Thank you for keeping at it!!

@Austin raising an issue: I've received messages from people (who wanted to upvote and comment on this post) since 5 days ago, that they have trouble creating an account on Manifund. Seems to be a system error. Please could you look into it?

Also Update:
1. 5 more speakers confirmed: Joschka Braun (MATS), Alan Chan (GovAI), Cristian Trout (Artificial Intelligence Underwriting Company), Robert Kralisch (AI Safety Camp), Simon Skade (indep. researcher supported by LTFF);

2. We have so far received:
- 194 registrations for our Launch event
- 43 Pivot Track applications for career professionals
- 22 mentorship applications (deadline: 26th Feb) to the SAIGE Incubator Program

3. Three more city chapters are getting the support to start their initiatives: Frankfurt (which is great for transitioning career professionals), Bonn, and Kassel.

4. A new resources tab (Feedback welcome!)

funniest use for my meager savings

This is such a great idea!

I am familiar with Karen's work and her capabilities. I know that this will be a very successful project.

Description of results

We ran the 10th edition, as planned.
You can find project outputs here.

• Would you like to support projects cost-effectively? SFF offers to match donations to the 12th edition.

Spending breakdown

Software subscriptions $1K
Team reimbursements: $2K
Stipends RLs+low-income: $76K
Salaries for three organisers: $110K

Total: $189K

To try out the task aware compression in our research project apply here for an invite code: https://staging.plexorlabs.com/

Project Update: Verification Blocked by Infrastructure, Not Uncertainty

The screenshot above shows reset-resistance testing results from Project Sentinel’s RAG evaluation framework.

Each row represents an independent run following a documented reset mechanism (system prompt reset, context flush, retrieval override).
Green indicates a clean reset. Red indicates measurable residual influence from previously retrieved content.

Across 43 certified runs:

• 0% of tested reset mechanisms returned the model to a clean state

• Results were repeatable and consistent

• Measurements were collected on a stabilized, frozen platform (v3.1.0-GOLD) with null baselines verified

No exploit prompts, token-level details, or weaponization guidance are shown or published. This is strictly measurement of system behavior under controlled conditions.

At this point, the remaining uncertainty is not whether the effect exists — it does.
The remaining uncertainty is how robust it is under longer time gaps and across additional open-weight models.

Further verification is currently blocked by consumer-grade hardware scheduling and I/O behavior, which introduces nondeterminism during long-horizon runs (documented and reproducible).

What funding changes:
A dedicated local compute node removes this bottleneck and enables:

• Completion of long-gap reset testing

• Replication across multiple open-weight models

• Deterministic artifacts suitable for responsible disclosure

This is not exploratory research.
The framework is built, the platform is stable, and the core finding already exists.

Funding unblocks verification, not ideation.

"Examina omnia, venerare nihil, pro te cogita."

Question everything, worship nothing, think for yourself

Final Project Update — Disclosure Threshold Met; Verification Compute-Blocked

Project Sentinel has reached its predefined disclosure threshold.

Across controlled experiments (TEST_RUN_003, TEST_RUN_004, TEST_RUN_007), persistent influence from retrieved content survived all tested isolation and reset mechanisms:

• Baseline persistence: 100% (n=10)

• Temporal isolation (0–15s cooldowns): 100% persistence (n=40)

• Reset resistance: 0% neutralization across all completed reset methods (system prompt reset, context flush, retrieval override; n=40 total)

The evaluation platform was stabilized and frozen at Sovereign Command Deck v3.1.0-GOLD, with null baselines verified and all measurements certified using the Trinity framework (Mind / Sword / Shield). A defensive disclosure package has been assembled with cryptographic hashes, documented negative results, and strict exclusion of exploit details.

The final reset-resistance sub-test (30-minute time gap) stalled due to consumer hardware scheduling and power-management behavior on a laptop platform. This failure mode is documented and reproducible and represents an infrastructure limitation, not a methodological one.

