Con-den-1: findings and tables

Con-den-1: findings

Descriptive census of convertible instruments fleet-wide, run 2026-08-26 by the Code desk on branch study/con-den-1, cut from origin/main at f3d557c. Task id #C-126, provisional.

The definitions and the moneyness bands were fixed in REGISTRATION.md and committed at f9923f5 before the first outcome value was read. That commit is the receipt for the blind ruling. Section 8 below lists every decision taken after the read, because a registration is worth only what its exceptions are willing to admit.

No seal, no gate beyond canon and the integrity check. The uncovered set publishes at its true size and is not narrowed to make a table look finished.

1. What the census covers

72 instruments across 13 issuers. 70 come from the instruments tab, 2 are reached through deal history alone and carry no terms.

The two coverage dimensions are reported separately because they fail for different reasons and an instrument can be measurable on one and not the other. 19 instruments are known to have delivered something; only 15 carry a share count from a filing or a deal row.

Delivery coverage is the binding constraint on everything below. Every rate in this note rests on 26 instruments. The intervals are wide, they overlap, and no result here should be read as settled. This is a census, and its first finding is how much of the fleet cannot yet be measured.

2. Delivery by moneyness at issuance, the registered primary

Moneyness at issuance orders these outcomes weakly, and it does not order them monotonically. Deep OTM instruments delivered at a higher rate than Near instruments, on both the count and the share-weighted measure. All three intervals overlap across their whole width. On this population, an issuer's choice of strike relative to spot on the pricing day carries little information about whether the instrument later put shares into the basic count.

That is a result about the classifier, not about convertibles, and section 5 shows the ordering itself is not robust to where the band cut is placed. It is what the registered primary measures, and it is reported as the headline because the registration says so.

3. Delivery by instrument class

Warrants are the largest class by instrument count and the least measurable: 6 of 30 carry a delivery figure. The one preferred row with coverage is a single instrument and the 0.9985 should be read as one observation, not a rate.

4. Sensitivity 1, time-weighted moneyness (labeled, never the headline)

Registered as a sensitivity and run as one. For each banded instrument, the mean of S_t / K across every trading day from issuance to 2026-08-26 or settlement, re-banded on that mean. 49 of the 50 banded instruments carry enough history; H100 is excluded because its spot rides a fallback cross that would have to be applied daily.

Migration between the two views is heavy, which is why the two disagree:

The life-of-instrument view orders the outcomes where the at-issuance view does not, 0.786 against 0.222 on the count and 0.8911 against 0.1659 on the share weight. The intervals still touch between 0.52 and 0.55, so this is a difference in the expected direction and not a separation.

This sensitivity is not promoted. It is reported because the registration required it to run, and because the gap between the two views is the most substantial thing the census found: what an instrument was worth over its life tracks delivery better than what it was struck at. Whether that survives a larger covered set is the open question, and 26 instruments cannot answer it.

5. Sensitivity 2, band-choice cuts (labeled, never the headline)

The registered 30 percent premium cut against the 20 and 40 percent alternatives, ITM boundary held at M = 1.00. Full tables in FINDINGS-TABLES.md.

The band choice changes the answer, and this is the most awkward result in the study. At the 20 percent cut the ordering is monotone, 1.000 then 0.667 then 0.412, and the classifier looks like it works. At the registered 30 percent cut and again at 40 percent the ordering inverts, and at 40 percent the inversion is at its widest, 0.200 against 0.700.

Five covered instruments sit between 30 and 40 percent of premium and they carry the whole effect: CASH3-WARRANT-S4 and -S5 at 30.7 and 35.9 percent, which delivered nothing, and MSTR_2028CN, MSTR_2032CN and MSTR_2030ACN at 34.0, 38.3 and 38.8 percent, which also delivered nothing. Moving the cut past them shifts five non-delivering instruments out of Deep OTM and into Near, which lifts the Deep OTM rate and depresses the Near rate at the same time.

