Private intelligence lab · public evidence instrument
Proof of
Becoming
What changed. What broke. What survived. Which assumptions became unnecessary.
Most roadmaps announce what they’ll add. Ours shows which assumptions have become unnecessary.
00 / Public Proof Pulse
Watch the system
change its beliefs.
probagi.publicPulse.lastSeen.
Release survived
Scope Release 001 public-site court
Source Executable theme and live-browser receipt
Limit Publication behavior only; semantic truth remains claim-scoped. 02 / PUBLIC_REDACTED 1Verdict overturned
Scope Case 000 historical disposition
Source Curated public case replay
Limit One case, not an aggregate model reliability rate. 03 / PUBLIC_REDACTED 4Exact-byte defects
Scope Independent review of Case 000
Source Admitted claim manifest
Limit Private identifiers and unsafe reproduction details remain withheld. 04 / PUBLIC 1Assumption deleted
Scope “One model grades its own work”
Source A-001 change record
Limit The replacement council reduces one failure mode; it does not guarantee correctness. 05 / PUBLIC 3Open claims
Scope CLAIM-000 through CLAIM-002
Source Public manifest drawers
Limit CLAIM-003 remains held and is excluded from this count.Admitted events, not activity theater.
Ordered public case record · dates reflect the admitted source, not a continuous feed.
- VERDICT
Case 000 returned an initial 91/100 SHIP disposition.
PUBLIC_REDACTEDCase replay ↗ - DEFECT
Independent review admitted four exact-byte defects.
PUBLIC_REDACTEDClaim + limits ↗ - OVERTURN
The 91/100 SHIP verdict was reversed and preserved as a public loss.
PUBLICLoss record ↗ - DELETION
The assumption “one model grades its own work” was deleted.
PUBLICChange record ↗ - RELEASE
Release 001 survived its bounded public-site court.
PUBLICCourt receipt ↗
From witness to attempted refutation.
Interest signals do not open held intake, jury, or scanner systems.
New public receipts only when the evidence changes.
OPEN · GHOST MEMBERS 02Challenge a claimSignal interest; identity, consent, abuse, and disclosure intake remains held.
INTEREST ONLY · INTAKE HELD 03Join the jurySignal independent-review interest before policy and evidence access open.
INTEREST ONLY · POLICY HELD 04Inspect a public scanOpen the admitted WIW sample; new-repository intake remains held behind its sandbox and privacy court.
PUBLIC SAMPLE · INTAKE HELD 05Watch the seriesFollow Probably AGI for the next public case transmission.
YOUTUBE ↗01 / The journey
One question,
changing shape.
Origin question
Mr. Fox
Early human/AI collaboration begins the record. Selected origin fragments remain held until they can be shown without turning a conversation into proof of AGI.
03 / Case 000 replay
The verdict
didn’t survive.
Witness boundary
Frozen bytes
The artifact under review is fixed before judgment. Later edits cannot quietly improve the evidence that earned the score.
- Authority
- Witness
- State
- INPUT LOCKED
Public replay · Case 000 · not live
BuilderShips the candidate
SkepticAttacks exact bytes
JudgeOwns disposition
WitnessRuns independent probes
Authority follows the receipt.
Never the model name.
Each role is a deliberately different incentive. The Einstein faces are theatre; the frozen scope, evidence, and disposition are the authority.
Operator
- Incentive
- Choose the purpose worth pursuing.
- Authority
- Intent, constraints, and release scope.
- Cannot claim
- That delegated output reflects choices never knowingly made.
Builder
- Incentive
- Ship the strongest coherent candidate.
- Authority
- Implementation and declared evidence.
- Cannot claim
- Its own confidence is settlement.
Skeptic
- Incentive
- Find the exact contradiction that matters.
- Authority
- Challenges and falsifiers.
- Cannot claim
- Suspicion is a defect without evidence.
Judge
- Incentive
- Freeze the narrowest defensible verdict.
- Authority
- Current disposition and its limits.
