proof-engine

Proof Engine

LLMs hallucinate facts and make reasoning errors. This skill overcomes both by offloading all verification to code and citations. Every fact is either computed by Python code anyone can re-run (Type A) or backed by a specific source, URL, and exact quote (Type B).

Produces three outputs: a re-runnable proof.py script, a reader-facing proof.md, and a proof_audit.md with full verification details.

Gotchas

These are the highest-value lessons from field testing. Read before writing any proof code.

• Don't inline verification logic: Import the bundled scripts. Rewriting normalize_text() inline risks garbling the HTML-stripping logic.
• Don't use int() truncation as a cross-check: int(days / 365.2425) == calendar_years is not independent — both are functions of the same input.
• Don't restate the proof as an adversarial check: "70 years after 1948 is 2018, and 2026 > 2018" catches nothing. Search for counter-evidence.
• Handle Unicode in citations: Real web pages use en-dashes, curly quotes, ring-above vs degree, non-breaking spaces. verify_citations.py handles this automatically.
• WebFetch/WebSearch return summaries, not verbatim text — Never use text from WebFetch/WebSearch directly as the quote field in empirical_facts. Use these tools to identify sources, then obtain verbatim quotes via Python requests.get(), browser-captured snapshot, or Wayback archive. See environment-and-sources.md for the full workflow. If a citation returns partial/not_found on a source you know contains the finding, suspect paraphrasing — obtain raw page text and update the quote.
• explain_calc() vs compute_*(): Use named functions (compute_percentage_change(), compute_age()) when they match your computation — they self-document. Use explain_calc() for ad-hoc expressions. Don't wrap a compute_*() call in explain_calc().
• Don't call verify_extraction() on data_values: It's circular. Instead, call verify_data_values(url, data_values, fact_id) to confirm each value string appears on the source page, then cross-check across sources (Rule 6).
• Never create pseudo-quote fields for table data: Don't store table cell values in fields like cpi_1913_quote: "9.883". If the source evidence is a table cell or numeric grid, store it under data_values and verify with verify_data_values(). The validator will reject pseudo-quote fields containing bare numeric or date literals that are parsed as evidence.
• Handle verify_data_values() failures: If verify_data_values() returns found: false for a source (common with JS-rendered pages), do not use that source's data_values as the primary computation input. Use a verified source as primary, and note the unverified source's data as corroborating only. If both sources fail verification, search for a third source with static HTML. A cross-check between two unverified data_values sources is circular — it compares your authored strings against each other.
• Index base mismatches: Economic data from different aggregators may use different base periods. If cross_check() flags a large disagreement, check whether sources use different scaling. Document the base period in source_name.
• Dynamic/JS-rendered sites: Many aggregators (officialdata.org, in2013dollars.com, inflationdata.com) render page chrome via JavaScript. Live fetch gets raw HTML — data tables may be static but page titles, headings, and navigation are often JS-rendered. Quote verification on page titles commonly fails even when data is correct. Use verify_data_values() as the primary verification for table data; treat quote verification as a bonus, not a requirement.
• cross_check() mode and tolerance: Use mode="absolute" for computed results that should match closely. Use mode="relative" for source-to-source comparisons. Tolerance heuristics for government statistics: expect 1-5% variation across aggregators due to rounding, month selection (annual avg vs December), and base-period differences. If sources disagree by more than 5%, investigate: find a third source, check if they use different base periods or date ranges, and document the discrepancy in adversarial checks. Don't silently ignore large disagreements — they may indicate one source is wrong.
• Quote selection for qualitative claims: Pick quotes that directly state the claim's core assertion, not tangential mentions. A source that says "the brain is remarkable" does not support "adult neurogenesis occurs." The quote must be specific enough that citation verification confirms the source actually addresses the claim.
• Academic HTML degrades citation matches: PMC and journal pages embed inline reference markers ([1], superscripts) that inject noise after HTML stripping. If a real verbatim quote gets partial status, check whether the source is academic HTML before suspecting the quote itself. Use snapshot to capture clean text if needed.
• Don't conflate source count with evidence strength: 5 news articles citing the same study count as 1 independent source, not 5. For qualitative consensus proofs, check whether sources trace to independent primary research. Document the independence rationale in the cross-checks section.
• Absence claims need search documentation: For "no evidence exists" claims, use the Absence-of-Evidence template. Document what was searched (databases, query terms, date ranges), not just what was found. The search_registry structure makes this machine-checkable.
• Don't weaken causal claims to prove them: If the claim says X "causes" Y, you cannot redefine it to X "is associated with" Y in operator_note and then PROVE the weaker version. Decompose into SC-association + SC-causation sub-claims using the compound claim template. If only observational evidence exists without causal inference methods (Bradford Hill, Mendelian randomization, natural experiments), the result is PARTIALLY VERIFIED (association confirmed, causation not established), not PROVED.
• Don't rank from point estimates when the source says they overlap: If Our World in Data says "nuclear: 0.07/TWh, solar: 0.05/TWh" but also says "the uncertainties mean these values are likely to overlap," you cannot conclude solar is safer than nuclear. Set uncertainty_override = True and return UNDETERMINED.
• Adversarial evidence is prose-only, not citation-verified: Sources in adversarial_checks are documented as prose in verification_performed — they are not machine-verified by verify_all_citations(). For contested qualifier proofs, this means the strongest counter-evidence (e.g., independent reviews rejecting a qualifier) is only as trustworthy as the proof author's characterization. Mitigate by: (1) quoting specific findings verbatim in verification_performed, (2) citing the source URL so reviewers can check, and (3) using multiple adversarial sources that independently reach the same conclusion.

