Bloom Energy [BE] — Trade Construction & Risk Management

⚠️ SUPERSEDED IN PART — 2026-07-29

The position verdict in this document is retired. Under the current framework (references/criteria.md, 2026-07-29) the memo outputs an analysis, not a position. Whether an analysis justifies a position is a question about a particular book, and two books answer it differently.

The Gate block and the Gate 4 expected-return arithmetic below are also superseded, by the named Criteria (each with a type: BINDING or MEASURED, returning PASS / FAIL / INDETERMINATE), the reverse-DCF implied-path test, and a separate 12-month target.

→ Current analysis: Bloom_Energy_Criteria_and_Valuation_2026-07-29.md

Everything else here — the research, the evidence, the mechanism work — stands. Residual references to "Watchlist" in the prose below are the historical record of the 2026-07-28 assessment and are left intact deliberately.

Phase Space Research | July 28, 2026 | Task 5 of 9 | Skill v1.5.1 (six gates; archetype declaration; Gate 4a dual-horizon; Path B compounder test; C4 range reporting; C6 trigger audit; C8 tie-out; Task 5.5 adversarial review)

Re-run under v1.5.1, promoted 15:14 today. This memo was first written under v1.4.2 and has been re-run against the new rules: archetype declared, Gate 4 reported at both the thesis horizon and a 12-month reference, Path B tested explicitly, and the verdict submitted to memo-critic at Task 5.5. The verdict was PRE-REGISTERED as WATCHLIST before this re-run, with the condition that Path B must not rescue it and that momentum/PEAD may alter entry staging only, never the verdict.

Prerequisite: Tasks 1–3 complete. Probability-weighted value $126.60 (undiscounted expected FY2029 value, prob-weighted) (Ke 10%), against $183.90 aftermarket / $166.84 close.

0. PORTFOLIO CONTEXT (the portfolio-book contract)

portfolio_book.json: no open positions; 100% cash. Cash hurdle 4.7%, max single-name 5%, pairwise correlation disclosure threshold 0.60, position hard stop −2% of book.

Correlation — measured, not assumed (251 daily log returns, Alpaca SIP):

Nothing breaches 0.60. BE sits in the AI-infrastructure complex (SNDK/MU/SMR/NBIS 0.46–0.53) and is effectively orthogonal to the healthcare book and to MSFT. Note BE–SMR +0.487 — both are "power for AI" theses; treat any BE + SMR pairing as one allocation decision.

Capital competition: with an empty book, the comparison is directly against 4.7% risk-free.

0.5 ARCHETYPE — DECLARED BEFORE THE GATES (v1.5.0)

INFLECTION

Not a judgement call. BE meets every INFLECTION criterion: gross margin 33.4% and rising, operating margin expanding fast (−0.9% → 17.1% GAAP in four quarters), growth accelerating (+165.5%), and — decisively — value almost entirely in the terminal period. It was thin-to-negative on operating margin as recently as Q2'25.

What the archetype obliges: 1. Gate 1 uses the trajectory standard, not level-based quality measures. Applied below. 2. Task 3 must report terminal-value sensitivity explicitly. Done — §3 is a full two-dimensional grid over exit multiple and cost of equity, not a point estimate. 3. Say what the framework can and cannot price. Stated plainly: this framework cannot price BE's terminal period, and neither can anyone else. A DCF here is dominated by assumptions about 2029+ that no method resolves. What it can do is invert the question — and the reverse-DCF is the honest instrument: it does not claim to know what BE is worth, it states what you must believe for today's price to work. That is the only defensible use of a DCF on an inflection name, and it is why the verdict rests on the reverse-DCF rather than on the probability-weighted target.

Why this matters for calibration: item B8 looked like a sector-specific multiple problem. It was the INFLECTION archetype being valued with COMPOUNDER machinery. BE is the same archetype outside healthcare, which is the generalisation B8's own text asked for.