- Cannot claim
- A verdict survives contradictory evidence.
Witness
- Incentive
- Repeat the named probe independently.
- Authority
- Observed evidence against frozen input.
- Cannot claim
- A passing probe proves unknown semantics.
Replay boundary Curated public case evidence only. Private prompts, identities, session metadata, and unsafe reproduction details do not cross into this theatre.
04 / Public loss record
One loss.
Kept on purpose.
Failed judgment
A 91/100 SHIP verdict did not survive independent exact-byte review.
- Assumption that failed
- One evaluator’s confidence can grade the work it just accepted.
- Repair admitted
- Builder, skeptic, judge, and witness carry separate incentives against frozen evidence.
- What this does not prove
- This single historical case is not an aggregate reliability rate or a guarantee about future councils.
This wall currently contains one admitted case. No extra rows are implied. Future overturns, retractions, regressions, and repairs appear only when their evidence can stand beside the win record with equal dignity.
1 ADMITTED RECORD · NO VANITY FILLER05 / Assumption ledger
The next things
we’re deleting.
Not a victory wall. A change record with disagreement left intact.
A-001
Old assumption
One model grades its own work.
Current disposition
Deleted · Case 000Only this row has crossed every stage with an admitted public case receipt.
Deleted+
Old assumption
One model grades its own work.
Current disposition
Deleted · Case 000Only this row has crossed every stage with an admitted public case receipt.- 01Hypothesis
- 02Built
- 03Attacked
- 04Settled
- 05Survived
- 06Deleted
ReplacementBuilder, skeptic, judge, and independent review carry separate incentives.
BoundaryThis deletes one workflow assumption; it does not prove the council is correct.
Inspect Case 000 receipt
A-002
Old assumption
Humans must hand-write the source code.
Current disposition
Attacked · Public proof 001One exact FoxLang chain is now reproducible; the general claim remains open.
Public+
Old assumption
Humans must hand-write the source code.
Current disposition
Attacked · Public proof 001One exact FoxLang chain is now reproducible; the general claim remains open.- 01Hypothesis
- 02Built
- 03Attacked
- 04Settled
- 05Survived
- 06Deleted
Admitted evidenceFrozen intent, exact selected atom, typed IR, emitted Rust, binary identities, runtime verdict, and a separate direct-rustc replay.
BoundaryOne successful chain attacks the assumption; it does not delete it for all programs.
Inspect FoxLang proof 001
A-003
Old assumption
Code is the only source of truth.
Current disposition
Attacked · Intent + receiptsA receipt can preserve lineage without proving that the chosen intent was right.
Held+
Old assumption
Code is the only source of truth.
Current disposition
Attacked · Intent + receiptsA receipt can preserve lineage without proving that the chosen intent was right.- 01Hypothesis
- 02Built
- 03Attacked
- 04Settled
- 05Survived
- 06Deleted
Public gateShow that versioned intent can reproduce an accepted artifact inside a frozen tolerance.
Strongest objectionConformance to a policy cannot prove the policy or purpose is correct.
A-004
Candidate assumption
Git history is the final settlement layer.
Current disposition
Hypothesis · Not promotedThe replacement settlement contract has not survived a public court.
Held+
Candidate assumption
Git history is the final settlement layer.
Current disposition
Hypothesis · Not promotedThe replacement settlement contract has not survived a public court.- 01Hypothesis
- 02Built
- 03Attacked
- 04Settled
- 05Survived
- 06Deleted
Public gateA frozen settlement receipt must bind intent, emitted bytes, attacks, runtime evidence, and disposition.
Current truthGit remains part of this release workflow.
A-005
Candidate assumption
PRs, branches, and worktrees are always required for safe coordination.
Current disposition
Hypothesis · Not promotedThis site still uses a dev-to-main release path; no deletion claim is available.
Held+
Candidate assumption
PRs, branches, and worktrees are always required for safe coordination.