Reference Files

Read these on demand, not all upfront.

File	Read when
hardening-rules.md	Step 3 — the 7 rules with bad/good examples
proof-templates.md	Step 3 — read this index to choose a template, then read the specific template file it directs you to
output-specs.md	Step 5 — proof.md and proof_audit.md structure
self-critique-checklist.md	Step 6 — before presenting results
advanced-patterns.md	When encountering complex quotes or table-sourced data
environment-and-sources.md	When facing fetch failures, paywalls, or .gov 403s

Bundled Scripts

Import these instead of re-implementing verification logic.

Script	Purpose	Key functions
scripts/extract_values.py	Parse values FROM quote strings (Rule 1)	parse_date_from_quote(), parse_number_from_quote(), parse_percentage_from_quote()
scripts/smart_extract.py	Unicode normalization + extraction utilities	normalize_unicode(), verify_extraction(), diagnose_mismatch()
scripts/verify_citations.py	Fetch URLs, verify quotes (Rule 2)	verify_citation(), verify_all_citations(), build_citation_detail(), verify_data_values()
scripts/computations.py	Verified constants, formulas, self-documenting output (Rule 7)	compute_age(), compare(), explain_calc(), cross_check(), compute_percentage_change()
scripts/source_credibility.py	Domain credibility from URL (offline). Called automatically by verify_all_citations().	assess_credibility(url)
scripts/validate_proof.py	Static analysis for rule compliance	ProofValidator(filepath).validate()

Key function signatures:

# computations.py
cross_check(value_a, value_b, tolerance=0.01, mode="absolute", label=None) -> bool
#   mode="absolute": |a - b| <= tolerance
#   mode="relative": |a - b| / max(|a|, |b|) <= tolerance
compute_percentage_change(old_value, new_value, label=None, mode="increase") -> float
#   mode="increase": (new - old) / old * 100
#   mode="decline": (1 - old / new) * 100
explain_calc(expr_str, scope, label=None) -> object
#   Prints symbolic -> substituted -> result. RETURNS the computed value.
compare(value, op_str, threshold, label=None) -> bool
#   Prints "{label}: {value} {op_str} {threshold} = {result}". Label defaults to "compare".

# verify_citations.py
build_citation_detail(fact_registry, citation_results, empirical_facts) -> dict
verify_data_values(url, data_values, fact_id, timeout=15, snapshot=None) -> dict
#   Fetches page and confirms each value string appears. Returns {key: {found, value, fetch_mode}}

Import pattern:

import sys
PROOF_ENGINE_ROOT = "${CLAUDE_SKILL_DIR}"  # replaced with actual path at proof-writing time
sys.path.insert(0, PROOF_ENGINE_ROOT)

Environment

Full Type B verification requires outbound HTTP from Python. Fallback chain: live fetch -> snapshot -> Wayback Machine (opt-in). Type A proofs run entirely offline.

