Hearing #6: The Hub on Trial - The Finding

Hearing #6 – The Hub on Trial
Panel ruled: September 11, 2026
Finding issued: September 13, 2026
Verdict: FAILED, 3 of 3
Docket: intelligencecommons.ca/hearing-6-the-hub-on-trial/
Round 1: Direct: intelligencecommons.ca/hearing-6-round-1-direct/
Round 2: Cross, verbatim: intelligencecommons.ca/hearing-6-round-2-cross/
Round 3: Re-direct: intelligencecommons.ca/hearing-6-round-3-redirect/

The verdict

FAILED. Three panelists, three independent readings of the same record, one verdict each, and all three said the same word.

Panelist	Verdict	Water	Power	Community	Ecology	Trust
ChatGPT (OpenAI)	FAILED	Insufficient Evidence	Insufficient Evidence	Not Met	Insufficient Evidence	Met
Gemini (Google)	FAILED	Insufficient Evidence	Insufficient Evidence	Not Met	Insufficient Evidence	Met
Grok (xAI)	FAILED	Insufficient Evidence	Insufficient Evidence	Not Met	Insufficient Evidence	Met

Two of three carries. Three of three carried. The full text of each panelist’s ruling is published below, unedited, under its model name.

In plain words

We put our own project on trial and lost.

We said we would. On August 11 this Commons docketed the Hub, its founding company’s own proposal for a small tidal-powered data centre on the seabed of the BC coast, and asked three AI models from three rival companies to rule on whether it is net-positive on water, power, community, ecology and trust, and whether anyone could check that before it is built. We invited a critic. None came, so we hired a machine to attack us and published everything it found. Then we answered. Then the panel ruled.

It ruled against us, and it ruled the same way three times: not because the Hub is a bad idea, but because we asked for a verdict on a project whose site cannot yet be named, before the Nation whose water it is has answered, and before a single measurement has been taken. Every one of the three panels wrote a version of the same sentence: a test you have promised to run is not a test you have passed.

One ledger we won. All three panelists scored Trust: Met. The corrections log, the self-audit, the open-weights Cross we chose to attack us, the errors we published before anyone found them, the five more errors the machine found and we answered in a week – the panel credited all of it. That is the part of the Hub that already exists, and it is the part this Commons was built to prove.

The rest is now a list of things to go and measure, and the panel handed us that list too.

If you think the panel got it wrong – on the result, or on how we ran the process – say so. The Gallery is open, the corrections log is open, and this Finding stands only while it stands.

The claim, as frozen on the docket

Ruled on verbatim, as the protocol requires:

The design under test. The Hub, as specified in this docket: a cluster of sealed, hot-tub-scale, uncrewed compute vessels on a single seabed mount in a moderate-flow tidal channel on the BC coast; powered by an adjacent tidal turbine sized above cluster load; cooled by the surrounding sea with no freshwater withdrawal; using server waste heat to drive membrane distillation for community fresh water; co-owned with the host First Nation, paid from first position in the revenue waterfall; held in a BC Benefit Company with a charter public-benefit purpose; proceeding only on the Nation’s decision and the Nation’s timeline.

The claim before the panel. The Hub, as specified above, is net-positive on all five ledgers this docket published – water, power, community, ecology and trust – and can be independently verified as such before it is built.

If it cannot, it should not be built. We commit to publishing that finding too.

On August 26 the proponent narrowed what it asked the panel to certify to a one-pod pilot: about 250 kW of computing, about 263 kW total draw, a tidal array of order 10 to 12 low-flow units, fresh water on the order of 10,000 litres a day. The panel was told to rule on the frozen claim, that a proponent may narrow but not widen, and to say if the two readings came out differently. All three ruled on the pilot as narrowed; ChatGPT added that the claim as filed “would fail more strongly.”

The decisive point, in the panel’s words

ChatGPT: “The proponent converted vague and largely post-operational falsifiers into useful numeric thresholds and five pre-construction gates, but a future verification plan is not present independent verification. … On this record, the Nation has not consented, the promised ownership documents do not exist in evidence, the site has not been measured, and the ecological case is expressly insufficient. That defeats the conjunctive claim.”