Two things follow. The non-monotonicity reported in section 2 is not robust to the cut, so it should not be read as a property of convertibles. And the registered cut is not the one most favourable to the classifier, which is the reason for fixing it in advance. Had the cut been chosen after seeing these tables, 20 percent was available and would have told a cleaner story than the truth supports.

The ITM row never moves, since the M = 1.00 boundary is untouched by the premium cut.

6. The Cal-D1 holder-sitting hypothesis, checked against the broader population

Cal-D1 scored a Deep ITM band at n 17 and returned a per-quarter conversion statistic of 0.0168, Wilson 0.0015 to 0.1648, on MTPLF, measured at the mark. The reading offered was holder-sitting: holders of deeply in-the-money convertibles did not convert.

The units do not match and no arithmetic here reconciles them. Cal-D1 measured a per-quarter conversion rate at the mark. This census measures a lifetime delivery rate banded at issuance. A lifetime rate aggregates many quarters, so the two cannot be compared directly, and the Cal-D1 checkpoint already carries an unexplained factor of roughly 8 between its lifetime prior and its per-quarter statistic. This study was not designed to resolve that factor and does not resolve it. What follows is a comparison of differently-united quantities, stated as one.

Taking the closest available analogue, the time-weighted ITM band, the aggregate does not reproduce holder-sitting: 11 of 14 covered instruments delivered, 0.8911 of potential shares. But the aggregate conceals the whole result, and the split is the finding:

Holder-sitting is issuer-specific on this population, not a population-wide regularity. Every covered ALCPB convertible that spent its life in the money delivered in full. Every MSTR convertible that spent its life in the money delivered nothing.

The MSTR zero is a structural gate, not a holder preference, and this is the qualification the hypothesis most needs. The Q2 2026 10-Q, accession 0001050446-26-000044, states that the notes were not convertible during H1 2026 and that no conversion requests were received. Unrestricted conversion opens 2028-03-15; before that date conversion is conditional on a 130 percent price test, a trading-price test, a redemption call, or specified corporate events. The holders of those six series were not declining to convert. For the observation window they could not convert. On conversion the company further elects cash, stock, or a combination at its own election, so under Definition 1 even a future conversion may deliver nothing.

An at-the-mark measure cannot see that gate, because the gate is in the indenture and not in the price. Any reading that treats an unconverted in-the-money instrument as evidence of holder behaviour will misattribute a contractual restraint to a holder choice wherever the two coincide. This does not overturn Cal-D1, whose population and units are different. It identifies a confound that a per-quarter conversion statistic measured at the mark cannot separate on its own, and it suggests that a conversion-eligibility flag would be worth carrying beside the moneyness.

The gate_type column is now in the census. It is populated on 37 of 124 instruments-tab rows and on no ALCPB convertible bond, so it cannot carry this distinction fleet-wide as of this run. That referral is answered; see the forward notes in section 11.

This diagnosis, the contractual gate with its accession, is carried forward as the D2 restatement direction. Section 11 records it.

7. The uncovered set, at its true size

46 of 72 instruments carry no delivery figure, and 22 carry no at-issuance band. Reasons, as recorded per row:

Four of these are findings in their own right rather than gaps in the sources:

• pre_listing_no_market, 5 instruments. OBTC3 and TSWCF struck convertibles and warrants before a public market in their common existed. At-issuance moneyness is undefined for these, not merely unmeasured, and no data collection will supply it. TSWCF's pre-IPO warrants later delivered 21,637,603 shares, 0.6672 of potential, against a 2.5 pence strike.
• moving_strike, 2 instruments. MTPLF's moving-strike warrants reset to the prevailing price, so they carry no fixed at-issuance strike to measure a premium against. One of them, MTPLF_MSW_27TH, has delivered 5,270,000 shares across four exercise events.
• nominal_strike_prefunded, 2 instruments. A pre-funded warrant carries a nominal exercise price because the holder has already paid the equity value. Measuring a premium against it returns a number with no meaning, so these are recorded and left unbanded rather than read as deeply ITM.
• delivery_size_unsourced, 1 instrument. OBTC3's 2030 convertible has a recorded conversion event carrying no share figure. Delivery is known to have happened and its size is not. It publishes as uncovered. Writing it as a zero would have asserted the opposite of what is known, which is the failure mode the registration fixed in advance.