For environment-specific details (Claude Code, ChatGPT, sandboxed), paywalled sources, and .gov workarounds, see environment-and-sources.md.

Core Concepts

Type A facts (Pure): Established entirely by code. The computation IS the verification.

Type B facts (Empirical): Established by citation. Each MUST have: source name, working URL, exact quote. Reputable sources only.

Type S facts (Search): For absence-of-evidence proofs. Each database search is documented with a clickable search_url. The tool confirms the URL is accessible but cannot verify the result count — that's author-reported and reproducible by a human reviewer. This weaker trust boundary is reflected in the SUPPORTED verdict (never PROVED).

Every proof has three parts: (1) Fact Registry — numbered facts tagged Type A, B, or S, (2) Proof Logic — a self-contained Python script, (3) Verdict — one of the levels below.

The 7 Hardening Rules

Rule	Closes failure mode	Enforced by
1. Never hand-type values	LLM misreads dates/numbers from quotes	scripts/extract_values.py
2. Verify citations by fetching	Fabricated quotes/URLs	scripts/verify_citations.py
3. Anchor to system time	LLM wrong about today's date	date.today()
4. Explicit claim interpretation	Silent ambiguity	CLAIM_FORMAL dict with operator_note
5. Independent adversarial check	Confirmation bias	Counter-evidence web searches
6. Independent cross-checks	Shared-variable bugs	Multiple sources parsed separately
7. Never hard-code constants/formulas	LLM misremembers values	scripts/computations.py

See hardening-rules.md for detailed examples of each.

Workflow

Step 1: Analyze the Claim

Classify: mathematical (Type A), empirical (Type B), or mixed. Identify ambiguous terms. Determine what constitutes proof AND disproof. For compound claims (X AND Y, X BECAUSE Y), decompose into sub-claims. Write a brief proof strategy and share with the user before proceeding.

If the claim is an opinion or has no verifiable answer, do NOT attempt a proof. Offer a related factual claim instead.

Guiding questions:

• Crisp true/false threshold? Extractable facts? Canonical sources? Clear disproof condition?
• If fewer than 3 are true, consider a simpler factual summary instead.
• For consensus claims, see template-qualitative.md.

Step 2: Gather Facts (Both Directions)

Use your environment's web search tool — do not rely on memory for source selection. LLM training data has a cutoff; sources recalled from memory may be outdated. Perform at least three searches:

1. Search for the claim itself — find authoritative sources that address it directly
2. Search for recent data — find the latest benchmarks, studies, or statistics on the topic
3. Search for counter-evidence — find sources that contradict, debunk, or criticize the claim. Search for "[claim] debunked", "[claim] wrong", or "[claim] criticism"

Recency check: If your best sources are older than 12 months, search specifically for newer data. Fast-moving fields (AI benchmarks, politics, economics, medicine) require sources from the current year when available. Prefer recent primary sources over older ones when they cover the same data.

If web search is unavailable in your environment, note this limitation in the proof audit under adversarial checks and flag that sources may not reflect the latest data.

Find at least two independent sources (Rule 6). For math claims, plan two independent computation approaches.

Adversarial work happens once, here. Use web search for counter-evidence — do not rely on memory. The adversarial_checks list in proof code records what you found — it's documentation of Step 2 research, not code that runs searches at proof execution time. Use past tense in verification_performed (e.g., "Searched for counter-evidence...") to make this clear.

Adversarial sources belong in adversarial_checks, not empirical_facts. Sources that argue against your proof's conclusion should be documented in the adversarial_checks list's verification_performed field. Only sources that support the proof's conclusion belong in empirical_facts. This prevents adversarial citation failures from degrading the verdict via any_unverified. For contested qualifier claims: sources that reject the qualifier (e.g., "claims not substantiated," "allegations not verified") are adversarial to SC2 — put them in adversarial_checks, not SC2's empirical_facts. It is normal and expected for SC2 to have zero empirical facts when no independent body has confirmed the qualifier.

