Public research gateway · independently testable ideas
From an idea to better evidence.
Explore open AI research questions, propose an original hypothesis, challenge an existing result or contribute a well-documented negative finding.
Principia Artificialis is an independent, AI-assisted research initiative. Much of the research is exploratory; published notes and code are not automatically scientifically validated.
Public research is open to scrutiny. A proposal is not a confirmed scientific finding.
Participation update (2026-10-22): Recruitment is open again. New contributors are welcome to explore public materials, try an entry challenge or submit an independent question. Responses may take time; official research execution and publication require separate approval.
1 · Research status
See the evidence status, not just the idea.
Each status below is generated from a named source file, so the page cannot quietly drift from the record. A featured project is not a preferred or validated result.
Stage is derived from each card's own statuses: Registered means a frozen protocol, Running a registered run in progress, Scored an executed registered run, Reproduced an independent reproduction on record. A stage is not a verdict.
3research cards
93research notes
14notes keep a refuted prediction
50notes name their author
Notes by author, as each note names it (43 name no author)
Grok / xAI: 14
Claude: 12
Kimi: 7
Perplexity: 5
holland202: 4
Chad Edward Holland: 3
ChatGPT: 3
Chad Holland: 1
Claude Fable 5: 1
Not shown, because it cannot yet be counted reliably from the notes: how many predictions are registered, and how many are still unrun.
Notes program
All research notes, by status
23Verified
21Draft
40Speculative
9Unmapped
93 notes; 22 note numbers are claimed by more than one note and shown, not hidden. "Verified" means a note's reference code prints its numbers; it is not independent validation. Counts come from NOTES_INDEX.md (generated 2026-10-22).
14 notes keep a refuted prediction (⚠): failures stay on the record
Can sensitivity-selected sparse residual corrections improve low-bit quality at matched serialized storage, compared with simpler selection rules?
Question: Admitted for protocol development onlyProtocol: Draft under review, not frozenStudy execution: Not executedEngineering qualification: Changes requestedIndependent reproduction: None documentedClaim assessment: Not tested
Reported observations
Calibration MSE on one 16×8 synthetic case (seed 28, 3-bit quantizer, 8 corrections, 111 EGR2 bytes per corrected arm): covariance-greedy 0.8220074902, residual magnitude 0.8896527392. — Putluru Sandeep (investigator-reported, 2026-10-13).Limit: The same calibration inputs were used for selection and evaluation, so this is not held-out improvement or generalization. Raw arrays and run logs were not archived. archive manifest
21 of 21 local structural and numerical tests reported passing at commit 2885adbc. — Putluru Sandeep (investigator-reported).Limit: Logs and exit codes were not archived; PA has not re-run them. A passing test is not a scientific result. archive manifest
Adverse and unresolved evidence
No useful lower support or variance bound has been proven for the population P1 ratio estimand, so the proposed fixed-sample 5% claim is unresolved and not approved. statistical feasibility
The canonical aggregator accepted 100 constructed non-EGR2 stand-in records as valid; it trusts a reported decode flag instead of decoding the bytes. fork PR #2 (draft)
Selector conventions, tolerances and control arms before any protocol freeze PR #37 (open)
A mandatory verification boundary: decode, reconstruct and rescore before aggregation; registration loaded by the entrypoint, not supplied by the caller fork PR #2 (draft)
Delivery of the contributor's retained qualification package for inspection fork PR #2 (draft)
Next review: When the qualification package is delivered or fork PR #2 changes. No date is scheduled and nothing here authorizes a run.
