Baseline
10 / 11 MEETS
The problem · F1
The tie already exists. One line throws it away.
When two categories share an identical feature profile, BayesianReasoner computes exactly what you'd hope — a 50/50 posterior. But Cataloger._normalize_one keeps only each sub-system's single top answer, so the tie is gone before the cataloger ever votes. The system built to prefer "reporting competing candidates over picking one arbitrarily" cannot represent the most basic competition of all.
What Bayesian computes
A genuine tie. Both live in the distribution.
What the cataloger keeps
cheetah discarded. Result: 1 candidate, share 1.00, ambiguity 0.00.
The fix · distributed mass voting
Carry the distribution into the vote — no sub-system changes.
The distributions already exist in the sub-system outputs; the cataloger just stops discarding them. Each sub-system casts votes across its pruned candidate set, weighted by its own distribution. Two genuinely-close categories then accumulate comparable weight, and the existing ambiguity = runner_share / top_share formula finally sees real data.
Worked example
leopard vs cheetah, end to end
| Stage | leopard | cheetah | Note |
|---|---|---|---|
| Bayesian posterior | 0.50 | 0.50 | identical profiles |
| Pattern similarity | 0.90 | 0.90 | both kept (≥ 0.5 × top) |
| Distributed tally | 1.40 | 1.40 | summed weights |
| Composite share | 0.50 | 0.50 | — |
| ambiguity = runner ÷ top | 0.50 / 0.50 = 1.00 | ≥ 0.70 → flag | |
| result | ambiguous: True · 2 candidates | surfaced on the return path | |
A clear winner behaves as before: its runner-up prunes away (below 0.5 × top), leaving one candidate and ambiguous: False. The change only fires on a genuine near-tie.
Surfacing it · F3
One truth, carried on every return path
Fixing the cataloger is invisible unless the neuron surfaces it. Before this fix, ambiguous was set only in the uncertain-path helper, so a confident near-tie returned at step 6 with no signal at all. Now ambiguous is a single composite-level field, carried out of every exit.
| Return path | ambiguous | candidates |
|---|---|---|
| Fast-path cache (step 1) | False | — (a confident cache hit is unambiguous) |
| Confident (step 6) | composite["ambiguous"] | composite candidates |
| Routing delegation (step 7) | external["ambiguous"] | propagated from the expert |
| Honest uncertainty (step 8) | composite["ambiguous"] | composite candidates |
Named constants
Three knobs, each under test
The thing this project is rightly wary of — unmotivated magic numbers. Each lives in one place as a named parameter, tested on both sides of its boundary.
candidate_keep_ratio
0.5
A candidate must be ≥ half the top's weight (per sub-system) to count. A clear winner prunes to one; a near-tie keeps both.
max_candidates_per_subsystem
3
Caps distribution-tail noise beyond the real contenders.
ambiguity_flag_threshold
0.7
Runner-up within 70% of the winner's share reads as genuinely competing. Makes explicit a value already implicit in the code.
Honest caveat. 0.7 is reasoned, not measured — naming it and testing both sides of the boundary is the mitigation, not a substitute for eventually tuning it against data.
How we know it works
Tests — and the guard that matters most
Intent scenarios
- voting.close_candidates_flagged GAP → MEET
- voting.ensemble_disagreement — new: Bayesian favors A, Pattern favors B → a real near-tie via disagreement, not identical profiles
- a confident-winner control → ambiguous: False
Cataloger unit tests
- pruning: confident → 1 candidate; near-tie → 2; cap at 3
- distributed tally & ambiguity computation
- ambiguity_flag_threshold boundary, both sides
- exclusions & hierarchy still respected
The load-bearing check. Unlike F2's isolated module, distributed voting perturbs every composite. The 10 MEETS must hold, each clear winner still reporting ambiguous: False. Any wobble is a signal to tune candidate_keep_ratio — never to weaken a scenario.
Decisions & boundaries
What's locked, what's out
Shipped
- Full F1 + F3: compute ambiguity right and surface it everywhere.
- Distributed mass voting; ambiguity is a property of the ensemble.
- Change lives only in cataloger.py + neuron.py.
Out of scope
- Sub-system internals; the needs_more_info flag.
- Renormalising vote mass across sub-systems (a known, bounded imprecision).
- Multi-hop routing ambiguity nuances beyond passing the flag through.