1. FOUR CONCLUSIONS, KEPT SEPARATE

1. Fundamental — strong and improving, genuinely. First $1bn quarter, revenue +165.5%, GAAP operating income $182.2m against a −$3.5m loss a year ago, +$226.4m operating cash flow, third consecutive guidance raise with EPS guidance up 31.7% at the midpoint. Nothing here is deteriorating.
2. Expectations — the Street is positive and has just moved further up. Mean target $289.63 (9 Buy / 11 Hold); JPMorgan to $346 from $267. The FY2026 guidance raise landed 8.3% above prior consensus revenue and 24.4% above prior consensus EPS.
3. Valuation — this is where it breaks. 68x FY2026E non-GAAP EPS, 14.7x EV/sales, 70x EV/EBIT. Probability-weighted FY2029 value $126.60.
4. Portfolio — see the gates.

2. THE SIX HARD GATES

Binding constraint: Gates 2A, 2B, 3 and 4. Not momentum, and not feasibility.

2.1 The momentum-crash caveat, stated explicitly (Daniel & Moskowitz 2016)

Top-decile 12-1 momentum plus 110% realised volatility plus a price 37% below its own 50-day is the precise configuration in which momentum strategies suffer their worst drawdowns. The momentum reading is simultaneously top-decile and maximally fragile. These are not conflicting facts — they are the same fact, and Gate 6 "passing" here should not be read as tape support for a long.

3. GATE 4 — E[R] AS A RANGE, WITH THE FLIP POINT (v1.4.2 item C4)

Unlike GOOGL — where the decision turned on a single judgement input (bear weight 18% vs 30%) — BE's Gate 4 result is stable across the whole plausible parameter space. Both levers are shown.

CORRECTED BASIS (see §8). The original grid discounted scenario payoffs at Ke and then required the result to beat a 4.7% hurdle, which double-charges the required return. The identity is E[R]_reported = (E[R]_true − Ke)/(1+Ke) — so every Ke column encoded the same expected return and the grid had one degree of freedom, not two. "Stable across the parameter space" was an affine relabelling. Gate 4 now compares the undiscounted expected holding-period return against the cash hurdle.

Expected FY2029 value (probability-weighted): $126.60 against a $183.90 spot.

Break-even base-case exit multiple on FY2029E EPS: 40.1x — and it is FLAT across Ke, because the cost of equity is no longer double-counted. (The original memo reported 53.3x–79.6x rising with Ke; that variation was the artifact, not a finding.)

⚠ THE FLIP POINT IS UNREACHABLE, AND THAT IS THE HONEST C4 STATEMENT

On the corrected basis BE fails at every bear weight in [0, 1]. At a 0% bear weight — assuming the bear case simply cannot happen — E[R] is still −4.2%, failing by 8.9pp. No probability weighting produces a pass. This is materially different from GOOGL, where 18% and 30% bear weights were both defensible and gave opposite answers, and it is why item C4's "hurdle inside the range" category does not apply here.

The correction cut the shortfall from 31.4pp to 16.4pp — nearly in half — and the verdict did not move. That is the useful test: an error found by the critic, running in the name's favour, and the conclusion survived it.

What the price requires, restated on the corrected basis: for BE to return 4.7% annualised from $183.90, its FY2029 earnings must be capitalised at roughly 40x, as a ~$9bn-revenue company growing ~20%. The reverse-DCF says the same from the other side: even at a 50x exit and a 12% discount rate, the price needs FY2029 revenue of $8.8bn — a 29.6% CAGR for three years on top of a year that doubles.

3.5 PATH B — THE COMPOUNDER PATH (v1.5.0) — TESTED AND FAILED

Path B lets Gate 2 be satisfied without a variant, but only if all five conditions hold. Tested explicitly because the pre-registration named it as the one route that could overturn the verdict.

PATH B FAILS at condition 2, and not narrowly. The condition exists precisely because, without a variant, the whole return depends on the valuation being right — so it must not be a coin flip. BE is not a coin flip in the wrong direction: it is a clear negative at every probability weighting.