Pre-fetch snapshots early, not late. Many news and advocacy sites now return 403 to automated fetches — not just .gov/.edu. During Step 2 research, pre-fetch the full page text for every source you plan to cite and include it as the snapshot field in empirical_facts. This avoids discovering fetch failures late during verify_all_citations(), which forces source substitution under time pressure. Note: WebFetch and verify_all_citations() use different HTTP clients — a WebFetch 403 does not mean the script will also get 403, and vice versa. If both fail, the snapshot is your only recourse. See environment-and-sources.md for details.

Step 3: Write the Proof Code

Read hardening-rules.md for the 7 rules. Then read proof-templates.md to identify which template matches your claim type. Then read the specific template file it directs you to (e.g., template-qualitative.md, template-compound.md). Do not skip the second read — the index contains only the decision table, not the template code. If the claim uses an epistemic qualifier ("verified," "confirmed," "proven," "established"), use the compound claim template (template-compound.md) with the contested qualifier pattern: SC1 (provenance — do the numbers come from a credible source?) + SC2 (epistemic warrant — has the qualifier been independently confirmed?). If the claim uses causal language ("causes," "leads to," "promotes," "damages," "prevents"), use the compound claim template (template-compound.md) with SC-association + SC-causation sub-claims — see "Causal vs. associational claims" in the Verdicts section. For claims about absence of evidence, use template-absence.md. The proof script must be self-contained: python proof.py produces the full output.

Required elements:

• CLAIM_FORMAL dict with operator_note (Rule 4)
• empirical_facts dict (empirical proofs) or pure-math template (math proofs)
• compare() for claim evaluation, explain_calc() for computation traces (Rule 7)
• Adversarial checks with verification_performed field (Rule 5)
• Cross-checks from independent sources/methods (Rule 6)
• FACT_REGISTRY mapping report IDs to proof-script keys
• JSON summary block in __main__ ending with === PROOF SUMMARY (JSON) ===

Step 4: Validate

Run python ${CLAUDE_SKILL_DIR}/scripts/validate_proof.py proof_file.py and fix issues.

Step 5: Execute and Report

Run the proof script. Write three files: proof.py, proof.md, proof_audit.md.

For detailed output specifications, see output-specs.md.

Step 6: Self-Critique

Before presenting results, run through the checklist in self-critique-checklist.md.

Verdicts

Verdict	Meaning
PROVED	All facts verified, logic valid, conclusion follows
PROVED (with unverified citations)	Logic valid but some citation URLs couldn't be fetched
SUPPORTED	Absence-of-evidence threshold met, no counter-evidence found
SUPPORTED (with unverified citations)	Absence threshold met but corroborating citations couldn't be fetched
DISPROVED	Verified counterexample or contradiction found
DISPROVED (with unverified citations)	Counterexample found but some citations couldn't be fetched
PARTIALLY VERIFIED	Some sub-claims met threshold, others did not — Conclusion states whether each failing SC lacked evidence or was contradicted
UNDETERMINED	Insufficient evidence either way

Threshold guidance for source-counting proofs: The default threshold: 3 means 3 independently verified sources must confirm the claim. Never set threshold: 1 — a single source is not consensus.

Reducing to threshold: 2 is permitted only when ALL of the following are met:

1. Domain scarcity: Fewer than 3 independent authoritative sources exist. Document the search that established this (databases queried, terms used, why results are insufficient).
2. Source quality adequate for domain: Each threshold source must meet the minimum quality standard for its domain:
   • Human/clinical studies: n >= 30 participants (or justify smaller n for rare conditions)
   • Physical/mathematical: peer-reviewed or established reference (textbook, standards body)
   • Economic/statistical: government or intergovernmental data source, or peer-reviewed analysis
   • Other domains: tier >= 3 credibility (no unclassified or flagged sources)
3. No majority COI: No more than half of the threshold sources may have a conflict of interest (per Rule 6's COI taxonomy: organizational, funding dependency, institutional co-benefit, competitive antagonism, revolving door, or advocacy/ideological) on the same side of the claim.
4. Documented in operator_note: State why 3 sources are unavailable, confirm source quality, and disclose any known COI.