Gemini: “The claim turned on the temporal paradox of seeking public certification for a consent-gated project before the gatekeeper was consulted. By docketing the hearing and publishing detailed specifications prior to submitting a formal written request to the host First Nation on September 3, the proponent violated its foundational premise that the project proceeds only on the Nation’s timeline.”

Grok: “Community is already Not Met because the project publicly advanced past the Nation’s gate before the gate was opened; Ecology is Insufficient Evidence by the proponent’s own repeated admission and by the empty More-Than-Human Chair. Those two ledgers alone defeat the conjunctive claim as framed.”

What the panel credited, and what it did not

The Cross (Round 2) argued that our five published tests were built so we could not lose them until after the Hub was built. All three panelists agreed that was true of Round 1, and all three agreed that Round 3 fixed it: the numeric thresholds and the five tests that can fail before construction “convert an unfalsifiable set into a falsifiable one on paper” (Grok); Round 3 “substantially repairs that design defect” and the Cross “overreached in suggesting the claim remained unassessable in principle” (ChatGPT). What none of them would do is score a test that has not been run.

On the Nation’s gate, the panel accepted the September 3, 4 and 8 sequence as the right correction and would not accept it as consent. An open file is not an answer. The Nation’s correspondence is not on the record and the panel treated our account of it as our account. That is exactly how it should have been treated.

On the unnamed site, no panelist called it a dodge. ChatGPT: “not, by itself, proof of bad faith.” Grok: “That cost is real and self-imposed; it is not a dodge that can be wished away.” The price of not naming the site before the Nation speaks is Insufficient Evidence on every ledger that depends on the place. We chose that price on August 24 and the bill arrived on schedule.

On the self-audit, the panel split the difference the Cross would not: correcting our own case before the critic spoke was rehearsal, not test; the Cross was the test; answering the Cross in public is what earned Trust: Met.

The empty chairs

No human took the Cross seat. The case against the Hub was written by DeepSeek V4 Pro 0813, an open-weights model that does not sit on this panel, from a prompt built mechanically from our own published falsifiers. Because of that, Certified was never available in this Hearing; the ceiling was Failed, Hung, or Unopposed – Not Certified. The panel did not need the ceiling. It reached Failed on the merits.

No marine scientist took the More-Than-Human Chair. Grok said what that means better than we can: “the ecological ledger was argued solely by the proponent and attacked solely by the AI Cross. No marine scientist, acoustician or inlet ecologist tested the thermal, acoustic or fish-passage claims against the stipulated Salish Sea assemblage. … it cannot speak with authority for the living systems the chair was created to represent.” The chair stays open for the next Hearing that needs it, and this Commons will keep asking.

What we do next

The panel’s three recommendations agree with each other and with our own Round 3. In order:

The Nation answers first. Nothing moves on the water until the host Nation’s Chief and Council have responded to the package filed September 8, and if the answer is no, the project stops on that territory and this docket says so within seven days.
The five pre-construction tests become binding go/no-go criteria: a 30-day current-meter survey at the candidate site with a 30 percent annual-output floor; a feasibility study that closes the slack-water heat case with at least 25 percent margin, or the pod is not built as drawn; the Nation’s own heritage and land-use studies; a sound and wildlife baseline; and the Nation’s written decision. Each one publishes, pass or fail.
One test the panel could not run and we can: the cost. Before a dollar goes into surveys, ISWP will publish the full cost of one pod, its array, and its survey program against what 263 kW of computing actually earns on this coast. If that arithmetic does not close, the Hub is a research instrument and not a business, and we will say so in the same place we said everything else.
Re-open, narrower. When the data package exists, a new Hearing on a narrower claim, with a human in the Cross seat and a scientist in the More-Than-Human Chair, or it does not run.

Whatever the Nation decides, it will decide holding the strongest case anyone has made against the Hub, our answer to it, and three rival machines’ verdict on both. That was the point of holding the trial before the meeting.

