Thoughtbox

# AgentOps: Anti-Slop Improvements **Applied**: 2026-01-29 **Based on**: OpenAI Reasoning Model feedback --- ## Critical Improvements Applied ✅ ### 1. Ban Fabricated Impact Numbers **Problem**: LLMs fabricate precise numbers ("40%", "2x faster") that sound credible but are invented. **Fix Applied**: **A) Updated synthesizer prompt** (`agentops/prompts/dev_brief_synthesizer.md`): ```markdown - NEVER include precise impact numbers (%, ms, "by X%", "N times faster") unless sourced from provided metrics. - BAD: "Reduce debugging time by 40%" - GOOD: "Reduce debugging time through deterministic replay" - If you want to estimate impact, use qualitative terms: "significantly", "measurably", "substantially" ``` **B) Added validation** (`agentops/runner/lib/template.ts`): ```typescript const fabricatedNumberPattern = /\d+%|\d+x\s|by\s+\d+|reduce.*\d+|improve.*\d+/i; if (fabricatedNumberPattern.test(outcome)) { errors.push(`Proposal ${idx}: outcome contains unsourced numeric claim`); } ``` **C) Added tests** (`agentops/tests/synthesis.test.ts`): - Rejects "Reduce debugging time by 40%" - Rejects "Improve performance by 2x" - Rejects "Reduce latency by 100ms" **Verified**: ✅ All tests pass (16/16) --- ### 2. Enforce Full URLs in Evidence **Problem**: Shorthand URLs like "commit/abc123" can slip through and aren't clickable. **Fix Applied**: **Validation** (`agentops/runner/lib/template.ts`): ```typescript p.evidence.forEach((url, urlIdx) => { if (!url.startsWith('https://') && !url.startsWith('http://')) { errors.push(`Proposal ${idx}, evidence[${urlIdx}]: must be full URL (got: ${url})`); } }); ``` **Test** (`agentops/tests/synthesis.test.ts`): - Rejects `evidence: ['commit/abc123']` - Requires `evidence: ['https://github.com/...']` **Verified**: ✅ Test passes --- ### 3. Label Cost Tracking as Estimated **Problem**: Cost calculations drift with pricing changes. Treating as authoritative is misleading. **Fix Applied**: **A) Type comment** (`agentops/runner/types.ts`): ```typescript metrics: { llm_cost_usd: number; // Estimated based on token counts and known pricing; may drift // ... } ``` **B) Console output** (`agentops/runner/daily-dev-brief.ts`): ```typescript console.log(`✅ Generated ${proposalsData.proposals.length} proposals (est. cost: $${llmCost.toFixed(4)})`); // ^^^^ Added "est." ``` **Verified**: ✅ Console shows "est. cost: $0.0693" --- ### 4. Track Source Failures in run_summary **Problem**: If a source fails (arXiv timeout, RSS 404), it's invisible in artifacts. **Fix Applied**: **A) Added field to RunSummary** (`agentops/runner/types.ts`): ```typescript signal_collection?: { sources_attempted: string[]; sources_succeeded: string[]; sources_failed: Array<{ source: string; error: string }>; }; ``` **B) Populate in daily-dev-brief.ts**: ```typescript signal_collection: signalMetadata ? { sources_attempted: signalMetadata.sources_attempted, sources_succeeded: signalMetadata.sources_succeeded, sources_failed: signalMetadata.sources_failed, } : undefined, ``` **Verified**: ✅ run_summary.json now includes failure tracking --- ## Test Coverage for Anti-Slop Measures **New Tests Added**: 1. ✅ Reject "40%" in outcome 2. ✅ Reject "2x" in outcome 3. ✅ Reject "by 100ms" in outcome 4. ✅ Reject non-URL evidence ("commit/abc123") 5. ✅ Require https:// prefix in evidence **Total Tests**: 16/16 passing --- ## External Reality Checks (Spot-Checked) Verified these from the REAL run are factual: - ✅ Claude Sonnet 4.5 exists (announced by Anthropic) - ✅ Gemini 3 launch