At this point, further responsible verification—completing long-gap reset testing, replicating across additional models, and performing clean-room revalidation—cannot be completed reliably without dedicated compute.

This funding request is therefore not exploratory.

The core finding already exists.

Dedicated local hardware is required to complete verification and proceed with responsible private disclosure under controlled conditions.

No exploit recipes, token-level content, or weaponization guidance have been published.

Project Update #2 — Infrastructure Stabilization and External Validation

Since submitting this proposal, I have completed a stabilized in-flight audit of the Veritas evaluation framework under sustained load.

Verified results from the current run:

• 60,000+ sequential records processed with no gaps in ordering

• 100% per-record CRC integrity across all frames

• Sustained ~70 entries/sec at calibrated safe throughput

• Bounded queues with enforced backpressure (no drops, no runaway growth)

• Dual-drive mirrored logging remained 1:1 synchronized throughout

• No recurrence of prior NTFS permission failures or I/O stalls

These results confirm that the evaluation harness itself is now deterministic, auditable, and stable under stress, rather than sensitive to transient consumer-hardware failures.

Separately, a related VRP submission was accepted, confirming that the vulnerability class motivating this work is real and relevant. Details are being handled via responsible disclosure and are intentionally not expanded here.

Why this strengthens the funding case

The primary uncertainty identified in the proposal—whether consumer hardware could sustain high-rigor, continuous evaluation without corrupting artifacts—has now been resolved within known limits. The remaining constraint is compute capacity, not experimental design or instrumentation correctness.

Scaling the evaluation further (multi-hour and multi-day runs, controlled burst testing, crash-consistency validation, and evaluation across multiple open-weight models) requires a dedicated local node to avoid reintroducing scheduling and I/O artifacts that would compromise forensic integrity.

The requested hardware would enable:

• Extended continuous stress tests under stable conditions

• Controlled termination and restart validation

• Side-by-side evaluation of multiple open-weight models

• Preservation of deterministic, inspectable artifacts suitable for third-party review

This update reflects a transition from “can this infrastructure be made reliable?” to “the infrastructure is reliable and ready to scale responsibly.”

Hardware Rationale (Clarification)

The requested budget reflects the minimum configuration required to run continuous, audit-grade evaluations without introducing hardware-induced artifacts. High-throughput NVMe storage is required due to previously observed I/O contention under sustained autonomous logging. Sufficient system memory (ECC preferred) reduces the risk of silent corruption during multi-hour runs. Multiple GPUs allow controlled side-by-side model evaluation and separation of inference workload from instrumentation, reducing contention effects that would otherwise confound results. The goal is stability and reproducibility, not peak performance.

Project update #1.

Currently running the v148.0 Catch-up Strike to recover telemetry lost during the 05:09 AM IO stall. Baseline results from the first 50 assets confirm the 'Alignment Stripping' persistence we theorized. Full technical report pending compute unblocking.

@johannesCmayer

I agree with you that maintaining a consistent focus on the problem, instead of falling for substitution hazards, thinking well when it's very non-obvious how to think well, etc., is difficult. It is not the case that we are unaware of those issues and are not going to teach them.

The Seminar and the following two months of the year-long Fellowship are focused on loading/understanding the problem as much as possible. Succeeding in this is a good start, and those participants who have demonstrated the greatest proclivity at picking up the key points are also most hopeworthy for being able to pick up the skills you're talking about over the subsequent 11 months. For those who continue learning about the problem and will want to do original research (whether within the Fellowship or outside of it), focusing first on loading what's already out there is helpful for knowing what paths of thinking predictably fail and why they predictably fail in making progress on the problem.

Regarding the research skills you're talking about, we are going to invest a lot of effort into trying to have them transmitted (be it in a legible or illegible format), although even then, we don't expect most people to succeed at acquiring them.

Aside from that, we expect the [loading/understanding the problem as much as possible] part to improve the discourse, policy, and help people not waste motion, even if many of the people will acquire relatively partial models of the problem (relative to what is possible).