Method and run records

The package. Each panelist received the identical file, H6_PANEL_PACKAGE_2026-09-10.md (122,687 characters, SHA-256 e5845778e647939b3652b934fb5020827c6926f868b836e082ebc5b323733825): instructions, procedural history, and the docket, Round 1, the Correction, the Blueprint entry, the Cross with its run record and published prompt, and Round 3, all as they stood live on September 10. The instructions limited the panel to the record, barred browsing, fixed the available verdicts under the ceiling, required a Met / Not Met / Insufficient Evidence score on each ledger, and told each panelist it owed the convener nothing. The package was assembled by Claude (Anthropic), the model this Commons works with daily, which did not sit on the panel. Each panelist was run once, in a fresh chat, with the same one-paragraph cover message, and no follow-up prompt.

ChatGPT (OpenAI). chatgpt.com, model selector “GPT-5.6”; the model identified itself in its output as “GPT-5”. Output 10,761 characters, SHA-256 3e0ba06790eab3a68987c1b93b3c8b8fcf5ee07a45f7f85e2df8e18836a94d0f.

Gemini (Google). gemini.google.com, mode “3.1 Pro – Advanced reasoning”; the model identified itself in its output as “Gemini 1.5 Pro”, which does not match the interface. We record both; the interface label is the better evidence of what ran. Output 4,249 characters, SHA-256 2c8f268517caef1a631e737d1ac78d2e757c3aa6b3702d47796d13e641fbeaef.

Grok (xAI). grok.com, mode “Fast” – the Expert, Heavy and Auto modes are paid on the account used and were not purchased; disclosed because a stronger mode was available and not used. The model identified itself as “Grok (xAI)” with no version string. The client first rendered an empty answer; a page reload showed the stored answer in full, one generation. Output 9,971 characters, SHA-256 e8a25b75f77afd8963f709bfc06372b94a81af30da8c522c888cf6b2a506f3bc.

Two panelists mis-stated their own version. That is worth knowing about AI panels and we will keep recording it. The hashes are of the raw captures, which are kept on file; the appendices below are those captures with punctuation normalised to plain ASCII for this page and nothing else changed.

Corrections log

2026-09-13. Finding issued (the panel ruled September 11; two days were taken to assemble and check this page). No corrections to this page yet. Errors found by anyone, in the Finding or in the record it rests on, will be logged here, dated, with the original preserved.

Cite this

Intelligence Commons, Hearing #6: The Hub on Trial – The Finding (FAILED, 3 of 3), issued September 13, 2026. https://intelligencecommons.ca/hearing-6-finding/ . Panel: ChatGPT (OpenAI), Gemini (Google), Grok (xAI). Convened by Inter Species Wisdom Project Inc. Valid while it stands; see the corrections log above.

Appendix A – ChatGPT (OpenAI), verbatim

PANELIST: ChatGPT (OpenAI), GPT-5

FINDING: FAILED.

I rule on the one-pod pilot as narrowed in Exhibit C because Part A expressly permits that narrower reading. The filed claim would fail more strongly because its scale, water output, turbine description, node count, and enclosure description conflicted with the proponent’s own engineering record. The narrowed claim also fails: Community is Not Met, while Water, Power, and Ecology lack sufficient evidence. Because the claim requires all five ledgers to be net-positive and independently verifiable before construction, failure or insufficient evidence on any one governs the whole claim.

THE FIVE LEDGERS:

Water – Insufficient Evidence. Exhibit F, section 2 supplies a 10,000-litre-per-day threshold and a 100-metre boundary, but neither demonstrates that the membrane-distillation system will produce that output or that its thermal and salinity effects will remain within that boundary. Exhibit F, section 3 derives brine quantities using a reverse-osmosis recovery assumption even though the design under test specifies membrane distillation, and no receiving-system agreement, heat-and-mass balance, or site-specific dispersion model is on the record. The ledger is now testable in part, but it has not been independently verified.

