napkin-math: Phase 2 extract — source-arithmetic preservation, threshold-pairing parity, source_text truncation

[neoneye]

Scope

Phase 2 extract-prompt rules. Three commits to the two extract skills (extract-parameters-from-digest and extract-parameters-from-full), plus a same-session regression check via the napkin_math validate_parameters.py against six baseline plan probes.

The PR is narrow on purpose. It does NOT include:

• The source-preservation audit (proposal Multiple api keys cleanup5 #141; design is on main, implementation is a separate follow-up PR).
• Compress-LLM run-to-run variance handling (also a separate follow-up).
• Prompt-hygiene cleanup of the older domain-specific example (european_prepper_active_buyers) in the extract-from-digest prompt (scoped follow-up).
• Any claim that v51 outputs are globally better than v49 outputs — see "Verification posture" below.

What changes

extract-parameters-from-digest/system-prompt.txt and extract-parameters-from-full/system-prompt.txt

All edits applied symmetrically to both files.

1. Source-arithmetic preservation rule (new section, slotted between "Combined viability gate preservation" and "Critical-output completeness rule"). Three named patterns under one principle:
   • Aggregate sum (R1.1) — aggregate = sum_of_components ONLY when the source states or clearly implies the total is computed from the named constituents. Independent caps / ceilings / committed budgets / funding envelopes stay as primitive thresholds in key_values and are tested via the threshold-pairing rule's spend-vs-cap margin.
   • Burn rate × duration (R2.3) — total = rate * duration when the source supplies the rate AND the duration.
   • Explicit decomposition block (R2.4) — preserve every named operand and the operation between them.
2. Threshold-pairing rule backfilled into extract-parameters-from-full. PR napkin-math: prompt cleanup for compress + extract #737 only added it to extract-parameters-from-digest; this closes the parity gap.
3. Tightened aggregate-sum wording to explicitly exclude budget caps / ceilings / envelopes from the sum-identity pattern. Reconciled the "discipline shared" paragraph with the cap-pressure paragraph (placement decision between recommended_first_calculations and derived_questions is not a relaxation of the preservation rule).
4. Source_text 20-word cap reinforced with truncation discipline. Existing rule said the cap exists; new rule says how to truncate when a source if/then conditional runs long: drop the consequence clause, keep subject + threshold, end with ellipsis if mid-sentence, count words before emitting.

All four edits are corpus-agnostic. No plan names, no literals from any baseline plan, no expected output ids, no domain-specific jargon.

Verification posture

Compress + extract were re-run on six baseline plans (paperclip, datacenter, crate_recovery, yellowstone, mars_gtld, euro_adoption). v51 outputs were produced fresh via the skill under the updated prompts. All six v51/parameters.json files validate against validate_parameters.py with valid=True, errors=0, warnings=0.

This is a regression check, not an improvement claim.

• Same-LLM same-session regeneration cannot prove behavioural improvement on rules the same LLM authored and is applying. The same model that wrote the rules is also obeying them.
• Honest behavioural verification of these rules requires a different LLM or fresh-context same-model extraction against the same digests. That work is a separate follow-up.
• The regression check proves only that the rules do not break valid extraction.

Known unresolved issues, NOT addressed in this PR

These are explicitly out of scope. Each gets its own follow-up PR.

• Paperclip OPC UA / p99 latency tripwires absent from v50/v51 (present in v49). Compress-stage variance, not an extract-prompt problem. Addressed by the compress retry/merge or lower-temperature follow-up.
• v49 absences across the other probes (e.g. yellowstone public_compliance_*, mars registration_volume_buffer_fraction, crate q_logistics_budget_margin) remain unclassified tradeoffs. Without proposal 141's dropped_signals schema and audit script, I cannot mechanically distinguish "rename" / "structural restructure" / "acceptable drop under cap pressure" / "silent regression". Implementing Multiple api keys cleanup5 #141 is the next-most-load-bearing follow-up.
• Behavioural validation on a different LLM has not been done.
• european_prepper_active_buyers still appears in the extract-from-digest prompt as a formula example. Prompt-hygiene follow-up.

Documentation

experiments/napkin_math/docs/20260520_plan.md updated to reflect the narrow scope (commit f9d90ebb):

• "Landed on main" subsection lists PR napkin-math: prompt cleanup for compress + extract #737 (merged) and PR docs(napkin-math): proposal 141 — source-preservation audit #739 (proposal 141 design merged, code not).
• "Open for merge" subsection lists this PR's three substantive commits.
• "Verification posture" subsection states the regression-check-not-improvement-claim explicitly.
• "Known limitations" subsection names the four unresolved issues above by name.
• "Next likely move" subsection names four separate follow-up PRs in preferred order.

Commit chain

• 4cda70ba — Source-arithmetic preservation rule + threshold-pairing parity backfill
• 19f927b7 — Tightened aggregate-sum scope; reconciled discipline-shared paragraph with cap-pressure
• 8f94c8cd — 20-word source_text truncation discipline
• f9d90ebb — Plan status section updated for PR napkin-math: Phase 2 extract — source-arithmetic preservation, threshold-pairing parity, source_text truncation #740 narrow scope

Test plan

• CI green (lint / tests / typecheck)
• validate_parameters.py passes against all six regenerated v51 outputs (already confirmed: 6/6 clean)
• Hand-spot all four edits read corpus-agnostically (no plan-specific literals introduced)
• PR description framing matches the merge-readiness criteria — narrow scope, regression check not improvement claim, unresolved issues explicitly deferred