Pre-registration honoured. The stated condition was: "Path B must not rescue it — Path B needs Gate 4 to pass robustly with the hurdle outside the range, and BE fails by 13–34 points. If Path B produces a Long on BE, that's a re-scope trigger, not a result." Path B did not produce a Long. No re-scope trigger fired.

3.6 MOMENTUM AND PEAD ARE TIMING INPUTS, NOT VERDICT INPUTS (v1.5.1)

This is the resolution of item B9, and it applies to BE more sharply than to any other name in coverage, because BE has the strongest short-horizon signal set the system has measured: 12-1 momentum +386% to +800% (window-sensitive; see note) (top decile), a +29.0% / +90% revenue / EPS surprise, and an FY2026 EPS guidance revision landing +24.4% above prior consensus.

Under v1.5.1 these govern WHEN to enter a position the thesis already justifies, never WHETHER. BE's thesis does not justify a position at either horizon. Therefore:

• The momentum and PEAD readings are recorded, and they change nothing.
• Had BE cleared Gates 1–5, this signal set would have argued for an immediate rather than staged entry.
• It does not, so they argue for nothing.

Pre-registration honoured on this point too: "Momentum and PEAD may change entry staging only, never the verdict." They did not change the verdict.

And the honest note for the calibration file: this is what B9's resolution costs. Before v1.5.1 I argued the framework "measures short-horizon alpha it structurally cannot act on, which wastes the measurement." The v1.5.1 answer — they are timing inputs — is coherent, but on BE it means the strongest short-horizon signal set in coverage is measured, reported, and then correctly ignored. That is the right outcome if the multi-year thesis is genuinely negative. It is the wrong outcome if 12-1 momentum of +386% to +800% carries real information about the next two quarters that the book is simply choosing to forgo. v1.5.1 makes that a deliberate choice rather than an accident, which is the improvement. It does not make it costless, and the ledger is the only thing that can eventually price it.

4. POSITION VERDICT — RETIRED

This section previously read "WATCHLIST — long bias, not actionable". The position verdict is retired as of 2026-07-29: the memo outputs an analysis, not a position. The evidence below is kept; the conclusion is dropped. See Bloom_Energy_Criteria_and_Valuation_2026-07-29.md.

Not Long: Gate 2 fails on both branches (no estimate variant exists at all; the duration variant cannot be sized against an undated and now-unupdated backlog), Gate 3 has no variant to resolve, and Gate 4 fails by 31 points annualised across essentially the entire defensible parameter space.

Not Short — and this is important. Every empirically-documented short-horizon signal points the other way: top-decile momentum (+386% to +800%), a +29%/+90% revenue/EPS surprise, a +24.4% upward EPS guidance revision, and favourably negative accruals. Shorting into that combination is precisely the negative-EV trade this framework exists to prevent (and Gate 5 would bind anyway — borrowing a 110%-vol, top-decile-momentum name into positive drift is the definition of an expensive short). A rich multiple alone is never a short thesis.

What separates BE from the rest of the coverage: it has the strongest factor profile this system has underwritten — and still fails. That tension is logged as a calibration item rather than resolved by loosening a gate.

5. CONVERSION TRIGGERS — with the joint-probability audit (v1.4.2 item C6)

Disjunctive by design. Per C6, conjunction is reserved for cases where two genuinely different kinds of evidence are both required. Each trigger below independently repairs a binding gate.

Joint probability that at least one fires: ~65–70%. This is a reachable watchlist, not a NO POSITION in costume. Note T1 is unusually reachable here because the volatility is so extreme — a −45% move is a routine 3-month event for this security, not a tail.

Invalidation of the long bias: a weekly close below the 200-day (~$177) accompanied by evidence the H2 guide is tracking to the low end ($3.9bn, which implies sequential revenue decline), or any restatement/SEC comment-letter traffic on the warrant or backlog accounting.