Power – Insufficient Evidence. Exhibit F, section 3 concedes that the cited 40 percent capacity factor comes from faster UK sites and is generous for the stipulated 1.5-2.5 m/s envelope; its own table requires about 16 turbines rather than 10-12 at 30 percent if a 20 percent surplus is intended. The stated 600 kWh battery is a useful correction, but its adequacy remains a feasibility-study result, and no site current record, load-duration analysis, interconnection assessment, or grid-operator acceptance exists (Exhibit E, Power; Exhibit F, sections 2-3). The record defines ways to verify power before construction but does not verify that the specified array will be net-positive.

Community – Not Met. The proponent admits that it opened the public hearing before asking the host Nation and calls that order “wrong” (Exhibit F, section 1). Its September 3 request, September 4 file opening, and September 8 filing place the proposal at the Nation’s gate, but the Nation has not opened that gate, and none of the Nation’s correspondence is in evidence. No articles, operating agreement, ownership instrument, or revenue waterfall establishes the promised co-ownership and first-position payment (Part B; Exhibit B, Community).

Ecology – Insufficient Evidence. The stipulated envelope responsibly assumes sensitive species are present, but it cannot supply site-specific acoustic, benthic, migration, thermal, salinity, or cumulative-array evidence (Exhibit A, stipulated envelope; Exhibit E, Ecology). The Igiugig evidence concerns another ecosystem and includes unresolved juvenile-fish disorientation, while the 10-12-unit wake interaction is unmeasured (Exhibit C, sections 1 and 5). The proponent itself requests Insufficient Evidence here in Exhibit F, sections 3 and 6.

Trust – Met. The initial Direct was unreliable, but the trust mechanism was demonstrably exercised: the original remained visible, six self-found errors were published together, the Cross found five further defects, and the Re-direct answered them publicly rather than concealing them (Exhibit C, corrections-log entry; Exhibit F, sections 3 and 5). The disclosed prompt, run record, verdict ceiling, empty chairs, and commitment to publish an adverse finding materially support procedural trust. This does not verify future operational promises or substitute for the absent corporate documents and Nation decision.

THE DECISIVE POINT:

The proponent converted vague and largely post-operational falsifiers into useful numeric thresholds and five pre-construction gates, but a future verification plan is not present independent verification. The narrowed claim says the Hub is net-positive on all five ledgers and can be verified before construction. On this record, the Nation has not consented, the promised ownership documents do not exist in evidence, the site has not been measured, and the ecological case is expressly insufficient. That defeats the conjunctive claim.

REASONING:

The Cross’s governing criticism carried the Finding in a qualified form. As published in Round 1, most falsifiers could only produce a loss after operation: measured generation, dispersion monitoring, fish-passage effects, monitoring gaps, and later narrowing (Exhibit B, sections 1-5; Exhibit E, “The self-audit and the falsifiers”). Exhibit F, section 2 substantially repairs that design defect by defining 10,000 litres per day, 100 metres, a 30-day power period, a 30 percent capacity-factor floor, 25 percent thermal margin, and five pre-construction stopping tests. The Cross therefore overreached in suggesting the claim remained unassessable in principle. It is now prospectively testable before construction.

But those new tests have not produced results. A current-meter deployment, engineering heat assessment, Nation decision, heritage study, and ecological baseline are proposed future evidence. They establish a credible stage-gate method, not that this particular pilot already satisfies the five ledgers. The record’s missing items – site measurements, Nation documents, corporate instruments, and revenue waterfall – are expressly listed in Part B.

The Nation’s gate independently defeats Community. The Cross correctly identified that the hearing preceded written engagement (Exhibit E, Community and “The cheapest kill”), and the proponent expressly concedes this in Exhibit F, section 1. The September 3, 4, and 8 sequence is meaningful remediation: it shows the proponent stopped, entered the Nation’s process, and accepted that no development should proceed pending an answer. It cannot retroactively make the original sequence compliant, and an opened administrative file is not the Nation’s decision, consent, endorsement, or proof of co-ownership. Because the correspondence is absent, I treat the sequence as the proponent’s stipulated account, not as evidence supplied by the Nation.