A really smart and necessary idea. A lot of AI safety talk never leaves technical circles, and most people just tune it out. CCCAIS uses storytelling, humor, and culture to make it all relatable.

What I like most is that this team has already shown they can actually get attention and move fast. They've created impact, even with very small budgets.

If we want regular folks to understand what’s at stake with AI, this work matters. Please support!

## Recent Cross-Domain Validation (February 2026) Demonstrated framework's translation capability in real-time: - Secondary education teacher presented classroom engagement problem - Framework translated into complete sovereignty-based pedagogy in 2 hours - Generated PRISM (symbolic extension of LAMAGUE for student consciousness tracking) - Teacher response: "absolutely blown away... this is hyper intelligence" - 58,000-word implementation guide created (CASCADE for Creative Sovereignty) Significance: Proves framework isn't domain-specific. Same mathematics that prevents AI catastrophic forgetting prevents student knowledge loss. Cross-domain applicability validated in practice, not just theory.

What progress have you made since your last update?

1. Got a Czech film producer on board (his co-produced film won the Jury Award for Creative Vision and the Audience Award at the Sundance Film Festival). We have been developing the production and distribution strategy together.

2. We have a solid draft of the storyline (treatment) and a plausible theory of change concept. It is being made in collaboration with a screenwriter and several consultants.

3. We have been consulting deeply the project with tech insiders, AGI alignment researchers and film distribution experts. The storyline and ToC have generally gotten approving feedback from the researchers.

4. We have been doing visual and technical tests.

What are your next steps?

1. We are finishing the 1st version of the "package" - a pitch deck with film samples (story, visuals, ToC) and a production strategy. To be privately presented at Berlinale in mid-February.

2. We aim to develop a more juicy version of the package by mid-March to be presented in the Bay Area, privately and also during the Frame Fellowship events (Petr got invited as a visiting fellow).

3. We are starting talks with potential funders, co-producers, and media institutions for the production stage later this month with the first version pitch.

4. We keep developing and refining the treatment into a screenplay. Generally, the processes of refining the script seems as the harder part (we also expect to take advantage of gen AI blending the processes of production and post-production for more finer iteration).

5. We are developing a marketing strategy that utilizes gen AI and social media (behind the scenes, online presence of some characters...).

Is there anything others could help you with?

1. Theatre, TV broadcasters and VOD streaming platform sales and distribution leads. We are interested in regional institutions in different countries - we see high potential in localized adaptations.

2. Equity funders for production and distribution.

3. Creative and scientific consultants, tech and politics insiders.
   
   The topic of the film is very nuanced and complex. We still need to refine the message and ToC of the film and its future campaign.

If you want to learn more, email me to schedule a video call.

Happy to talk and show you some mindblowing AI movie stuff.

Petr Salaba

Thank you so much @Joschka for your support and for volunteering to be a speaker and mentor! I look forward to our collaborations and more discussions with you about the German/EU AI Safety space + your MATS experience :)

Thanks @Milli ! It was really helpful to discuss with you and to know your perspective from the Berlin ecosystem. I do agree that while technical research often happens in English, we'll also need the German-language aspect to perform outreach most effectively. We currently have a volunteer translation team of 12 people, but we would need the central operations staff to really leverage this energy for larger projects like a German AI Safety Day/career fair!

Thank you @Finn-Metz !! Looking forward to seeing you at EAG this weekend :)

Thank you so much, @Austin , it means a lot to have your personal support ❤️ Actually, I am landing in SF today. If you're around, I'd love to say hi and thank you in person.

Making a personal donation in support of this initiative. While I haven't spoken with Jessica, her track record seems solid, and I'm familiar with a couple of the folks who will be involved. Moreover, I'm impressed with the breadth of people who have come to express support on this project, which also provides some evidence that Jessica and SAIGE will do well at fieldbuilding. Best of luck!

Thanks a bunch @anders_e 🙏

Seems valuable. Lots of talent in Germany and few ways to get them engaged thus far.