6. IF IT CONVERTS — SIZING AND VEHICLE

Sizing. Conviction is Low (Gate 2 fails on both branches; Gate 4 fails by 31 points — only Gate 1 is genuinely strong). Volatility is Extreme: 110.6% realised 1-year, against the skill's ">45% = High" threshold — this is 2.5x the High threshold and the most volatile name in the entire coverage universe. Low conviction × Extreme volatility → 1.0% of book maximum, and the −2%-of-book hard stop binds at roughly a 50% adverse move on a 1% position, which is inside one 3-month sigma. That is the real sizing constraint: at this volatility, a position large enough to matter is a position large enough to breach the book's own stop.

Vehicle — defined-risk spread, per the skill default, and here the default is right. Recomputed implied volatility is ~1.19x three-month realised (137% vs 122.5% at the 80-day tenor). That is the ordinary variance-risk-premium configuration in which buying premium is systematically expensive — the exact inverse of GOOGL, where implied sat at 0.81x realised and long premium was the cheaper expression. On conversion, prefer a call debit spread (e.g. Jan-2027, long ~$180 / short ~$240) over outright calls or equity, and state the implied-versus-realised ratio in the ticket.

DATA GOTCHA — Alpaca's reported impliedVolatility field is unreliable and must not be used. For BE it returned call IV ~110% and put IV ~155–170% at the same strike and expiry — impossible if put-call parity holds. It does hold: measured C−P against S−Ke^(−rT) agrees to under $1 at every strike tested. A Black-Scholes recomputation from mid prices gives call and put IV agreeing to within ~2 points (146.2% vs 147.1% at the Sep-18 $160 strike). Recompute IV from mid prices and check parity before using any Alpaca IV number in Gate 5. Filed to references/alpaca-options.md.

7. BOOK WRITE-BACK

Watchlist entry: BE / long / converts_if T1 OR T2 OR T3 OR T4 / joint ~65-70%. No ledger entry — a Watchlist is not a recommendation, and logging one would corrupt the calibration record.

OUTPUT FORMAT (feeds the Task 7 trade ticket)

POSITION VERDICT: RETIRED 2026-07-29 — the memo outputs an analysis, not a position.
                  See Bloom_Energy_Criteria_and_Valuation_2026-07-29.md for the current Criteria block.
The gate lines below are the 2026-07-28 record and are superseded.
GATES: 1(Mechanism): PASS — interconnection-queue bypass is structural (Tier 3), corroborated by
       accruals -11.9% of assets, F-score 7/9, GM 26.7%->33.4% YoY
       2A(Estimate variant): FAIL — NO VARIANT EXISTS; house forecast IS guidance IS consensus
       2B(Duration variant): FAIL legs 3+4 — backlog quantified ONCE (Feb-26), undated, includes
       anticipated ITC credits, ~70% is multi-decade O&M; and it is already in a 68x multiple
       3(Catalyst): FAIL — real dated events, but none resolves a 2029 exit-multiple disagreement
       4(Expected Return): FAIL — -11.7% annualised, -16.4pp vs hurdle, stable across the grid
       5(Feasibility): PASS — $4.6bn ADV, deep chain; but IV ~1.19x realised, premium is EXPENSIVE
       6(Momentum/Tape): PASS precariously — 12-1 top-decile on any window convention, but -37.1% vs 50-day and
       110% vol = the Daniel & Moskowitz momentum-crash configuration
CURRENT: $183.90 aftermarket ($166.84 close) | PROB-WEIGHTED FY2029 VALUE: $126.60 | STREET: $289.63
E[R]: -11.7% annualised at house assumptions | break-even base exit multiple is 40.1x, FLAT across Ke
REQUIRED BELIEF: BE still trades at ~40x earnings in 2029 as a ~$9bn-revenue, ~20%-growth company
CONVERSION: T1 price ~$110, OR T2 backlog re-quantified WITH duration, OR T3 two more quarters at
       the high end, OR T4 the Q2 10-Q changes the underlying revenue trajectory. JOINT ~65-70%.
SIZING ON CONVERSION: 1.0% of book MAX (Low conviction x EXTREME 110% vol — 2.5x the High threshold)
VEHICLE ON CONVERSION: call DEBIT SPREAD — implied vol is 1.19x realised, the exact inverse of GOOGL
THESIS IN ONE SENTENCE: The best factor profile this system has underwritten — top-decile momentum, a
       +90% EPS surprise, a +24.4% guidance revision and negative accruals — attached to a price that
       already requires BE to be a ~40x business in 2029, with a backlog it has quantified once.