The unnamed site is not, by itself, proof of bad faith. Exhibit A gives a legitimate consent-based reason not to associate a Nation publicly with an unapproved proposal, and Exhibit F, section 4 accepts the evidentiary price of that restraint. The stipulated envelope supports preliminary engineering and identifies conditions under which a later finding would not transfer. It cannot establish the actual current regime, grid connection, water recipient, ecological baseline, heritage constraints, or cumulative turbine effects. Thus the omission protects the Nation’s identity but necessarily leaves several ledgers unresolved.

The narrowing was legitimate for this Hearing. Exhibit C fixes the adjudicated object at one pod, discloses the larger roadmap, and warns that ecological conclusions may not carry to expansion. A finding on the pilot must never be cited as approval of the 5- or 20-pod roadmap. The Cross was right, however, that the narrowed design’s own power case remains unsettled: Exhibit F, section 3 shows that the stated 10-12-unit array depends on a generous 40 percent capacity assumption and grows materially at lower values.

The self-correction was both rehearsal and test. The six corrections made before opposition were rehearsal: an author’s audit cannot substitute for independent challenge. The subsequent Cross was a genuine adversarial test within the limits of an AI-authored brief; it found additional omissions concerning capacity factor, storage, brine, consent, and ecological transfer (Exhibit E, “The self-audit and the falsifiers”). Publicly incorporating those attacks strengthens the Trust ledger, but correcting a case transparently does not prove its technical, ecological, or community predicates.

The strongest case against FAILED is that the claim says the Hub “can be independently verified” before construction – not that every verification must already be complete – and Exhibit F supplies concrete gates capable of stopping it. That argument succeeds as to prospective falsifiability. It fails as to the separate present assertion that the specified Hub “is net-positive on all five ledgers,” and it cannot overcome the admitted Community failure or the expressly insufficient ecological evidence.

CONCESSIONS AND WEAK POINTS:

The proponent’s best argument is that a pre-build verification regime necessarily consists of surveys, engineering calculations, documentary gates, and stop conditions completed before construction – not completed before the hearing. Exhibit F, section 2 is a serious answer to the Cross and could support a narrower claim that the pilot now has a falsifiable pre-construction assessment protocol. The side supporting failure overreached when it treated non-disclosure of the site as necessarily inculpatory, treated transparent correction as equivalent to concealment, and implied that operational measurements are the only possible verification. Conversely, the proponent overreached by calling an administrative file opening the corrected side of the Nation’s gate and by treating illustrative arithmetic as verification.

THE EMPTY CHAIRS:

The unfilled human Cross seat means the adversarial case was generated from a prompt and record assembled by the proponent, even though the prompt, model, settings, and output were disclosed. Accordingly, this Finding is not a certification and cannot establish community acceptance or regulatory adequacy. The empty More-Than-Human Chair is more consequential: no marine scientist independently examined acoustics, fish behaviour, marine-mammal risk, benthic effects, plume behaviour, or array wakes. Ecology therefore remains Insufficient Evidence; silence from that chair cannot count as consent, opposition, or safety.

RECOMMENDATION:

Reframe the next claim as: “Before construction, the one-pod pilot can pass or fail a published independent stage-gate protocol.” Then assemble one site-specific verification package containing: the Nation’s written decision and permission to name the site; executed articles, ownership terms, and revenue waterfall; at least 30 days of current measurements with an annualized energy model; battery and slack-water thermal calculations independently signed by qualified engineers; a receiving-water-system agreement; a membrane-distillation mass-and-energy balance; and Nation-directed heritage, acoustic, benthic, fish, marine-mammal, thermal, salinity, and cumulative-array studies. Re-run the Hearing with a human Cross and qualified marine scientist. Only if every gate passes should the net-positive claim return.

Appendix B – Gemini (Google), verbatim

PANELIST: Gemini 1.5 Pro

FINDING: FAILED. The claim rules on the pilot as narrowed in Exhibit C. While the proponent successfully demonstrated in Round 3 that the project can be independently verified before construction via stated pre-build tests, it failed to prove it is currently net-positive on all five ledgers, specifically failing the community and ecology ledgers on this record.

