DESIGN_STRATEGIST.md §2. Four read-only view tools the strategist uses
to ground its decision each round.
graph_overview() — hypotheses table (log_odds, conf, edges_in,
distinct_sources, recent_flip), sources table,
pending leads. distinct_sources is the
critical signal: a hypothesis with 23 edges
but only 1 distinct_source has fragile cross-
source independence and is a candidate for
a corroboration-seeking lead.
source_coverage(src) — per-source ✓/✗ against an expected-artefact
catalogue. Catalogue is heuristic hints,
NOT a forced checklist. Footer reminds the
strategist to investigate ✗ items only when
an active hypothesis depends on them — this
is the "应试能力存在但不被绑死" guardrail.
marginal_yield(N) — new phenomena / edges / status flips per
recent round. Two consecutive zero-yield
rounds = strong signal to declare complete.
budget_status() — usage vs caps (tool_calls, rounds, wall
clock). Pacing warnings at 70% / 90%.
tools/strategy.py also exports EXPECTED_ARTEFACTS, a per-source-type
table of (name, detector, value_for) entries. Detectors are
substring patterns on tool name + args; the matcher resolves at
call time against graph.tool_invocations. Catalogue covers iOS /
Android / Windows disk / media-collection / archive source types.
All four tools registered in tool_registry, listed as read-only in
llm_client.READ_ONLY_TOOLS for parallel execution. They go through
the invocation-logging wrapper so the strategist's reads are
themselves auditable (the wrapper does NOT cache them — graph
state changes between calls).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
DESIGN_STRATEGIST.md §1. Foundation for the Phase 3 strategist loop.
Lead now carries four annotations that let the orchestrator measure
marginal yield per lead and dedupe strategist proposals:
- proposed_by (agent that proposed it: "strategist", "filesystem", …)
- motivating_hypothesis (hyp-id the lead is meant to corroborate/refute)
- expected_evidence_type (edge type the lead's worker should produce)
- round_number (0 = Phase 1 lead, ≥1 = strategist-proposed)
add_lead idempotently dedupes strategist proposals on
(motivating_hypothesis, expected_evidence_type, target_agent, source_id)
to prevent the "strategist loops on the same lead" failure mode.
New InvestigationRound dataclass records per-round provenance: before/
after hypothesis status snapshots, phenomena + edge count deltas, and
the strategist's decision_rationale. ``new_phenomena_count``,
``new_edges_count``, ``status_flips`` are derived properties that the
marginal_yield tool will use.
start_investigation_round / complete_investigation_round /
get_investigation_round / latest_round / leads_from_round complete the
lifecycle. complete is idempotent on already-closed rounds.
Lead.from_dict is forward-compat for state files written before this
commit. InvestigationRound persists as a top-level list in
graph_state.json (auto-save + load_state both wired).
EvidenceGraph also gains graph.budgets and graph.run_start_monotonic
fields that the budget_status view (S2) will read; orchestrator
populates them in S5.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
First full-case run (runs/2026-05-20T20-15-04/) produced hypotheses
with log_odds +31 (8 direct_evidence + 15 supports). That's the
naive-Bayes independence assumption breaking down: 15 different
phenomena all "supporting" the same hypothesis from one source are
not 15 independent pieces of evidence, they're highly correlated.
DESIGN.md §4.5 last bullet flagged this as a "未实施旋钮" — this
commit implements it.
Rule: the k-th edge of a given (hyp_id, edge_type) contributes
log_lr_base / k instead of log_lr_base. Cumulative is harmonic
sum H_N, bounded by ~ ln N. Single-edge hypotheses unaffected
(k=1 → /1 → no change). Replaying the 2026-05-20 graph's 108
edges under the new rule pulls the top hypothesis from +31.0 →
+8.75; the smallest active hypothesis from +4.0 → +2.08.
Also adds rank + log_lr_base to confidence_log entries so the
math is auditable from the persisted graph.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Timeline agent on the 2026-05-20 full run produced 0 phenomena: initial
round hit max_iterations=60 cap before recording, forced retry then hit
max_iterations=10 cap because every grounding-rejected call burns one
iteration in the new gateway. Two changes restore depth without re-
introducing the original "agent wanders off and never records" failure:
1. Raise retry cap 10 → 30. With grounding auto-rescue (prev commit)
most rejections heal on the first retry, but some still need 2-3
turns; 10 is empirically too tight, 30 leaves headroom.
2. Narrow the retry tool surface to RECORD + graph-write +
read-only-graph-query tools. Investigation tools (list_directory,
sqlite_query, parse_registry_key) are dropped on retry so the agent
can't restart its search loop — the retry is explicitly "record
what you already found, then stop".
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
First full-case run (runs/2026-05-20T20-15-04/) produced 83 GroundingError
rejections, almost all from a single failure mode: LLM cites a plausible-
looking inv-XXXXXXXX that doesn't exist, while the fact's value is in fact
present verbatim in one of its real tool outputs. The agent knew which
tool it read from, it just mis-typed the citation id.
Two-layer fix in evidence_graph.validate_fact_grounding:
Layer A (silent heal): when the cited inv-id misses, search the same
agent / task's invocations for one whose output contains the value
(strict or normalised substring). If exactly one matches, rewrite
fact.invocation_id in place and accept. Multi-match is NOT auto-
rescued — the candidate ids go back to the LLM so it picks deliberately.
Layer B (informative retry): GroundingError now appends the agent's
recent invocation ids and a brief tool-call summary, so the LLM has
the real ids in front of it for the next attempt rather than
fabricating again from memory.
Both layers preserve the design invariant: the fact's value must still
be present in a real tool output — nothing new can land grounded that
wasn't already verifiable. Cross-agent / cross-task isolation is also
preserved (rescue candidates filtered on agent + task_id).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