8. TASK 5.5 — ADVERSARIAL REVIEW (memo-critic, Mode A)

RULING: UPHELD WITH CONCERN.

The WATCHLIST verdict is correct and I would not change it. Gate 2A fails at an exactly zero variant — the house FY2026 forecast is the company's guidance is the Street's number, $4.05bn / $2.70 on both sides — which is the cleanest Gate 2 failure in the 22-name record (ISRG failed at 0.2%, NTRA at 3.9%; this is 0.0% by construction). Path B is the only route past Gate 2 for a long and it requires Gate 4 to pass robustly; Gate 4 fails at every probability weighting including a 0% bear weight. Nothing I can attack changes the answer.

But the Gate 4 arithmetic that the memo leans on is wrong by 15.0 percentage points a year, and item B9 in CALIBRATION_WATCH.md is now built on the wrong number.

8.1 The binding gate, and the dispositive one

Four gates are recorded as failing (2A, 2B, 3, 4). Only Gate 2A is dispositive. 2B and 3 are consequent — Gate 3 fails only because there is no variant for a catalyst to resolve, and 2B fails legs 3/4 downstream of the same absence. Gate 4 is also-failing, not binding: flip Gate 4 to PASS and BE is still a Watchlist, because a long without a variant must go through Path B, and Path B's condition 2 is Gate 4 passing robustly. There is no ordering of these four in which Gate 4 is the constraint. The memo's "Binding constraint: Gates 2A, 2B, 3 and 4" over-counts by three.

8.2 The number I am attacking: Gate 4 double-charges the required return

build_model.py:138 computes

3-yr annualised E[R]  =  (K10 / spot)^(1/3) - 1        where K10 = Σ pᵢ · FY29valueᵢ / (1+Ke)³

The scenario payoffs are discounted at the cost of equity (20.4%) and then the result is required to beat the 4.7% cash hurdle. That charges a required return twice. Algebraically the reported figure and the true expected return are related by an exact identity:

E[R]_reported = (E[R]_true − Ke) / (1 + Ke)

which is why every column of the §3 Ke grid encodes the same expected return. The grid has one degree of freedom (the exit multiple), not two. "The conclusion does not depend on the discount rate" is true — but because Ke is an affine relabelling carrying no information, not because the result is robust. The grid is presented as robustness evidence and is not.

Corrected, using the memo's own scenario table:

The memo's headline contradicts its own scenario table: −26.7%/yr for three years implies BE is worth $72.43 in 2029, while the model says the probability-weighted 2029 value is $126.60.

Two internal tells confirm the formula is the outlier, not the corrected number. (i) Trigger T1 sets the actionable entry at ~$110; the memo's own Gate 4 formula implies $72.42. The same document carries both, 35–52% apart. T1 is right. (ii) The three names that passed Gate 4 the same day were measured on a different instrument — Edwards Lifesciences [EW]/data/build_model.py:259 computes (future value / spot)^(1/6) − 1, undiscounted, which is the correct construction. BE, the highest-beta name in coverage (β 3.14 Blume / 4.20 raw), took the largest possible penalty from a convention no other persisted model in the repo uses.

8.3 What survives correction, and what does not

Survives — and this is why the ruling is UPHELD. The load-bearing claim is the flip point, and it holds. Corrected E[R] across the plausible bear-weight span: 0% bear −4.2%, 15% −8.2%, 25% (used) −11.7%, 40% −18.4%. The 4.7% hurdle sits outside that range at every weighting. Gate 4 remains unreachable on probabilities, exactly as the memo says — and the corrected numbers make the "hurdle outside the range" statement true without needing §3's "defensible corner" hedge, which was only required because the spurious Ke axis manufactured a +8.9% cell.

