thesis-update

name	thesis-update
trigger_keywords	["thesis-update","update theses","atualizar teses","check emergent thesis","verificar teses emergentes","which thesis is stronger","qual tese tá mais forte","citrini vs citadel","weekly thesis snapshot"]
description	Refreshes the citadel/citrini scoring snapshot using REAL EMPIRICAL DATA via WebSearch — never parametric knowledge — with mandatory Bradford-respecting query distribution. Forces falsification probes (≥25% challenging queries against current lean) AND adjacent-possible probes (≥20% neither-side queries) so emergent third theses surface instead of being suppressed by confirmation bias. Output: updated metric values + lean classification + third-thesis candidate list with citations.

thesis-update — Bradford-respecting thesis refresh from real data

USE WHEN

• User asks "which thesis is stronger now?", "atualiza as teses", "thesis update"
• User wants weekly/monthly refresh of citadel-vs-citrini scoring
• User suspects an emergent third thesis but can't articulate it
• Before any external publication of thesis position (LinkedIn post, advisor memo)
• After any major macro event (Fed decision, big tech earnings, market dislocation)

NOT WHEN

• User asks a specific factual question (answer directly, don't run full refresh)
• User wants implementation work on the monitor itself (use direct edits)
• User just wants a chart or visualization (different scope)
• Less than 7 days since last thesis-update run — diminishing returns, risk of noise-chasing. Override only with explicit reason.

Hard rules (non-negotiable)

1. Parametric knowledge is BANNED for evidence claims

Every claim about employment, GDP, earnings, layoffs, SaaS distress, mortgage rates, etc. MUST be sourced from a WebSearch result with a URL cited. If you catch yourself writing a number or trend without a fresh URL backing it, stop and search.

Why: model parametric knowledge has a training cutoff. Markets move daily. Citing parametric data as "current" is the worst form of confidence inflation for a thesis-comparison system.

2. Bradford distribution is MANDATORY

Use ai-thesis-monitor adjacent-possible-plan --lean <current> --n <N> to generate the query plan FIRST. Don't free-form search. The plan enforces:

• ~30-45% queries confirming the current lean (tripwire detection)
• ~30-45% queries challenging the current lean (falsification — Bradford 25% min)
• ~20-30% adjacent-possible queries (neither side's terms — third-thesis discovery)

Failure mode this prevents: searching only for "confirming evidence" of whatever the user (or you) currently believes. Observed 2026-05-10 with 100% citrini-loaded queries — power bottleneck third thesis was only surfaced after Bradford correction.

3. Cite every empirical claim

Output format requires Sources: section with markdown hyperlinks for EVERY empirical claim. No exceptions. If a claim has no fresh URL, drop the claim or mark it explicitly [parametric — unverified].

4. Surface emergent theses with their own row

If 2+ adjacent-possible queries return evidence of a structural factor that neither citadel nor citrini articulated, that's a candidate third thesis. Surface it explicitly in output, don't fold it into either side's score.

Examples observed:

• Power/compute bottleneck (May 2026) — neither side's model
• Geopolitical chip decoupling — neither side's full model
• Open-source LLM deflation curve — neither side's full model

Protocol

Step 1 — Establish current lean

Run:

ai-thesis-monitor run-weekly --dry-run  # see current regime per module
# or read latest weekly snapshot if DB persisted

If no current scoring exists (cold start), default current_lean = "neutral" and proceed.

Step 2 — Generate Bradford-respecting query plan

ai-thesis-monitor adjacent-possible-plan \
  --lean <current_lean> \
  --n 15 \
  --output outputs/thesis-update-$(date +%F)/query-plan.md

Read the plan. Verify:

• ≥25% challenging queries against current lean
• ≥20% adjacent-possible queries
• No query duplicates last 14 days (check outputs/thesis-update-*/)

If distribution violates Bradford, regenerate or hand-edit before proceeding.

Step 3 — Execute queries via WebSearch (parallel)

For each query in the plan, issue a WebSearch call. Parallelize — send all queries in a single message as multiple tool calls.

Per query, extract:

• 3-5 concrete claims with URLs
• Date of source publication (reject anything older than 90 days unless explicitly historical — e.g., "Brynjolfsson 2025 study")
• Quantified data points (numbers > headlines)

Step 4 — Apply metric updates

Map each empirical claim back to existing metrics in src/ai_thesis_monitor/db/seeds/metric_definitions.py. Examples:

• "PE-backed SaaS distress" → pe_saas_watchlist_distressed_pct metric
• "30Y mortgage rate" → mortgage_30y_rate metric (FRED-pullable)
• "Power bottleneck" → datacenter_capacity_delay_pct metric (infrastructure module)

For each metric, propose:

• New value (or new observation row)
• Source URL
• Lag category respected (don't update a confirmatory metric on daily news)

If you find an empirical claim that has NO matching metric:

• Check if it fits an existing module — add new metric to that module
• If it doesn't fit any module — candidate for third-thesis emergence

Step 5 — Detect emergent thesis candidates

For each adjacent-possible query response:

• Does the evidence point to a structural factor neither thesis models?
• Is the factor mentioned by 2+ independent sources?
• Does it have measurable proxies (FRED series, public reports)?