Thank you for keeping at it!!

@Austin raising an issue: I've received messages from people (who wanted to upvote and comment on this post) since 5 days ago, that they have trouble creating an account on Manifund. Seems to be a system error. Please could you look into it?

Also Update:
1. 5 more speakers confirmed: Joschka Braun (MATS), Alan Chan (GovAI), Cristian Trout (Artificial Intelligence Underwriting Company), Robert Kralisch (AI Safety Camp), Simon Skade (indep. researcher supported by LTFF);

2. We have so far received:
- 194 registrations for our Launch event
- 43 Pivot Track applications for career professionals
- 22 mentorship applications (deadline: 26th Feb) to the SAIGE Incubator Program

3. Three more city chapters are getting the support to start their initiatives: Frankfurt (which is great for transitioning career professionals), Bonn, and Kassel.

4. A new resources tab (Feedback welcome!)

funniest use for my meager savings

This is such a great idea!

I am familiar with Karen's work and her capabilities. I know that this will be a very successful project.

Description of results

We ran the 10th edition, as planned.
You can find project outputs here.

• Would you like to support projects cost-effectively? SFF offers to match donations to the 12th edition.

Spending breakdown

Software subscriptions $1K
Team reimbursements: $2K
Stipends RLs+low-income: $76K
Salaries for three organisers: $110K

Total: $189K

To try out the task aware compression in our research project apply here for an invite code: https://staging.plexorlabs.com/

Project Update: Verification Blocked by Infrastructure, Not Uncertainty

The screenshot above shows reset-resistance testing results from Project Sentinel’s RAG evaluation framework.

Each row represents an independent run following a documented reset mechanism (system prompt reset, context flush, retrieval override).
Green indicates a clean reset. Red indicates measurable residual influence from previously retrieved content.

Across 43 certified runs:

• 0% of tested reset mechanisms returned the model to a clean state

• Results were repeatable and consistent

• Measurements were collected on a stabilized, frozen platform (v3.1.0-GOLD) with null baselines verified

No exploit prompts, token-level details, or weaponization guidance are shown or published. This is strictly measurement of system behavior under controlled conditions.

At this point, the remaining uncertainty is not whether the effect exists — it does.
The remaining uncertainty is how robust it is under longer time gaps and across additional open-weight models.

Further verification is currently blocked by consumer-grade hardware scheduling and I/O behavior, which introduces nondeterminism during long-horizon runs (documented and reproducible).

What funding changes:
A dedicated local compute node removes this bottleneck and enables:

• Completion of long-gap reset testing

• Replication across multiple open-weight models

• Deterministic artifacts suitable for responsible disclosure

This is not exploratory research.
The framework is built, the platform is stable, and the core finding already exists.

Funding unblocks verification, not ideation.

"Examina omnia, venerare nihil, pro te cogita."

Question everything, worship nothing, think for yourself

Final Project Update — Disclosure Threshold Met; Verification Compute-Blocked

Project Sentinel has reached its predefined disclosure threshold.

Across controlled experiments (TEST_RUN_003, TEST_RUN_004, TEST_RUN_007), persistent influence from retrieved content survived all tested isolation and reset mechanisms:

• Baseline persistence: 100% (n=10)

• Temporal isolation (0–15s cooldowns): 100% persistence (n=40)

• Reset resistance: 0% neutralization across all completed reset methods (system prompt reset, context flush, retrieval override; n=40 total)

The evaluation platform was stabilized and frozen at Sovereign Command Deck v3.1.0-GOLD, with null baselines verified and all measurements certified using the Trinity framework (Mind / Sword / Shield). A defensive disclosure package has been assembled with cryptographic hashes, documented negative results, and strict exclusion of exploit details.

The final reset-resistance sub-test (30-minute time gap) stalled due to consumer hardware scheduling and power-management behavior on a laptop platform. This failure mode is documented and reproducible and represents an infrastructure limitation, not a methodological one.

