[design·data·P2] The driver taxonomy has one level, so several holdings riding the same sub-driver inside one sector read as ordinary sector exposure

[atomchung]

Privacy boundary

Found on an owner-live walkthrough with real holdings. De-identified: no ticker,
amount, or date appears.

The gap

DRIVER_FALLBACK assigns one broad sector label per instrument, and
max_sector_pct is computed over those labels. Inside a single broad bucket,
holdings that ride genuinely different cycles and holdings that ride the same
cycle are indistinguishable.

The observed case: a book held several names under one broad hardware sector
label, of which a subset are pure plays on one component cycle — the same supply,
the same pricing curve, the same downturn. To the engine they are simply members
of the same sector bucket, no closer to each other than to any other member. To
the owner they are one bet held twice.

The owner named it directly while reading the card:

there's no way to tell that these two are essentially the same play — they're
both memory.

Why it is not just a nicer label

Three engine-owned conclusions rest on the granularity:

1. max_sector_pct understates the tightest real concentration, because the
   sub-driver group inside the bucket is invisible.
2. The card's headline risk line names the broad sector as the concentration.
   A user who reads that and diversifies within the bucket — away from the
   named sector's most obvious member and into another member sharing the tighter
   driver — has followed the card and increased the risk it flagged.
3. consider's post-trade concentration numbers inherit the same bucket, so
   an add that doubles down on a sub-driver reads as ordinary sector exposure.

This is the mirror image of unclassified_drivers, which the honesty ledger
already discloses: that key says concentration may be understated because some
instruments carry no driver at all. Coarse buckets understate it for instruments
that carry one.

Not the same as the existing look-through issue

#599 covers ETF look-through — a single instrument whose internal composition is
invisible. This is the opposite direction: several distinct instruments whose
shared driver is invisible because the taxonomy has only one level.

Possible shapes, none ruled in

• A second, narrower driver level below the sector, populated for the cases where
  it changes an answer rather than exhaustively.
• Keeping one level but allowing an instrument to carry more than one driver tag,
  with concentration computed over tag intersections.
• Leaving the taxonomy alone and disclosing the limitation, the way
  unclassified_drivers does — cheapest, and honest, but it leaves conclusion 2
  above intact.

Acceptance criteria

• A ruling on which shape, recorded here.
• Whichever is chosen, a book holding several same-sub-driver names inside one
  sector must not be able to read the card's diversification advice and act on it
  in a way that concentrates the tighter driver further.
• If disclosure is the chosen answer, the honesty ledger gains a key for it and
  the card states it; silence is not an option once the limitation is known.

Related

• [design·P1] The market read is bound to weekly_review, so the holdings-only cold-start user — the one its gate was written for — can never receive it #772 — what a holdings-only user is owed; the sub-driver is part of "what is
  this exposure sensitive to".
• [bug·snapshot·P1] The snapshot card withholds five facts the route already supports — the adapter zeroes every card field instead of filling the subset it can #771 — the same route's withheld facts.
• [feat·data] consider has no ETF look-through — a concentrated single-sector active ETF is invisible to ai_pct/max_sector_pct #599 — ETF look-through, the other direction of the same limitation.

[atomchung]

Unblocked — the risk definition is ruled (2026-08-02)

#781 is settled: risk includes what is about to happen to a position, and the owner calls the forward-looking catalyst the core concern. Current structure alone is an incomplete risk answer; the recorded book is the selector that decides which forward-looking events matter to this user.

Ruling recorded at #781 comment 5156514056. Two constraints it carries into this issue:

1. A dated published fact is in; a forecast or price target is out. Boundary 2 of AGENTS.md is unchanged and now carries most of the weight.
2. Deterministic first. Anything retrievable deterministically (scheduled earnings dates, index events, ex-dividend, lockup expiry) enters through the engine with a validator. The researched half waits on the live voice/answer gate owed by [process·voice·P2] A live consider answer passes through no voice check — three output-voice violations shipped in one owner-live session with a fully green trace #782 — the engine's numbers pass gates, the agent's own sentences currently pass none.

Re-scope this issue's candidate shapes against that definition before proposing one; the shapes recorded here were written when the definition was still open.

[atomchung]

Referenced from the #776 convergence proposal (comment 5156586064), §4.

That proposal is deliberately not blocked on this ruling — forward-dated events group by whatever concentration unit the engine already computes, so a second taxonomy level is inherited for free if it lands. What it does give you is the value of this decision stated as one concrete sentence rather than in the abstract:

• with one level, the strongest sayable thing is "two of your holdings report this week";
• with a second level it becomes "two holdings riding the same cycle report in the same week" — which is the version that changes a decision.

That is the sub-driver's payoff under #781's forward-looking definition of risk, and it is additive to the three conclusions already listed in this issue's body.