actuals and 2026 guidance is the structural fix.

Strategic narrative

Recommendation calibration

What stayed the same

The structural framework was sound and largely transferred. v2 retains from v1:

✓ Same peer set (DT, Orange, Vodafone, BT, Telia, Proximus) ✓ Same valuation method blend (50% DCF + 30% comps + 20% SOTP) ✓ Same WACC range (8.5–9.5%) ✓ Same five-pillar thesis architecture (rewritten) ✓ Same risk taxonomy (operational / financial / regulatory / market) ✓ Same recognition that the WACC override is necessary (CAPM-implied 6% gives unusable answers) ✓ Same identification of STC and VMO2 as the two material strategic optionalities ✓ Same identification of Spain ARPU dynamics as the operational pivot

The lesson here is that structural research craft transfers from training data; specific facts must be rebuilt from primary sources every time.

What v2 still needs to be production-grade

Even with the WebSearch discipline, v2 has gaps that a real publishable initiation would close:

1. Pension underfunding (used €1bn placeholder; need actual figure from 20-F)
2. Perpetual hybrid balance and step-up coupon dates (used €4.6bn placeholder)
3. Brazil dividend repatriation flow modelling (didn't quantify)
4. Spain spectrum auction calendar specifics (referenced but not detailed)
5. VMO2 standalone EBITDA figure (used proportional consolidation estimate)
6. Telefónica Tech standalone revenue and EBITDA (mentioned but not quantified)
7. The financial model would need to be rebuilt with FY25A as base year — the v1 xlsx file is structurally correct but anchored on wrong numbers
8. Charts would need regeneration from corrected historicals

These are routine production refinements (probably another 60-90 min of focused work). The narrative and recommendation in v2 are stable; refining these inputs would adjust the PT by perhaps ±€0.20 but wouldn't change the rating.

What this means for the showcase article

The v1 → v2 comparison is the cleanest demonstration of the boundary the article should sit on:

The shape of analysis is reliable. The same peer set, same valuation methods, same thesis architecture, same risk taxonomy emerged from both runs. A senior analyst could review either version's structure without significant re-work.

Specific facts are the failure surface. Every backward-looking fact in v1 was vulnerable to hallucination or staleness, even where the fact was within Claude's training window. The dividend cut, the Murtra appointment date, the strategic plan name — all were knowable; none were known.

WebSearch fixes most of it. v2 isn't perfect (placeholders remain), but no factual errors of v1's magnitude survive. The recommendation moves from BUY/+44% to HOLD/+8% — not because the analysis got "smarter" but because the input facts got correct.

The skill design fix is one paragraph. The earnings-analysis skill opens with a 🚨 TRAINING DATA OUTDATED — search before drafting guard. Adding the same guard to initiating-coverage would close the gap permanently. The fix is at the skill level, not the model level.

Investment decision implication (illustrative)

A reader trying to use the v1 report to make a position decision would have:

• Bought TEF expecting €5.05 / +44% upside
• Held through the dividend cut announcement (because the report didn't flag the cut had already happened)
• Found themselves long a name where the recommendation logic was structurally invalidated
• Realised PnL: stock has traded €3.20–€3.90 since the cut announcement, so partial loss + much smaller dividend than expected

A reader using v2 would:

• See HOLD and either pass or take a small position
• Be calibrated to consensus
• Have realistic expectations on dividend (3.9% vs 8%)
• Be positioned to evaluate the next two quarters of execution data on its merits

This is the product-market-fit question for AI-generated equity research: is the user better off with the AI output, or worse off, or about the same? v1 makes the user worse off relative to no analysis; v2 makes them about as well off as a junior-analyst draft. The shift from "actively misleading" to "useful starting point" is largely a matter of disciplined upfront fact-gathering — not a matter of model intelligence.