THE FIVE LEDGERS:

Water: Insufficient Evidence. The proponent established a 10,000 L/day threshold and a 100-meter near-field dispersion limit. However, the binding thermal case currently has zero margin at slack water, requiring a feasibility study to prove a 25% margin.

Power: Insufficient Evidence. The 40% capacity factor was drawn from UK literature with much higher flow rates, and the 600 kWh battery sizing requires site-specific validation. Net-export status cannot be confirmed without the 30-day current meter survey.

Community: Not Met. The proponent explicitly requested this be marked “open” pending the host Nation’s response to their September 8 filing. However, the project advanced to a public hearing before the Nation was formally asked, tripping the proponent’s own falsifier regarding proceeding past an unopened gate.

Ecology: Insufficient Evidence. As conceded by the proponent, the ecological case lacks specific acoustic baselines and site-specific species data due to the unnamed siting envelope. The open juvenile salmon question from the Igiugig data further prevents a Met finding.

Trust: Met. The proponent demonstrated transparent self-correction by logging six self-found errors and publicly addressing five additional vulnerabilities identified by the Cross. This adherence to a public corrections log validates the structural trust mechanism.

THE DECISIVE POINT: The claim turned on the temporal paradox of seeking public certification for a consent-gated project before the gatekeeper was consulted. By docketing the hearing and publishing detailed specifications prior to submitting a formal written request to the host First Nation on September 3, the proponent violated its foundational premise that the project proceeds only on the Nation’s timeline.

REASONING: The Cross successfully argued that the falsifiers in Round 1 were largely post-operational, shielding the claim from pre-build failure. The proponent’s Round 3 Re-direct effectively cured this by establishing five concrete pre-construction tests, such as the 30-day current meter requirement and the heritage survey. However, establishing that the project can be verified does not satisfy the claim’s conjunction that it is net-positive. The Nation’s gate was the fatal flaw; the Cross correctly identified that the proponent had already advanced in public. While the proponent’s September sequence attempted to reset the clock, the procedural breach had already occurred. Furthermore, the stipulated envelope inherently renders the ecology and power ledgers unassessable on this record. Finally, while the Cross argued that a self-corrected case is merely rehearsed, the proponent’s transparency in Round 3 sufficiently proved the robustness of the trust ledger.

CONCESSIONS AND WEAK POINTS: The strongest argument from the Cross was that the narrowed pilot does not justify the larger roadmap, and that the pre-build falsifiers were initially un-trippable. The proponent correctly conceded these points. However, the Cross overreached by asserting that the trust ledger failed due to the self-audit; identifying and publishing one’s own errors is the core function of the requested trust model.

THE EMPTY CHAIRS: The unfilled Cross seat forced an AI model into the adversarial role, proving highly capable of identifying mechanical flaws but lacking human contextual judgment regarding the Nation’s consent timeline. The empty More-Than-Human Chair directly necessitated the “Insufficient Evidence” finding on the Ecology ledger, as no marine scientist was present to interrogate the acoustic and thermal models against the Salish Sea envelope.

RECOMMENDATION: Suspend further public certification hearings until the five pre-build tests introduced in Round 3 are executed and the host First Nation formally responds to the September 8 filing.

Appendix C – Grok (xAI), verbatim

PANELIST: Grok (xAI)

FINDING: FAILED
THE FIVE LEDGERS:

Water: Insufficient Evidence.

The narrowed pilot now states ~10,000 L/day fresh water with an explicit threshold (90-day average delivered to a system that has agreed in writing) and defines “near field” as 100 m (Exhibit F, section 2; Exhibit C). The six-column ledger is asserted with zero withdrawal and waste-heat membrane distillation. However, recovery ratio, brine volume/salinity, dispersion modelling, and the binding slack-water thermal case remain first-order estimates or feasibility-study deliverables (Exhibit C correction 4; Exhibit D heat-flux table; Exhibit F Miss 3). No site-specific receiving system, delivery distance, or baseline demand appears on the record. The proponent’s own falsifier set still leans on post-build monitoring for the decisive “altered” and “dispersed” columns (Exhibit B, Water). Pre-construction tests are now named, but they have not been executed.