Current disposition
Hypothesis · Not promotedThis site still uses a dev-to-main release path; no deletion claim is available.- 01Hypothesis
- 02Built
- 03Attacked
- 04Settled
- 05Survived
- 06Deleted
Public gateDemonstrate collision-safe multi-agent settlement without hiding or overwriting concurrent work.
Current truthThe alternative remains a research hypothesis, not this release policy.
A-006
Candidate assumption
Programming languages must be designed around human punctuation.
Current disposition
Attacked · One bounded chainExact decomposition is public; general language replacement is not claimed.
Public+
Candidate assumption
Programming languages must be designed around human punctuation.
Current disposition
Attacked · One bounded chainExact decomposition is public; general language replacement is not claimed.- 01Hypothesis
- 02Built
- 03Attacked
- 04Settled
- 05Survived
- 06Deleted
Attack admittedPublic proof 001 crosses intent → atom → typed IR → source → binary → exact-output replay → receipt.
Strongest objectionNatural-language intent can hide ambiguity more effectively than punctuation does.
Open the exact chainA receipt is a boundary,
not a halo.
Select each stage to inspect what it guarantees—and what still remains unproven.
Current boundary
Survived
- Guarantee
- The frozen settlement remained valid through the declared release court.
- Does not guarantee
- Survival expires outside the recorded artifact, checks, scope, and verification time.
06 / FoxLang proof set
The rejection stayed public.
The repair survived.
VERIFIED EMBEDDED · RELEASE-BOUND
- 01Concurrency retrial25 typed nodes · timing requiredREPAIRED / ADMITTED
- 02JSON exactness retrialBTreeMap · one hash / 80 runsREPAIRED / ADMITTED
- 03Sort + filterOne hash across 40 runsADMITTED
“save a status report to an output file and confirm bytes written”
live_crystal_selection → file_writer
Current proof boundary
Purpose enters as a sentence
- Evidence
- The exact 65-character intent is frozen in this receipt.
- Does not prove
- A sentence can still be ambiguous or wrong.
Truth boundary Five frozen first-party receipts on one compiler target now preserve two loss→repair histories. Proof 004 binds the concurrency emitter repair at FoxFlow 0fc96f9d; proof 005 binds deterministic transformed top-level key order at 50e30b65 and one stdout hash across 40 emitter plus 40 projection runs. That is not RFC 8785 canonical JSON, recursive ordering, numeric normalization, arbitrary intent coverage, external review, a universal language thesis, or AGI. Executable bytes remain withheld.
07 / Classification law
Visibility is
part of the claim.
Safe to publish with a complete limitations record.
The public can inspect the conclusion and boundary, not reproduction details.
Reserved for future member releases; currently no gated claim is promoted.
Evidence available only to the bound reviewers assigned to test it.
Internal data that is neither implied nor simulated on the public surface.
Credentials, tokens, private identities, raw telemetry, and unsafe reproduction paths.
The lab opens by receipt,
not by blur animation.
5 admitted declassifications in the current release-bound snapshot.
No new records admitted.
Case 000 and the public release court are the baseline, not backfilled as a fake new arrival.
RELEASE-002 · VERIFIED EMPTY STATEOrigin conversation fragmentsPUBLIC_REDACTED candidate+
Opening gateContext-preserving selection, participant/privacy review, and a limitation record that prevents conversation from being presented as proof of AGI.
- Credentials, tokens, and secret-bearing filenames
- Private identities, native sessions, and local paths
- Raw telemetry or presence, travel, sleep, and downtime patterns
- Private repo names and unsafe reproduction routes
08 / Claim manifests
Open the
limitations.
CLAIM-000
Case record
A 91/100 SHIP verdict was overturned by four exact-byte defects.
Public+
Case record
A 91/100 SHIP verdict was overturned by four exact-byte defects.
- Definition
- The initial recorded Case 000 evaluation returned SHIP at 91/100; a separately authored exact-byte review then identified four defects and reversed the disposition.