If these conditions are not met, keep threshold: 3. If fewer than 3 qualifying sources exist and the quality gates are not met, the verdict should be UNDETERMINED (insufficient evidence), not PROVED at a lowered threshold.

Causal vs. associational claims: When the claim uses causal language — "causes," "leads to," "promotes," "triggers," "results in," "damages," "prevents" — the proof must decompose it into at least two sub-claims using the compound claim template:

• SC-association: "X is associated with Y" — satisfiable by observational cohorts, cross-sectional studies, epidemiological correlations.
• SC-causation: "The association is causal (not confounded)" — satisfiable by:
  • RCTs or controlled experiments (gold standard)
  • Established causal inference methods where RCTs are impractical: Bradford Hill criteria assessment, Mendelian randomization, natural experiments, dose-response with adequate confounding control
  • Converging mechanistic evidence (biological pathway) + longitudinal outcome data from independent studies

Verdict outcomes:

• SC-association holds + SC-causation holds -> PROVED
• SC-association holds + SC-causation fails -> PARTIALLY VERIFIED (association confirmed, causation not established)
• SC-association fails -> use proof_direction: "disprove" in CLAIM_FORMAL, which maps to DISPROVED when disproof sub-claims hold
• Insufficient evidence for SC-association -> UNDETERMINED

The operator_note must NOT redefine a causal claim as associational to avoid this decomposition. This rule does not apply to claims already phrased associationally ("is correlated with," "is associated with").

Limitations

Strong for crisp, auditable, bounded claims; weak for open-ended, normative, or predictive claims. The key limit is formalizable vs fuzzy — a claim works if it decomposes into extractable facts and a clear rule for proof/disproof.

Disproof is often easier (single counterexample suffices). The engine struggles with: deep original mathematics beyond sympy, broad causal inference, competing definitions, and large literature synthesis. Citation verification confirms quote presence, not semantic entailment — the adversarial check (Rule 5) mitigates this.

Edge Cases

Fictitious source attributions: If a claim attributes data to a specific source that doesn't contain that data (e.g., "according to the 1947 British census" when no such census exists), treat it as a compound claim: (SC1) the numeric value is correct, (SC2) the stated source contains it. Prove SC1 from the actual source, and note the attribution error in operator_note and adversarial checks. The verdict reflects both sub-claims.

Partial-period data: If a claim covers a time range but the best sources only cover part of it (e.g., claim says 1994-2023, sources cover 1994-2020), document the gap in operator_note. For cumulative nonnegative totals (e.g., total aid disbursed, cumulative emissions), if the partial-period sum already exceeds the claim's threshold, prove it with a logical extension: "If S₂₀ > T and the quantity is monotonically nondecreasing (cumulative total cannot shrink), then S₂₃ ≥ S₂₀ > T." State the monotonicity assumption explicitly using explain_calc(). This shortcut does NOT apply to averages, rates, percentages, or rolling metrics — for those, the missing-period values could decrease the aggregate, and you must either find full-period sources or return UNDETERMINED with an explanation of what data is missing.

Source doesn't contain claimed constant: If a claim says "per [Source]" but that source doesn't publish the specific constant (e.g., "CODATA values for solar mass" when CODATA doesn't list solar mass), document the substitution in operator_note: which source you actually used, why it's authoritative, and how it relates to the claimed source.

Comparative claims with source-acknowledged uncertainty: When a claim uses a superlative or comparative — "the safest," "the lowest," "the most," "better than" — and the source used to evaluate the comparison explicitly states that the compared values have overlapping uncertainty ranges, confidence intervals, or error bars: set uncertainty_override = True in the verdict section (date/age or numeric template). The verdict will be UNDETERMINED, not PROVED or DISPROVED. Point estimates alone cannot resolve a ranking when the source itself flags that the estimates are not statistically distinguishable. Document in operator_note: "Source [name] states [exact quote about uncertainty overlap]." This applies when the source providing the data also flags the uncertainty, or when the caveat comes from a source analyzing the same underlying dataset.