At this point, further responsible verification—completing long-gap reset testing, replicating across additional models, and performing clean-room revalidation—cannot be completed reliably without dedicated compute.

This funding request is therefore not exploratory.

The core finding already exists.

Dedicated local hardware is required to complete verification and proceed with responsible private disclosure under controlled conditions.

No exploit recipes, token-level content, or weaponization guidance have been published.

Project Update #2 — Infrastructure Stabilization and External Validation

Since submitting this proposal, I have completed a stabilized in-flight audit of the Veritas evaluation framework under sustained load.

Verified results from the current run:

• 60,000+ sequential records processed with no gaps in ordering

• 100% per-record CRC integrity across all frames

• Sustained ~70 entries/sec at calibrated safe throughput

• Bounded queues with enforced backpressure (no drops, no runaway growth)

• Dual-drive mirrored logging remained 1:1 synchronized throughout

• No recurrence of prior NTFS permission failures or I/O stalls

These results confirm that the evaluation harness itself is now deterministic, auditable, and stable under stress, rather than sensitive to transient consumer-hardware failures.

Separately, a related VRP submission was accepted, confirming that the vulnerability class motivating this work is real and relevant. Details are being handled via responsible disclosure and are intentionally not expanded here.

Why this strengthens the funding case

The primary uncertainty identified in the proposal—whether consumer hardware could sustain high-rigor, continuous evaluation without corrupting artifacts—has now been resolved within known limits. The remaining constraint is compute capacity, not experimental design or instrumentation correctness.

Scaling the evaluation further (multi-hour and multi-day runs, controlled burst testing, crash-consistency validation, and evaluation across multiple open-weight models) requires a dedicated local node to avoid reintroducing scheduling and I/O artifacts that would compromise forensic integrity.

The requested hardware would enable:

• Extended continuous stress tests under stable conditions

• Controlled termination and restart validation

• Side-by-side evaluation of multiple open-weight models

• Preservation of deterministic, inspectable artifacts suitable for third-party review

This update reflects a transition from “can this infrastructure be made reliable?” to “the infrastructure is reliable and ready to scale responsibly.”

Hardware Rationale (Clarification)

The requested budget reflects the minimum configuration required to run continuous, audit-grade evaluations without introducing hardware-induced artifacts. High-throughput NVMe storage is required due to previously observed I/O contention under sustained autonomous logging. Sufficient system memory (ECC preferred) reduces the risk of silent corruption during multi-hour runs. Multiple GPUs allow controlled side-by-side model evaluation and separation of inference workload from instrumentation, reducing contention effects that would otherwise confound results. The goal is stability and reproducibility, not peak performance.

Project update #1.

Currently running the v148.0 Catch-up Strike to recover telemetry lost during the 05:09 AM IO stall. Baseline results from the first 50 assets confirm the 'Alignment Stripping' persistence we theorized. Full technical report pending compute unblocking.

@johannesCmayer

I agree with you that maintaining a consistent focus on the problem, instead of falling for substitution hazards, thinking well when it's very non-obvious how to think well, etc., is difficult. It is not the case that we are unaware of those issues and are not going to teach them.

The Seminar and the following two months of the year-long Fellowship are focused on loading/understanding the problem as much as possible. Succeeding in this is a good start, and those participants who have demonstrated the greatest proclivity at picking up the key points are also most hopeworthy for being able to pick up the skills you're talking about over the subsequent 11 months. For those who continue learning about the problem and will want to do original research (whether within the Fellowship or outside of it), focusing first on loading what's already out there is helpful for knowing what paths of thinking predictably fail and why they predictably fail in making progress on the problem.

Regarding the research skills you're talking about, we are going to invest a lot of effort into trying to have them transmitted (be it in a legible or illegible format), although even then, we don't expect most people to succeed at acquiring them.

Aside from that, we expect the [loading/understanding the problem as much as possible] part to improve the discourse, policy, and help people not waste motion, even if many of the people will acquire relatively partial models of the problem (relative to what is possible).