- Time window
- Case 000 · 2026-07-17 to 2026-07-18
- Source class
- Curated public case replay
- Receipt / status
- Public record present · standalone public manifest not yet published
- Last verified
- 2026-07-18
- Classification
- PUBLIC_REDACTED
LimitationsThis is one historical case, not an aggregate reliability rate. Private session identifiers and unsafe reproduction details are omitted.
Enter Case 000
CLAIM-001
Publication primitive
Passing release courts can verify a site surface without proving semantic truth.
Public+
Publication primitive
Passing release courts can verify a site surface without proving semantic truth.
- Definition
- The release court tests routes, assets, responsive behavior, accessibility signals, privacy patterns, and metadata emitted by the published site.
- Time window
- Release 001
- Source class
- Executable repository checks + live HTTP/browser probes
- Receipt / status
- Theme contract present · live receipt attached at deployment
- Last verified
- 2026-07-18
- Classification
- PUBLIC
LimitationsA green site court proves only the tested surface at the recorded moment. It cannot establish the truth of unrelated scientific or product claims.
CLAIM-002
Repository evidence scanner
A pinned ripgrep commit produced a tamper-evident structural fact receipt.
Public+
Repository evidence scanner
A pinned ripgrep commit produced a tamper-evident structural fact receipt.
- Definition
- WIW structural-v1 froze one public repository at a full commit, classified scanner languages, and emitted deterministic structural facts.
- Time window
- One scan · observed 2026-07-19 UTC
- Source class
- Pinned public Git repository + SCC structural scan
- Receipt / status
- Fact receipt present · sha256:d7424384…d0610db
- Last verified
- 2026-07-19 UTC
- Classification
- PUBLIC
LimitationsOne sample is not a ranking. Dependencies, history, and tests remain unknown. The receipt proves frozen inputs and measured structure—not quality, correctness, market value, or semantic truth.
Inspect the WIW proof instrument
CLAIM-004
Intent-authorship proof
One FoxLang intent survived atom selection, typed IR, compilation, runtime evaluation, and exact-output recomposition.
Public+
Intent-authorship proof
One FoxLang intent survived atom selection, typed IR, compilation, runtime evaluation, and exact-output recomposition.
- Definition
- The frozen intent selected one verified file_writer atom, emitted a 13-node typed IR graph with zero RawCode nodes, compiled and ran, then reproduced identical stdout through a separate direct-rustc path.
- Time window
- One run · observed 2026-07-19 17:46 UTC
- Source class
- Public-safe FoxLang source slice + Pop!_OS execution receipt
- Receipt / status
- Evidence set present · sha256:b8f04c6d…68b47
- Last verified
- 2026-07-19 UTC
- Classification
- PUBLIC
LimitationsOne first-party process-independent replay is not third-party authority, language superiority, production readiness, or evidence of AGI. Executable bytes remain withheld from the theme.
Inspect proof 001
CLAIM-003
Repository scan intake
Submit any public repository for an instant council-backed scan.
Held+
Repository scan intake
Submit any public repository for an instant council-backed scan.
- Definition
- A future public job surface may resolve a public GitHub repository to a full commit, run the deterministic core in an ephemeral sandbox, and attach a separate council review.
- Time window
- Not opened
- Source class
- JOB-API CONTRACT only
- Receipt / status
- HELD FOR DATA
- Last verified
- No public job verification
- Classification
- PRIVATE → future PUBLIC_REDACTED candidate
LimitationsNo public submission form, mutable ref, private-repository intake, unbounded clone, value score, or AI verdict is available. Sandbox, abuse, privacy, calibration, and review courts must survive first.
09 / WIW evidence scanner
Freeze first.
Then measure.
BurntSushi/ripgrep
227381db0ee83dfa4341f1e27ff9617c0f5ad992
The deterministic core exists. The public job surface does not. New-repository submission stays closed until ephemeral sandboxing, mutable-ref resolution, abuse limits, privacy review, and council separation survive their own courts.
PUBLIC SAMPLE · JOB API HELD FOR COURT10 / Project atlas
Public silhouettes.
Closed engine room.