Of the 26 delivery figures, 15 are VERIFIED against a filing or a deal row and 11 are EST. The 11 are ALCPB and OBTC3 rows where the instruments tab records a converted status but no event carries the count; full conversion at the struck rate is inferred and marked as an inference. Every ALCPB delivery figure in section 6 is of this kind, so the 1.0000 there is an inferred full conversion, not a filed share count.

8. Decisions taken after the read

Each of these was made after outcome values had been seen. None changes a definition or a band.

1. Scope discovery extended to the warrant classes regardless of the is_convertible flag. The flag is inconsistent within a single class: DDC_ANSON_WARRANTS carries TRUE and CASH3-WARRANT-S1 carries FALSE for the same kind of instrument. Keying scope on the flag alone would have dropped 23 warrants on a data-entry artifact.
2. MSTR's six convertible notes are joined across a row pair. The claims row carries face and strike, the share_delivery companion carries share_count. The census row is the claims row and the companion is joined onto it, so the six are counted once. Keying scope on is_convertible alone would have missed them entirely, since the companion rows carry the flag blank.
3. The split-basis guard is evidence-based and replaced a plausibility window. A first cut flagged any moneyness outside 0.05 to 3.0 as a basis error. That was wrong. It flagged ALCPB, MTPLF and TSWCF rows whose issuers carry no split at all and whose high moneyness is the real finding: a strike fixed at a program's inception and drawn down months later, or a pre-IPO strike measured after listing. The guard now reads the split events and flags only where a split falls between issuance and today and the issuer's strike basis is unestablished. That leaves 3 NAKA rows flagged, across a 1-for-40 reverse split on 2026-05-22. MSTR is exempted on evidence: its December 2020 notes priced at a $397.99 conversion price and the tab carries 39.8, exactly one tenth, so its strikes are recorded split-adjusted and match the series.
4. pre_listing_no_market was separated from a missing quote, and pre-funded warrants were given their own reason. Both distinctions were visible only once the data was in hand.
5. A delivery event carrying no share figure yields no delivered figure. The first cut summed such events to zero and published the zero. That contradicted the registration and was corrected.

9. Data-quality observations for the data desk

Reported, not fixed. Nothing in this study writes to the sheet or to any served surface.

• SWC.L quotes in pence, and the site already handles it. The archived price spine records the symbol as GBp and the live Yahoo metadata agrees, while config/tickers.js declares feedCurrency: 'GBP'. That declaration is correct rather than wrong: api/stock-price.js:94 divides LSE quotes by 100 before serving them, with a comment noting that the label varies between GBX, GBp and GBP by session, so feedCurrency describes what the endpoint returns and not what Yahoo sent. No defect, and it is recorded here because an earlier cut of this note reported one. The factor of 100 is live only for code reading Yahoo directly, which is what this census does, and the census divides by 100 explicitly for that reason.
• stock_price_at_deal is not reliably in the instrument's currency. H100's convert issuance row carries 0.724 against claims_currency SEK, which is the EUR quote, since the SEK strike is 8.48. This census drives spot from the currency-declared price series and uses the deal price only as a cross-check, for that reason.
• NAKA strikes sit on mixed share bases. Across the 1-for-40 reverse split the recorded strikes do not reconcile to the adjusted series on a single basis, so 3 NAKA rows publish unbanded.
• BSTR appears in the deals tab with a convertible and a preferred issuance and has no instruments-tab row. Two census rows carry it with terms blank.
• OCA_B01_BLOCKSTREAM and OCA_B01_BLOCKSTREAM_CONVERTED carry different strikes, faces and dates and are treated here as distinct tranches. If they are two records of one instrument the ALCPB potential-share total is overstated. This is a question for the ALCPB anchor, not a finding.