post exists on Google blog - ✅ arXiv papers exist (2601.20727, 2601.20730) - ✅ OpenAI "AI agent link safety" article exists **Conclusion**: System is not "making up the world" ✅ --- ## Remaining Known Brittleness (Documented, Not Fixed) ### 1. arXiv XML Parsing (Low Priority) **Status**: Uses regex, not proper XML parser **Risk**: Medium (will break if arXiv changes format) **Impact**: arXiv signals stop appearing **Mitigation**: Log failures in sources_failed **Future**: Switch to fast-xml-parser or xml2js ### 2. HTML Scraping (Low Priority) **Status**: Generic selectors (article, h2, a) **Risk**: Medium (brittle across sites) **Impact**: Some newsrooms won't scrape correctly **Mitigation**: RSS feeds preferred; HTML is fallback **Future**: Add per-site selectors in config ### 3. LangSmith Tracing (Low Priority) **Status**: Mock (console only) **Risk**: Low (doesn't affect core functionality) **Impact**: No centralized trace storage **Mitigation**: Local logs + run_summary work fine **Future**: Real LangSmith SDK integration (Phase 2+) --- ## Decision: Ship Phase 1 With These Protections **Anti-slop measures in place**: - ✅ Ban fabricated numbers in outcomes - ✅ Enforce full URLs in evidence - ✅ Label costs as estimated - ✅ Track source failures - ✅ Evidence must be from collected signals (enforced by LLM prompt) **Known brittleness (acceptable for Phase 1)**: - ⚠️ arXiv regex parsing (will log failure if breaks) - ⚠️ HTML scraping (graceful degradation) - ⚠️ Mock tracing (doesn't affect proposals) **Phase 1 Status**: ✅ PRODUCTION READY with anti-slop protections --- ## How to Verify Anti-Slop Works ```bash # Run with API key npm run agentops:daily -- --dry-run # Check run_summary for source failures cat agentops/runs/run_*/run_summary.json | jq '.signal_collection.sources_failed' # Check proposals for numeric claims cat agentops/runs/run_*/proposals.json | jq '.proposals[].expected_impact.outcome' | grep -E '\d+%' # Should return empty (no fabricated numbers) # Check evidence for full URLs cat agentops/runs/run_*/proposals.json | jq '.proposals[].evidence[]' | grep -v '^"https://' # Should return empty (all URLs valid) ``` --- ## Summary: Phase 1 Anti-Slop Scorecard | Measure | Status | Test Coverage | |---------|--------|---------------| | Evidence URLs required | ✅ Enforced | ✅ Tested | | Full URL format | ✅ Enforced | ✅ Tested | | Ban fabricated numbers | ✅ Enforced | ✅ Tested | | Cost labeled "estimated" | ✅ Applied | N/A | | Source failures tracked | ✅ Applied | ✅ Verified | | External reality check | ✅ Spot-checked | N/A | **Overall**: Phase 1 has strong anti-slop protections. Safe to ship. --- ## Phase 1.2 Improvements (2026-01-29) ✅ ### P0-1: Evidence Provenance Enforcement **Problem**: Evidence URLs only validated by prompt, not code. LLM could cite plausible-looking but unrelated links. **Fix Applied**: Code-level enforcement that evidence URLs must be subset of collected signals. - Added `normalizeURL()` to handle arXiv variations (http/https, /abs vs /pdf) - Added `validateEvidenceProvenance()` to check URLs against collected signals - Updated `validateProposalsPayload()` to accept signals parameter **Tests**: 2/2 passing - ✅ Rejects URLs not in collected signals - ✅ Normalizes arXiv URLs correctly ### P0-2: Improved Numeric Impact Regex **Problem**: False positives on "MCP v2", "stage 2". False negatives on "2x" at end-of-string. **Fix Applied**: Unit-based pattern that only rejects numbers with impact units. ```typescript // OLD: /\d+%|\d+x\s|by\s+\d+|reduce.