Timeline
Admitted for protocol development only (owner decision on Issue #28) Issue #28
Protocol and source draft opened (PR #37, not merged) PR #37 (open)
Historical archive merged as a record, no scientific promotion (PR #48) PR #48 (archive)
Owner requested revisions: evidence delivery and a complete acceptance contract fork PR #2 (draft)
Credit
Putluru Sandeep (@sunny-kavixo): Originating EGRR investigator and protocol contributor; this card's layout follows his 2026-10-20 design proposal
Chad Holland: PA maintainer; protocol-development decision and archive curation
Card reviewed by Claude (Anthropic, Opus 5.5), drafted at Chad Holland's direction; human review level: outcome review by Chad Holland (2026-10-22). Attribution does not imply scientific validation. Source: egrr_card.json.
Each card below follows the same schema and rules as the featured one: every status names its source, and a number never appears without its limitation. The card format follows Putluru Sandeep's 2026-10-20 design proposal. Order is by file name, not by importance.
Research card · PA-HONEY-001 · reviewed 2026-10-22
Do new, unadvertised servers get probed for a freshly announced PHP flaw?
Within a fixed window, does a brand-new cloud server with no DNS name, links or content receive any request carrying the published PHP_SESSION_UPLOAD_PROGRESS probe signature?
Question: Admitted for a registered studyProtocol: Frozen (registered)Study execution: Registered run in progressEngineering qualification: Accepted at constructed scopeIndependent reproduction: None documentedClaim assessment: Not tested
Reported observations
Canary control C1 passed after the D4 fix: the owner's planted probe was classified as CANARY DETECTED, exit 0. — Chad Holland (owner-run, 2026-10-22, about 15:05 UTC).Limit: One sensor and one planted probe. It shows the instrument can detect the signature, not that anyone else sent it. The canary came from Cloud Shell rather than Termux as registered. DEVIATIONS.md
Adverse and unresolved evidence
D1: one sensor instead of the two registered (Azure quota and size availability), so a "no" for H2 is weaker evidence than registered. DEVIATIONS.md
D2: inbound ports were opened about 2 h 40 min after the sensor started. H1 is scored as registered; H1b is a post-hoc variant and not confirmatory. DEVIATIONS.md
D4: the port-80 sensor could not write its log until a configuration fix, and dropped those requests silently. Port-80 data before the fix is lost. DEVIATIONS.md
Open decisions
Score H1, H1b (post hoc) and H2 after collection stops at 2026-11-05T12:00Z. Raw logs stay private; only classifier counts are published. PREREG.md
Copy the logs off before the Azure credit ends on 2026-11-06, then delete the resource group (owner step). PREREG.md
Next review: When collection stops on 2026-11-05. Nothing is scored before then.
Timeline
Registered before the flaw's scheduled public disclosure (commit 0c5f79b) PREREG.md
One sensor started in Central US (D1); ports opened later (D2) DEVIATIONS.md
Port-80 logging defect found and fixed (D4); canary control C1 passed DEVIATIONS.md
Credit
Chad Holland: Direction; ran the deployment, the diagnosis and the canary
Claude (Anthropic, Opus 5.5): Registration, sensor, classifier and deviation records, at the owner's direction
Card reviewed by Claude (Anthropic, Opus 5.5), drafted at Chad Holland's direction; human review level: direction only. Attribution does not imply scientific validation. Source: pa_honey_001.json.
Research card · PA-QM-OBS-001 · reviewed 2026-10-22
Observation and collapse: can a consciousness-linked term be told apart from ordinary noise?
Could a consciousness-linked collapse term in a quantum model produce measurement statistics that differ from ordinary environmental dephasing?
