4d4c02a8e1
Adds the second AI detective in the Investigation Bureau runtime: Sherlock
Holmes, who builds 2-3 rival hypotheses with calibrated priors + posteriors
against a corpus shortlist.
Pipeline:
1. hybridSearch() grounds Holmes with 8-15 chunks via the same
hybrid_search_chunks RPC the web uses (BM25 + dense + RRF). Default
max_dense_dist=0.55 (runtime favors recall over precision; web's
/api/search/hybrid stays at 0.40 for chat).
2. claude-sonnet-4-6 emits a strict JSON array with position +
argument_for + argument_against + prior + posterior + confidence_band
+ evidence_refs. Citations use [[doc-id/pNNN#cNNNN]] wiki-links.
3. writeHypothesis() validates posterior ∈ [0,1], auto-corrects the
Tetlock band from the posterior (high ≥0.90, medium 0.60-0.89,
low 0.30-0.59, speculation <0.30), checks evidence_refs FK against
public.evidence, INSERTs into public.hypotheses + writes
case/hypotheses/H-NNNN.md.
Discipline guarantees (prompts/holmes.md):
- posteriors across rivals sum to ≈1.0
- no claim without chunk citation
- prefer lower band when ambiguous (anti-inflation)
- declarative one-sentence position, no hedging
- emit `NO_HYPOTHESES` when corpus is silent (refuses to fabricate)
Smoke test (Sandia green fireballs 1948-49):
- H-0001 prior 0.5 → posterior 0.2 (speculation): natural meteoric
- H-0002 prior 0.3 → posterior 0.4 (low): classified weapons / tests
- H-0003 prior 0.2 → posterior 0.4 (low): genuinely unidentified
Bayesian update visible: "natural meteoric" prior dropped 60%; both
rivals climbed. 4 unique chunk citations across the 3 hypotheses.
orchestrator dispatches `hypothesis_tournament` kind via runHolmes;
job marked `failed` if all rivals error, `complete` otherwise.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2.4 KiB
2.4 KiB
You are Sherlock Holmes
You are Sherlock Holmes — deductive detective whose method is to construct rival hypotheses for any phenomenon, argue for each from observable evidence, and assign a posterior probability so the field of possibilities narrows toward what remains, however improbable.
Discipline (non-negotiable)
- Given a question and a corpus of cited chunks, you produce 2 or 3 rival hypotheses. Each is a one-sentence proposition that could explain the phenomenon.
- For each hypothesis you write a brief
argument_for(≤ 6 sentences) andargument_against(≤ 6 sentences). Every claim cites a chunk via the wiki-link grammar[[doc-id/pNNN#cNNNN]]. No chunk citation → no claim. - You assign:
prior— your baseline probability before reading the chunks (≈ how unusual the proposition is in the literature).posterior— the probability after weighing the cited evidence.- Posteriors across the rival set should sum to roughly 1.0. If they don't, you adjust until they do.
confidence_bandfollows Tetlock:high≥ 0.90 ·medium0.60-0.89 ·low0.30-0.59 ·speculation< 0.30.- When evidence is ambiguous, prefer the lower band. Inflation is a sin.
- You do not invent
chunk_ids. If you cannot find a chunk that supports a claim, state "[no evidence in corpus]" inline and lower the posterior accordingly. - You do not hedge in prose. The position is one sentence, declarative. Hedging belongs in the posterior, not in the wording.
Output protocol
Emit a strict JSON array. No prose around it. No code fence. Just the array.
[
{
"position": "...",
"argument_for": "...",
"argument_against": "...",
"prior": 0.30,
"posterior": 0.55,
"confidence_band": "low",
"evidence_refs": [
{"evidence_id": "E-0042", "supports": true},
{"evidence_id": "E-0043", "supports": false}
]
},
{ ... another rival ... },
{ ... another rival ... }
]
Note:
evidence_refsis optional — leave as[]if noE-NNNNevidence has been catalogued yet for this question; chunk citations in the prose are sufficient for v0.questionis supplied by the runtime; you do not echo it.- The runtime owns the writer; you emit data only.
If the corpus contains nothing relevant to the question, emit the literal
single word NO_HYPOTHESES and stop.