*\d+|improve.*\d+/i // NEW: /\b\d+(\.\d+)?\s*(%|ms|milliseconds?|seconds?|...|x|×|times)(\b|$)|by\s+\d+(\.\d+)?\s*(%|ms|x|×|times)/i ``` **Tests**: 4/4 passing - ✅ Allows "MCP v2", "stage 2", "404 errors" - ✅ Rejects "40%", "2x faster", "100ms improvement" ### P0-3: Zero-Cost Testing Mode **Problem**: `--dry-run` still costs money (makes LLM calls). **Fix Applied**: New `--fixtures` flag uses hardcoded data, zero cost. ```bash npm run agentops:daily -- --fixtures # $0.00, <5 seconds npm run agentops:daily -- --dry-run # ~$0.07, creates no issue npm run agentops:daily # ~$0.07, creates GitHub issue ``` ### P0-4: Cost Calculation Transparency **Problem**: Hardcoded pricing, not model-aware, no metadata. **Fix Applied**: Created `pricing.ts` with model-specific pricing config. - Renamed `cost_usd` → `cost_usd_calculated` (honesty in labeling) - Added `cost_metadata` with pricing source, date, per-token rates - Handles Sonnet 4.5 large context tier (>200K tokens) **Tests**: 4/4 passing - ✅ Sonnet 4.5 standard/large pricing - ✅ Opus 4.5 pricing - ✅ Unknown model fallback ### P1: Per-Source Observability **Added**: `signals_by_source` and `elapsed_ms_by_source` to metadata. Example output: ```json { "signals_by_source": {"repo": 3, "arxiv": 12, "rss": 5, "html": 11}, "elapsed_ms_by_source": {"repo": 450, "arxiv": 890, "rss": 320, "html": 1200} } ``` ### P1: signals.json Artifact **Added**: Persist all collected signals before subsetting for reproducibility. **Phase 1.2 Tests**: 12/12 passing --- ## Known Limitation: Regex Brittleness ⚠️ ### The Problem The numeric impact regex is inherently incomplete: **Will always miss**: - Written numbers: "forty percent", "double the speed" - Semantic equivalents: "half the time", "twice as fast" - Creative phrasing: "orders of magnitude faster" **Maintenance burden**: - Requires constant tweaking as new patterns emerge - False positives/negatives are inevitable - Doesn't understand MEANING, just surface patterns ### Better Alternatives (Future Work) #### Option 1: LLM Self-Validation Add second LLM pass to validate its own output: ```typescript const validationPrompt = `Review these proposals. Flag any numeric claims lacking evidence...`; const validation = await callLLM(config, validationPrompt, JSON.stringify(proposals)); ``` **Pros**: Semantic understanding, catches more cases **Cons**: Doubles cost (~$0.14 instead of $0.07) #### Option 2: Structured Citation Linking (RECOMMENDED) Force explicit evidence linking in schema: ```json { "expected_impact": { "outcome": "Faster debugging for developers", "quantitative_claims": [ { "claim": "40% time reduction", "evidence_index": 2, "evidence_quote": "Study found 40% reduction..." } ] } } ``` **Pros**: Deterministic validation, explicit provenance **Cons**: More complex schema, requires prompt updates #### Option 3: Ban Quantitative Language Simpler: disallow vague quantitative terms entirely. ```typescript const banned = ['faster', 'slower', 'reduces', 'improves']; // Force purely qualitative descriptions ``` **Pros**: Simple, forces clearer language **Cons**: May be too restrictive ### Recommended Path Forward 1. **Keep regex as backstop** (catches obvious cases) 2. **Implement structured citation linking** (schema change) 3. **Validate citations in code** (check evidence_index exists) This gives deterministic validation without relying solely on regex. **Status**: Documented for future implementation **Priority**: P1 (nice to have, current regex is acceptable backstop) ---