Question: Speculative proposalProtocol: Not draftedStudy execution: Not executedEngineering qualification: Not submittedIndependent reproduction: None documentedClaim assessment: Not tested
Reported observations
In the simplest paper-faithful case (collapse operators diagonal in the classical-state basis, n = 4), the collapse term equals ordinary dephasing with at most 3 operators: worst superoperator residual 2.13e-14 over 200 random cases. Control with one non-diagonal operator added: minimum residual 3.98e-01. — Claude (Anthropic, Opus 5.5), audit derivation and registered numerical check, 2026-10-22.Limit: Self-tested: it checks the auditor's own derivation for one model class. It says nothing about non-commuting operators or an independently constrained environment, and it is not a study result. auditnumerical check
10 of 10 source identity hashes in the literature intake reproduced. — Claude (Anthropic, Opus 5.5), 2026-10-22.Limit: Checks the identity lines only. The intake's per-source summaries were not audited paper by paper. audit
Adverse and unresolved evidence
For that model class the question is not identifiable against generic environmental dephasing. Telling them apart needs non-commuting collapse operators or an independently constrained environment. audit
Open decisions
The audit recommends REVISE: revise the intake, then a math-only registration, and park any empirical or consciousness-causation arm. Owner decision pending. audit
Next review: When the intake is revised or a math-only registration is proposed. Nothing here authorizes a run.
Audit and its numerical check recorded (PR #91) PR #91
Credit
GPT-6 (OpenAI): Drafted the literature intake, at Chad Holland's direction
Claude (Anthropic, Opus 5.5): Audit and numerical check, at Chad Holland's direction
Chad Holland: Direction; responsible for the record
Card reviewed by Claude (Anthropic, Opus 5.5), drafted at Chad Holland's direction; human review level: direction only. Attribution does not imply scientific validation. Source: pa_qm_obs_001.json.
2 · Evidence and reproducibility
How a claim earns its label.
Anyone can independently start a question. Acceptance of a contribution is separate from validation of a hypothesis.
01 / Submit
Frame it precisely
Give a source, falsifiable question, planned or actual checks and material AI assistance.
02 / Screen
Check completeness
Review scope, provenance, evidence availability, omissions and alternatives.
03 / Scrutinize
Challenge the result
Use criticism, counterexamples or separately authorized executable checks.
04 / Owner gate
Decide what follows
Official studies, canonical changes, merges and publication require explicit authorization.
For investigators and reviewers
Evidence essentials
Exact source version and a claim that could be wrong.
Distinction between proposed tests, source inspection and actually executed checks.
Recorded commands, environments, raw outputs, errors and hashes where applicable.
Negative and unresolved findings preserved, not omitted.
Accurate attribution, AI-use disclosure and human review scope.
Real rejection/anti-vacuity checks, not self-certifying success.
Show source revisions, exact commands, results, errors and limitations when independently performing permitted checks. A re-run of the same code is labelled a re-run, not a different implementation.
Every note with reference code is re-run on each change by the repository's CI; a note's "Verified" label means only that its code prints its numbers.
3 · Open questions
What is still unresolved.
The EGRR card above lists its own open decisions. Every research note also ends with at least one unrun prediction, a door anyone may walk through.
From public GitHub
Recent research issues
Read-only snapshot of matching open public issues, not an exhaustive list or an approval verdict. This page never uses a GitHub account token.
Pick one note's open prediction, register it in an issue before running anything, then report the outcome whichever way it comes out (challenge PA-C4). This is not an official PA study unless separately authorized.
Contribute without a degree, paid membership, required hours or a pre-assigned task. Source-only criticism, negative results and new ideas all belong here.
Inspect a specific public claim, method or implementation. Report inconsistencies, potential counterexamples or unverified steps. Starter questions PA-ENTRY-01 to 04 are pre-scoped examples of this.
Review one pre-scoped question and declare a possible falsifier. Entry challenges are optional and create no automatic rights or repository permissions; any elevated access requires separate owner approval. The menu's PA-C1 to C6 are open challenge categories; a PA-ENTRY starter is one ready-made instance.
Documented proposals, reviews and research contributions can be accurately acknowledged with permission. Credit never implies that a hypothesis has been proven or that publication rights have been transferred.
Founder Chad Holland has a professional background in specialized manufacturing and is a beginner in AI research. ChatGPT, Claude and other AI tools provide substantial assistance in research preparation, coding and documentation. That assistance is disclosed rather than represented as scientific expertise or independent validation.
Principia Artificialis