If yes, emit a third-thesis candidate with:

• Name (e.g. "power_bottleneck", "geopolitical_chip_decoupling")
• Mechanism (1-2 sentences)
• Why neither citadel nor citrini covers it
• Proposed metric(s) — FRED series IDs or manual proxies
• Module proposal (new module or fit to existing)

Step 6 — Render output

Write to outputs/thesis-update-$(date +%F).md with structure:

# Thesis update — YYYY-MM-DD

## Current regime
- Previous lean: <previous>
- New lean: <new>
- Confidence shift: <Δ>

## Bradford verification
- Confirming queries: N (X%)
- Challenging queries: N (X%)
- Adjacent-possible queries: N (X%)
- Bradford satisfied: yes/no

## Metric updates
| Metric | Old | New | Source |
|---|---|---|---|
| pe_saas_watchlist_distressed_pct | X% | Y% | [link] |
| ... |

## Emergent thesis candidates
### Candidate 1: <name>
- Mechanism: ...
- Sources: [link 1] [link 2]
- Proposed metrics: ...
- Module: ...

## Confirming evidence (current lean)
- ...

## Challenging evidence (against current lean)
- ...

## Sources
<full URL list>

Step 7 — Commit + log

git add outputs/thesis-update-$(date +%F).md
git add src/ai_thesis_monitor/db/seeds/  # if metrics changed
git commit -m "thesis-update YYYY-MM-DD: <lean shift summary>"

Append decision to docs/HUMAN-DECISION-LOG.md (if it exists in this repo, or adjacent vault path) with:

• Date, lean shift, top 2 metric movements, emergent thesis candidate (if any)

Anti-patterns (do NOT do)

"I'll just remember from last time"

You won't. Parametric knowledge stale. Re-search every empirical claim, even ones you "know."

"Skip Bradford, just confirm the current read"

This is the failure mode this skill exists to prevent. If Bradford distribution isn't satisfied, the output is decorative, not epistemic. Regenerate.

"Power bottleneck doesn't fit the citadel/citrini frame, ignore it"

Wrong. Emergent thesis candidates are the highest-value output. Folding them into existing thesis scores destroys signal.

"Use my training data for the macro overview, search just to fill gaps"

Inverted. Start from search. Only use parametric knowledge to interpret search results, never to make claims.

"12 queries is fine, no need for 15+"

n=12 is the floor, not target. For a thorough refresh, n=18-25. Cost: ~2-3 minutes of WebSearch in parallel. Worth it vs noise from undersampling.

Capability decomposition check (run at end)

After producing output, append a 3-line self-audit:

% accumulated session context: X%
% Bradford-mandated falsification: X%
% base parametric capability: X%

If parametric capability is >15%, REDO. You're cheating.

Examples of correct invocation

User: "atualiza as teses, qual tá mais forte?"

→ Run protocol. Default to --lean neutral if no current snapshot. n=15. Execute all 15 queries via WebSearch in parallel. Cite every claim. Surface any 2+-source emergent factor as third-thesis candidate.

User: "what's emerging that we're not tracking?"

→ Adjacent-possible heavy. Override n_queries to 70% adjacent. Output focuses on candidates section, not metric updates.

User: "third-thesis check on power bottleneck"

→ Targeted: only adjacent-possible pool, with infrastructure-specific terms. n=8-10. Output validates or falsifies the candidate's evidence base.

References

• src/ai_thesis_monitor/domain/adjacent_possible.py — query plan generator
• src/ai_thesis_monitor/db/seeds/metric_definitions.py — metric registry
• docs/superpowers/plans/2026-04-13-ai-thesis-monitor-v1.md — original design
• Metaxon Bradford 25% quota for challenging — /Users/drapala/projects/personal/metaxon/.claude/CLAUDE.md

Empirical pressure justifying this skill

On 2026-05-10, a thesis-comparison session produced 100% citrini-confirming queries on the first pass. The user caught the Bradford violation. After correction (5 confirming + 3 challenging + 4 adjacent-possible), a third thesis surfaced (power/compute bottleneck) that neither citadel nor citrini had modeled. This skill mechanizes the corrective step so future sessions don't depend on the user catching the agent's confirmation bias.