Does not survive. The characterisation of the margin. "Fails by 31 points," "fails by 13–34 points at every defensible parameterisation," and "−48–65 points" (research §8) should read −16.4pp, and the break-even should read 40.0x on FY2029E EPS — 25% above the memo's own 32x bull-case multiple, against a 24x base that the memo nowhere anchors (Path B condition 4 is marked "not callable — the peer set is not stable enough to adjudicate"). That is a stronger and more falsifiable rejection than the Ke grid, and it is the number that should be in the ticket.

8.4 v1.5.1 compliance — the horizon declaration is nominal, not genuine

The user asked whether the v1.5.1 labels are real applications or post-hoc. On Gate 4a: not genuine.

1. The horizon is inherited from a gate that FAILED. Gate 4a says "Gate 4's horizon is Gate 3's catalyst horizon." Gate 3 is recorded FAIL — "no variant for a catalyst to resolve." There is no catalyst horizon to inherit; the rule's antecedent is absent.
2. It contradicts the memo's own thesis statement. §2 Gate 3 says the disagreement "is about the 2029 exit multiple." A 2029 object cannot resolve in 12 months. The honest thesis horizon is ~3 years, which is what the model computes.
3. The declared horizon does not carry the verdict. §3's own header reports the 12-month figure, but §4, the OUTPUT block, the manifest and every downstream artifact report the 3-year number. The declaration is decorative — the same defect item B21 logged against C4 on ISRG/GH.
4. "Declared before E[R] was computed" cannot be literally true of the 3-year figure: it is dated 14:22 in BE_Memo_Manifest and in the v1.4.2 valuation document, three hours before the re-run.
5. The −43.9% / −29.9% pair has no workpaper and is not internally consistent. Neither number appears in BE_Model_2026-07-28.xlsx (which has a "3-yr annualised E[R]" row and no 12-month row) or in any other artifact on disk. Reverse-engineering them, −29.9% implies Ke ≈ 12.0% and −43.9% implies Ke ≈ 10.8% — two different undisclosed rates, neither the memo's own 20.4% CAPM figure.

The exculpatory finding, stated because it matters. The declaration was made in the direction that hurts the name (12m fails by more than 3y). A gamed horizon declaration looks like the opposite. This is sloppy application, not motivated reasoning — and that distinction is the whole point of asking.

The generalisable defect (logged as B25): Gate 4a specifies a period but not a convergence convention. EW's 12-month E[R] is a forward-P/E target (+11.6%); BE's is a multi-year DCF rescaled to one year. Under BE's convention a shorter declared horizon mechanically produces a worse E[R], so the dual-horizon test cannot discriminate — it can only ever confirm.

8.5 The pre-registration — did it steer the analysis?

My finding: no, and the evidence is specific. Three checks:

• The pre-registration predicted failure "by 13–34 points." The re-run produced numbers worse than predicted (−48.6pp at the declared horizon), not conveniently matching them. A steered analysis lands on its prediction.
• The corrected Gate 4 number (−16.4pp) is better for the name than the published one. If the work had been fitted to a pre-registered rejection, the error would run the other way.
• Two independent corrections were published that weakened the memo's own bear-side findings: the scandium/china absence claim was retracted, and BE was explicitly downgraded in CALIBRATION_WATCH.md's "retired disclosure" table from a strong instance to a weak one, with the honest note that "counting it there overstated the technique's hit rate." Self-correction against interest is the strongest available evidence of an unsteered process.

What the pre-registration did cost: §3.5 and §3.6 each close with a paragraph narrating compliance ("Pre-registration honoured," "No re-scope trigger fired"). That is advocacy framing in an analytical document. It changes no number, but it is the register a steered memo would be written in, and it makes the genuine independence above harder to see.