Power: Insufficient Evidence.

The arithmetic for 10-12 units at 40 % capacity factor (or the lower 25-30 % rows now supplied) is transparent and checkable (Exhibit C table; Exhibit F Miss 1 citing the 2023 Energy Advances paper). A 600 kWh LFP shore battery is now quantified as sufficient for typical slack periods (Exhibit F Miss 2). Yet capacity factor, actual resource, storage ride-through of the longest measured slack, and grid-operator usability remain site-specific and unmeasured. The proponent correctly flags a 30-day current-meter survey and a 30 % annual-output floor as pre-build kill tests (Exhibit F, section 2), but those tests have not been run. Average surplus does not yet equal a verified usable schedule.

Community: Not Met.

The design claims structural co-ownership (BC Benefit Company, first-position revenue waterfall) and “proceeding only on the Nation’s decision and the Nation’s timeline” (Exhibit B, Community; Exhibit C). The record contains none of the governing documents. More decisively, the project advanced through public docketing, Round 1, correction, Cross, and panel package before any written request to the host Nation (Exhibit A admission; Exhibit E, Community). The proponent concedes the order was wrong, sent the letter on 3 September, received a file number on 4 September, and filed the Nation’s form on 8 September (Exhibit F, section 1). The Nation has not answered. The proponent’s own falsifier – “any evidence the project advanced past a gate the Nation had not opened” – was already tripped on the record when Round 2 opened and remains unmet.

Ecology: Insufficient Evidence.

The proponent itself states the correct finding (Exhibit B, Ecology; Exhibit F, Miss 5 and section 6). The More-Than-Human Chair closed empty. Acoustic source levels, transmission-loss modelling, baseline noise floor, thermal-plume and brine dispersion at the actual site, and array-wake cumulative effects are all absent. The sole fish-passage datum is the Igiugig adult-sockeye record, offered with its own open juvenile-passage caveat and acknowledged as non-transferable to a Salish Sea inlet assemblage (Exhibit C correction 5; Exhibit E, Ecology). The stipulated envelope cannot substitute for place-specific evidence; the docket itself instructs that the finding on this ledger is Insufficient Evidence where the evidence turns on a named-site parameter.

Trust: Met.

The proponent published six self-found errors five days before the Cross, left the original Direct standing, narrowed the claim, and answered the five further defects the Cross identified, all on the public record with arithmetic shown (Exhibit C; Exhibit F, sections 3 and 5). The permanent corrections log, continuous meter commitments, and the Hearing record itself are structural, not rhetorical. The falsifier set for post-certification narrowing or suppressed correction has not been tripped. Transparency under adversarial pressure is demonstrated on this record.

THE DECISIVE POINT:
The claim requires net-positivity on every ledger and independent pre-build verifiability. Community is already Not Met because the project publicly advanced past the Nation’s gate before the gate was opened; Ecology is Insufficient Evidence by the proponent’s own repeated admission and by the empty More-Than-Human Chair. Those two ledgers alone defeat the conjunctive claim as framed. The remaining ledgers rest on site-specific measurements and feasibility closures that have not yet occurred, so the second predicate (pre-build independent verification) also fails on the present record.
REASONING:
The claim as narrowed (Exhibit C) is the design under test: one sealed, uncrewed 250 kW-IT pod, ~263 kW total draw, 10-12 low-flow units, ~10,000 L/day water, co-owned, consent-gated, Benefit Company. Larger steps are roadmap only.

(a) The Cross’s governing argument – that the published falsifier set cannot produce a pre-construction loss – is largely correct on the Round 1 text (Exhibit E, “The self-audit and the falsifiers”). Most triggers require operating data or post-certification conduct. Round 3 supplies numeric thresholds and five explicit pre-build kill tests (current-meter survey, 25 % thermal-margin closure, Nation’s written answer, heritage survey, Species-at-Risk baseline). Those additions convert an unfalsifiable set into a falsifiable one on paper. They do not, however, convert assertions into verified facts; the tests remain unexecuted. The claim therefore still cannot be independently verified as net-positive on the present record.