10. Limitations

• 26 delivery observations govern every rate in this note. Wilson intervals are reported on all of them and they are wide.
• 11 of the 26 delivery figures are inferred from a converted status rather than read from a filing.
• Issuance dates from instruments.launched are month-precision on 20 rows. A month of price movement can cross a band boundary, and no attempt was made to source exact pricing dates from filings.
• The census is unweighted by size. ALCPB contributes 23 of 72 rows and dominates the ITM counts; MSTR contributes the largest potential-share balances.
• Terms as struck are taken as recorded. Anti-dilution adjustments after issuance are not traced.
• Delivered-share counts from the instruments tab cover Tier 1 only, per #A-66. No Tier 2 authorization is netted into any figure here.

11. Forward notes, ruled 2026-08-26

Both notes were ruled by the operator on receipt of this study and are recorded before the file freezes. They state where the findings travel. Neither changes a figure above.

11.1 The MSTR zero of three is the D2 restatement direction

The contractual-gate diagnosis in section 6 is carried forward as the direction for D2's restatement, with the accession cited as its source: Q2 2026 10-Q, accession 0001050446-26-000044, which states that the notes were not convertible during H1 2026, that no conversion requests were received, and that unrestricted conversion opens 2028-03-15.

What the direction rests on: three MSTR convertibles spent their life in the money and delivered nothing, and the reason is in the indenture rather than in the price. A measure taken at the mark cannot separate an instrument whose holders declined to convert from one whose holders were barred from converting. Where the two coincide, the at-the-mark reading attributes to holder behaviour what the conversion condition produced. D2's design assumption should be restated against that, not against the aggregate delivery rate, which conceals it.

The caution from section 6 travels with the direction. This census measures a lifetime delivery rate banded at issuance and Cal-D1 measured a per-quarter conversion rate at the mark. The two are differently united, the Cal-D1 checkpoint already carries an unexplained factor of roughly 8 between its lifetime prior and its per-quarter statistic, and nothing in this study resolves that factor. The direction is the diagnosis, not the arithmetic.

11.2 The gate_type referral is answered

Referral 6 asked for a conversion-eligibility signal carried fleet-wide. It is answered: gate_type populates fleet-wide by anchor dispatch, with the vocabulary ruling to follow from copilot. This study is the consumer that priced it, and is cited as such.

What this study establishes about the column's value, for whoever writes the vocabulary: the column was populated on 37 of 124 instruments-tab rows at the time of this run and on no ALCPB convertible bond, which is exactly the pair the section 6 result turns on. MSTR's six convertibles carry gate_type holder, and the ALCPB convertibles that delivered in full carry nothing. The single distinction that changes the reading, whether an instrument could be converted during the window, was therefore available on one side of the comparison and absent on the other, and had to be recovered by reading a filing note rather than a column.

The vocabulary ruling is copilot's and is not anticipated here. One observation is offered as input and nothing more: holder on the MSTR rows marks who holds the power, and the fact that decided this study was whether the power was exercisable at all during the observation window. Those are separate facts, and the instruments-tab note for MSTR_2028 already keeps them apart, recording the 2028-03-15 free-conversion date as a holder power while leaving lockup_expiry_date blank because that column is registered as a holder restraint. A vocabulary that collapses the two would lose the distinction this census needed.

Artifacts

Con-den-1: findings tables (generated)

Census as of 2026-08-26. Rows: 72.

Coverage

Headline: delivery by moneyness at issuance (registered primary)

Delivery by instrument class

Sensitivity 1, time-weighted life-of-instrument moneyness (labeled, not the headline)

49 of 50 banded instruments carry enough price history to time-weight.

Delivery by time-weighted moneyness

Migration between the two views

Sensitivity 2, band-choice cuts (labeled, not the headline)

Premium cut at 20 percent

Premium cut at 30 percent (registered)

Premium cut at 40 percent