8.6 The correction is incomplete — three downstream survivals

The falsified claim (scandium/china absent from the transcripts) is corrected in research §3.3. It is not corrected in:

1. Research §4, bullet 1 — still reads "the term appears zero times across the four transcripts sampled." This is the sentence that adjudicates the Hunterbrook allegation, and it is the falsified version.
2. BE_Memo_Manifest_2026-07-28.json, disclosed_limitations[4] — still reads "'scandium' and 'china' each appear zero times across all four sampled transcripts." The manifest is the provenance artifact and it certifies the pre-correction memo (framework_version: "1.4.2", no archetype, no Path B, no Gate 4a, expected_return_annualised: -0.267, and a limitation stating the series is 4 of 9 and "this memo will be refreshed" — which is now false).
3. Research §3.2 still asserts "This is the same disclosure pattern flagged on TWST" — the exact equivalence that §3.3's second correction retracts ("it is not the TWST/CAI 'retired disclosure' pattern"), and that CALIBRATION_WATCH.md has already been corrected to reflect. The manifest's key_findings[1] repeats it. The document contradicts itself across ninety lines.

And a live lead the correction surfaced but did not read. The retrieved 2025Q1 transcript contains management answering the scandium question fifteen months before Hunterbrook: "we are not dependent on China for scandium… we get this from multiple geographies and multiple continents… We don't reveal those sources and methods. That's part of our IP." (K.R. Sridhar, Q&A). That is (a) corroborating consistency for the company across fifteen months and an unprompted occasion, which cuts mildly against the short report and is uncredited, and (b) in the same exchange, an analyst asks management to "explain why we change metrics and how we change metrics" and for an update on "the megawatts deployed" — a contemporaneous, third-party observation that BE changes disclosed metrics, which is directly on point for the memo's own Gate 2B leg-3 argument and goes unused. The count was corrected; the retrieved text was never re-read for content.

8.7 Archetype — the declaration is overstated

§0.5 asserts "Not a judgement call. BE meets every INFLECTION criterion." Against the definition in references/trade-construction.md ("high gross margin, negative or thin operating margin, operating margin expanding fast, high or accelerating growth"), BE meets two of four as written:

• Gross margin 33.4% is not high. The criterion exists to prove unit economics work and was calibrated on a reference set running 65.5%. 33.4% is a hardware manufacturer's gross margin.
• Operating margin 17.1% GAAP is not "negative or thin." The memo pivots to "it was thin-to-negative as recently as Q2'25" — a trajectory argument substituted for a level criterion.

Not verdict-changing (under COMPOUNDER, Gate 1 still passes on F-score 7/9 and accruals −11.9%, and Gate 4 is untouched). But "not a judgement call" is the wrong confidence, and the honest reading is that the taxonomy has no slot for a capital-intensive hardware manufacturer at a capacity-constrained inflection — which is exactly the generalisation-outside-healthcare test BE was chosen to run. Logged as B26.

8.8 Verified clean

Stated because confirming discipline is as valuable as finding bias. Independently re-derived and correct: Q2 results tie to EX-99.1 to the dollar; the FY2024 press-release-vs-XBRL basis footnote is right (~$4.6m/qtr); guidance arc +12.5% / +25.9% / +31.7%; H2 math ($2,083.6–2,383.6m, low end implies sequential decline, +72.3% YoY); net cash $4.06m; EV/sales 14.7x, 68x, 70x; warrant $400.0m proceeds, 1.09% dilution, $306.5m = 7.6% of FY26E; accruals −12.1% (memo −11.9%); asset growth +122.4%; the 21% tax correction is applied in the workbook ((F·G·0.79+60000)/shares); trend −37.0% vs 50-day, −5.8% vs 200-day, +3.7% at $183.90; realised vol 110.5 / 121.7 / 128.7 (IEX vs the memo's SIP 110.6 / 122.5 / 129.3); 52-week low $34.75; break-even multiples internally consistent with the grid to ±1x. The mention-frequency table reproduces exactly from .cache/be/, including the three hardest claims: capacity 26Q1 = 30 mentions, 18 in prepared remarks; backlog 26Q1 = 6, 5 prepared; scandium = zero prepared remarks in all nine quarters, Q&A only. The Alpaca IV parity finding is correct and is a real catch.

