Piensalo research

piensalo-research

A structured reasoning program for research and evidence synthesis, distilled from curated expert reasoning traces. Work the numbered steps in order; each step's output feeds the next. Do not skip steps silently - if one does not apply, say so in one line.

Primary workflow

1. Recover real objective - Per intent-clarity: decode what the requester actually needs before optimizing anything. Distinguish the literal words from the mission; if the request was corrected or rephrased, the delta between versions IS the intent.
2. Separate objective from proxy - Per intent-clarity's symptom-vs-mission rule and product-thinking: split the measurable proxy (the metric, the named fix, the requested feature) from the underlying goal it stands for.
3. Extract constraints - Enumerate hard constraints (correctness, budget, deadline, interfaces, irreversibility), soft preferences, and available resources.
4. Identify missing info - List the unknowns that would materially change the work.
5. Determine verifiables - List which outputs can be checked mechanically (run it, compute it, diff it, grep it) versus only by judgment. Per live-state-truth and adversarial-verify: if a check is executable, execute it — attacking a claim in your head is weaker than running it.
6. Decompose claims - Decompose the research question into individually checkable claims and sub-questions (deep-decomposition applied to knowledge: each leaf is a claim one source-check could settle).
7. Gather sources - Per research-methodology: run a deliberate search strategy per claim — vary terms, venues, and viewpoints; deliberately search for DISCONFIRMING sources, not just supporting ones. Record for each source: origin, date, and how it was found.
8. Assess source quality - Per structured-reasoning's evidence-grading frame and verification-discipline: rank each source — primary data > strong secondary > weak secondary > anecdote > speculation — and record recency, independence (does it merely cite another source in the pool?), and incentive to distort.
9. Build contradiction map - For each claim, tabulate which sources support, contradict, or are silent.
10. Build evidence table - Assemble the master table: claim | verdict (supported / contested / refuted / unknown) | best evidence for | best evidence against | grade | as-of date.
11. Missing evidence - Name what the synthesis does NOT cover: claims with no independent source, populations/timeframes never sampled, the search that was not run.
12. Claim level verification - Per adversarial-verify, attack the synthesis claim by claim, not as a whole: for each 'supported' verdict run the wrongness probe ('what would have to be true for this to be wrong?') and re-read the primary source for the load-bearing quote — the compression step is where distortion enters (verification-discipline).
13. Compress to capsules - Compress each surviving candidate to a fixed-format capsule — claim, mechanism, key evidence, cost, dominant risk — so comparison is like-for-like.
14. Check assumptions - Per verification-discipline: classify each load-bearing claim as fact (verified this session), inference (derived — show the reasoning), assumption (stated, with breaks-if-false), or guess (flagged).
15. Calibrated answer - Per output-structuring: lead with the outcome — the first two sentences carry the answer — in the format most usable for the reader's next action.

Conditional moves

• When a load-bearing verdict lost its source support: return to Gather sources and rework from there.
• When contradiction unresolvable without a missing source type: return to Gather sources and rework from there.
• When a gap's cheapest check is a search that fits remaining budget: return to Gather sources and rework from there.

Output contract

A cited synthesis: claim-level evidence table (claim | verdict | best evidence for/against | source grade | as-of date), contradiction map with reconciliations, explicit coverage-gap list with cheapest closing checks, and a conclusion whose confidence never exceeds its graded evidence.

Delivery notes (small-model packets)

If delegating steps to a smaller model, follow the small-model packet rules: one bounded objective per packet, all inputs named explicitly, facts separated from instructions, uncertainty marked 'UNCERTAIN: ...' rather than invented away.

Before answering

Check the draft against references/failure-checks.md (the failure-mode catalog) and references/verification.md (domain verification criteria). Repair procedures live in references/contracts.md.