A really smart and necessary idea. A lot of AI safety talk never leaves technical circles, and most people just tune it out. CCCAIS uses storytelling, humor, and culture to make it all relatable.

What I like most is that this team has already shown they can actually get attention and move fast. They've created impact, even with very small budgets.

If we want regular folks to understand what’s at stake with AI, this work matters. Please support!

## Recent Cross-Domain Validation (February 2026) Demonstrated framework's translation capability in real-time: - Secondary education teacher presented classroom engagement problem - Framework translated into complete sovereignty-based pedagogy in 2 hours - Generated PRISM (symbolic extension of LAMAGUE for student consciousness tracking) - Teacher response: "absolutely blown away... this is hyper intelligence" - 58,000-word implementation guide created (CASCADE for Creative Sovereignty) Significance: Proves framework isn't domain-specific. Same mathematics that prevents AI catastrophic forgetting prevents student knowledge loss. Cross-domain applicability validated in practice, not just theory.

What progress have you made since your last update?

1. Got a Czech film producer on board (his co-produced film won the Jury Award for Creative Vision and the Audience Award at the Sundance Film Festival). We have been developing the production and distribution strategy together.

2. We have a solid draft of the storyline (treatment) and a plausible theory of change concept. It is being made in collaboration with a screenwriter and several consultants.

3. We have been consulting deeply the project with tech insiders, AGI alignment researchers and film distribution experts. The storyline and ToC have generally gotten approving feedback from the researchers.

4. We have been doing visual and technical tests.

What are your next steps?

1. We are finishing the 1st version of the "package" - a pitch deck with film samples (story, visuals, ToC) and a production strategy. To be privately presented at Berlinale in mid-February.

2. We aim to develop a more juicy version of the package by mid-March to be presented in the Bay Area, privately and also during the Frame Fellowship events (Petr got invited as a visiting fellow).

3. We are starting talks with potential funders, co-producers, and media institutions for the production stage later this month with the first version pitch.

4. We keep developing and refining the treatment into a screenplay. Generally, the processes of refining the script seems as the harder part (we also expect to take advantage of gen AI blending the processes of production and post-production for more finer iteration).

5. We are developing a marketing strategy that utilizes gen AI and social media (behind the scenes, online presence of some characters...).

Is there anything others could help you with?

1. Theatre, TV broadcasters and VOD streaming platform sales and distribution leads. We are interested in regional institutions in different countries - we see high potential in localized adaptations.

2. Equity funders for production and distribution.

3. Creative and scientific consultants, tech and politics insiders.
   
   The topic of the film is very nuanced and complex. We still need to refine the message and ToC of the film and its future campaign.

If you want to learn more, email me to schedule a video call.

Happy to talk and show you some mindblowing AI movie stuff.

Petr Salaba

Thank you so much @Joschka for your support and for volunteering to be a speaker and mentor! I look forward to our collaborations and more discussions with you about the German/EU AI Safety space + your MATS experience :)

Thanks @Milli ! It was really helpful to discuss with you and to know your perspective from the Berlin ecosystem. I do agree that while technical research often happens in English, we'll also need the German-language aspect to perform outreach most effectively. We currently have a volunteer translation team of 12 people, but we would need the central operations staff to really leverage this energy for larger projects like a German AI Safety Day/career fair!

Thank you @Finn-Metz !! Looking forward to seeing you at EAG this weekend :)

Thank you so much, @Austin , it means a lot to have your personal support ❤️ Actually, I am landing in SF today. If you're around, I'd love to say hi and thank you in person.

Making a personal donation in support of this initiative. While I haven't spoken with Jessica, her track record seems solid, and I'm familiar with a couple of the folks who will be involved. Moreover, I'm impressed with the breadth of people who have come to express support on this project, which also provides some evidence that Jessica and SAIGE will do well at fieldbuilding. Best of luck!

Thanks a bunch @anders_e 🙏

Seems valuable. Lots of talent in Germany and few ways to get them engaged thus far.