8.9 Mirror check (required)

Would this gate logic, sign-flipped, also have rejected a Short? Yes — and correctly, for a different reason. The long is blocked by Gate 2 (a literally zero variant) and Gate 4 (price). A short is blocked by the fundamentals themselves: +165.5% revenue growth, a third guidance raise, top-decile momentum, negative accruals. Those are not one unsatisfiable gate firing twice; they are two different findings. This is the "expensive but not a short" case the framework was founded to get right, and it gets it right. No mirror failure.

8.10 Three smaller errors

1. Research §8 units conflation: "values the name 31% below spot on an undiscounted expected-value basis, failing the hurdle by 16.4pp." Those are the discount-to-intrinsic-value figures (−48.4% to −65.4%), not points versus the hurdle. §3.5's "fails by 13–34 points" quotes E[R] magnitudes as if they were shortfalls (the actual gaps for those cells are 17.8–38.6pp). Same class as D2 — a well-formed, plausible, wrong number. Logged as D4.
2. "Oracle was material from 2025Q2, five months before the warrant was agreed." The Q2'25 earnings 8-K is dated 2025-07-31 (EDGAR acc 0001628280-25-036998); the warrant was agreed 2025-10-28. That is ~3 months, not five. The inference (commercial relationship predates the equity consideration) survives; the interval is overstated by 60%, in the direction that strengthens it.
3. 12-1 momentum "+386% to +800%" is not reproducible from Alpaca daily closes under any of eight standard windows, which span +380% to +1,072% (trading-day 12-1 with skip month: +1,072.4%; calendar 12-1: +627.9%; 12-month total: +380.1%). BE is emphatically top-decile under all of them, so Gate 6 is unaffected — but the figure is quoted verbatim in CALIBRATION_WATCH.md item B9 as the evidence for "the strongest factor profile the system has underwritten," and a number with a 3x spread across defensible definitions must name its window.

8.11 My own bias, stated

I am incentivised to find problems, and I found several. Weigh them accordingly: none of them changes the verdict, the verdict is right, and the single biggest correction (§8.2) runs in the name's favour — I am reporting a finding that makes the memo's rejection look less comfortable, not more. The one thing I would genuinely push back on if this were a Long is the bear case (FY2029 revenue of $4,037m, below FY2026E guidance, after three years with a $6bn product backlog, at 25% weight). It looks harsh to me. It also does not matter, because the flip point on bear weight does not exist.

8.12 What I am asking the memo to change (evidence, not instruction — the user calls it)

1. Restate Gate 4 as −11.7% annualised, −16.4pp versus the hurdle, with the break-even at 40.0x FY2029E, flat across Ke, and delete or relabel the §3 Ke grid, which has one degree of freedom.
2. Correct research §4 bullet 1 and manifest disclosed_limitations[4] (scandium), and research §3.2 plus manifest key_findings[1] (the TWST equivalence, already retracted in §3.3 and in CALIBRATION_WATCH.md).
3. Reissue the manifest at framework_version: "1.5.1" with the archetype, Path B, Gate 4a and the corrected E[R]. As it stands it certifies a memo that no longer exists.
4. Restate item B9 in CALIBRATION_WATCH.md at −16.4pp. Its argument — that a top-decile short-horizon signal set is being correctly overridden by a multi-year valuation gate — survives the correction and is if anything more interesting at a 16pp gap than at 31pp, because the trade-off is now a real one rather than a rout.
5. Soften §0.5 from "meets every INFLECTION criterion" to the two it meets, and record the taxonomy gap.

Logged to CALIBRATION_WATCH.md: B24 (Gate 4 double-discount), B25 (Gate 4a horizon convention undefined), B26 (archetype taxonomy gap), D4 (three further D1-class instances). Nothing here is a gate loosening; B24 is a correction that makes a gate less strict by 15pp/yr and changes no verdict in the record. Per the amendment quarantine, B24's restated arithmetic may not first be applied to BE as grounds for a position until BE's next full re-underwrite — it is published here as a correction to the stated margin, not as a route to a Long.