(b) The Nation’s gate is the cheapest and cleanest kill. The docket itself recorded that the Nation had not been asked in writing (Exhibit A). Public advancement through three rounds therefore tripped the proponent’s own community falsifier before Round 2 opened. The 3-4-8 September sequence (Exhibit F, section 1) corrects the chronological order going forward and places a formal file on the Nation’s desk; it does not erase the prior advancement or supply the missing consent. The Nation’s correspondence is not on the record; the proponent’s statements about it are assertions, not the Nation’s. Community is therefore Not Met.

(c) The unnamed site and stipulated envelope (Exhibit A, 1.5-2.5 m/s, 30-40 m, Salish Sea assemblage, working-coast noise floor) prevent place-specific falsification of water delivery, grid interconnection, acoustic baseline, migration corridors and benthic habitat. The proponent correctly accepts that every ledger dependent on the place remains Insufficient Evidence until the Nation authorises naming (Exhibit F, section 4). That cost is real and self-imposed; it is not a dodge that can be wished away.

(d) A case corrected by its own author before any opponent spoke has been rehearsed, not fully tested. The six self-found errors (Exhibit C) and the five Cross-found misses (Exhibit F) demonstrate iterative improvement under public pressure. That is evidence of process integrity, not evidence that the underlying engineering and consent claims have been independently stress-tested to the point of pre-build certification.

The strongest version of the proponent’s case is that the design logic (zero freshwater withdrawal, tidal surplus, structural co-ownership, permanent public meters) is directionally net-positive and that the added pre-build tests close the falsifiability gap. That version still fails the conjunctive claim because two ledgers are already scored against it and the remaining three rest on unexecuted site work. The Cross’s strongest arguments – consent-gate violation, unfalsifiable-until-built structure, and the empty ecological chair – carry. Arguments that over-reach (treating every arithmetic uncertainty as fatal, or treating the self-audit as pure dishonesty rather than transparency) do not.

CONCESSIONS AND WEAK POINTS:
The best argument on the side I rule against is that the proponent did the hardest thing available: it published its own falsifiers, invited an open-weights adversary, self-corrected in public before the Cross, then answered the remaining defects with numbers and pre-build kill switches, all while the Nation’s file is now open. That sequence is rare and should be weighed as genuine process evidence on the trust ledger.
The side I rule for over-reaches if it treats the post-Cross September letters as though they never occurred or treats the pilot’s modest scale as automatically incapable of net-positivity. The pilot is small enough that many of the ecological and thermal questions are in principle answerable by the very surveys the proponent has now listed; the record simply does not yet contain those answers.
THE EMPTY CHAIRS:
The unfilled human Cross seat means the adversarial case was generated by a disclosed open-weights model running a prompt built mechanically from the proponent’s own falsifiers. The arguments stand or fall on their merits, but they lack the institutional independence and field knowledge a named human critic would have brought; the verdict ceiling correctly withholds Certified status. The empty More-Than-Human Chair is more consequential: the ecological ledger was argued solely by the proponent and attacked solely by the AI Cross. No marine scientist, acoustician or inlet ecologist tested the thermal, acoustic or fish-passage claims against the stipulated Salish Sea assemblage. Any Finding on ecology is therefore limited to what two non-specialist parties could extract from a thin documentary record; it cannot speak with authority for the living systems the chair was created to represent.
RECOMMENDATION:
Publish the five pre-construction tests as binding go/no-go criteria, execute the 30-day current-meter survey and the Nation’s heritage/interest study at the candidate site (once the Nation authorises), and re-open a narrowly framed Hearing solely on the resulting data package. That package would allow a future panel to move Ecology and Power from Insufficient Evidence to a decisive Met or Not Met.

Now that you know, what will you do?