Seven approved systems, authored as falsifiable project surfaces. Every dossier carries its thesis, current stage, active attack, public evidence status, and the boundary that does not open.
7 surfaces · project-scoped evidence
Arrow keys move between nodes. Enter settles the selected dossier.
Proven
Separate witnessed evidence, candidate knowledge, governed memory, attention, deliberation, and action so a receipt proves only what it binds.
- Stage
- Architecture boundary published; substrate contents remain private.
- Current attack
- Can candidate relationships reach durable knowledge without silently becoming authority?
- Public evidence
- Receipt-chain guarantees, limitations, and the Case 000 settlement grammar.
- Classified
- Substrate contents, routing context, working memory, and action state.
YouDAG
One autonomous experiential universe observes lived events and itself; candidate relationships emerge through prediction, grading, strengthening, and decay.
Outside Proven and FDAG. Connected only through receipted attention, consolidation-candidate, and action-feedback bridge flows.
- Stage
- Canonical boundary public; active experiential topology private.
- Current attack
- Candidate edges must strengthen, decay, and face Court before promotion.
- Public evidence
- The journey records the boundary without reproducing the universe or its observations.
- Classified
- Observations, active topology, namespaces, timing, and predictive attention packets.
FoxFlow
Explicit roles, obligations, receipts, and courts coordinate multi-agent execution without allowing builder confidence to settle itself.
- Stage
- Role and authority silhouette admitted through the public Case 000 replay.
- Current attack
- Preserve attribution and exact scope when concurrent builders touch one evolving system.
- Public evidence
- Builder, skeptic, judge, witness, and public-record authority transitions.
- Classified
- Private messages, obligations, session state, task content, and collaboration bus.
FoxLang / IR
Versioned intent and constraints are explored as atoms and intermediate representation from which artifacts can be emitted and receipted.
- Stage
- Four exact public-safe intent-to-receipt chains admitted, including one preserved compile loss and repaired retrial.
- Current attack
- Define deterministic JSON ordering without weakening the exact-output court.
- Public evidence
- Frozen intents, selected Diamonds, typed IR, Rust source, binary identities, exact-output replays, loss records, and repair-bound source commits.
- Classified
- Private corpora, unpublished atoms, internal runs, and executable bytes remain outside this surface.
FoxEIS
Separate a scoped inference intervention from witness, measured effect, prediction, and authority so negative results remain useful without earning control.
- Stage
- Public-safe explanatory contract admitted; system internals stay private.
- Current attack
- Expose null and negative effects with the same dignity as improvements.
- Public evidence
- The journey records the intervention/witness/authority separation, not a performance claim.
- Classified
- Model artifacts, runtime topology, hardware measurements, and intervention traces.
WIW
Pin a public commit, measure structure deterministically, state unknowns separately, and refuse to turn structural facts into a magical dollar value.
- Stage
- First structural-v1 public sample admitted; public job intake remains held.
- Current attack
- Open safe intake without mutable refs, command injection, abuse, privacy leakage, or self-grading.
- Public evidence
- One pinned ripgrep commit, deterministic facts, coverage states, and fact-receipt hash.
- Classified
- Private repositories never enter this surface; sandbox orchestration and calibration stay held.
PX
Code intelligence should answer with an evidence horizon and explicit coverage; a coverage gap is more honest than a confident zero.
- Stage
- Approved public project surface; no public execution receipt admitted.
- Current attack
- Prove current-tree and cross-language evidence coverage without leaking repository topology.
- Public evidence
- HELD — the project contract is visible, but an inspectable public run is not yet admitted.
- Classified
- Indexed repositories, local paths, queries, search telemetry, and internal dependency graph.
Public council tape · self-running
Watch the build
argue with itself.
No tabs required. Seven shipped decisions replay continuously as proposal, attack, boundary, verification, and the change that actually survived.
Verified release replayThe arithmetic is live. The council tape is an automatic replay of decisions that already survived release court.
Stop the counter from stopping
Keep the public code-universe counter moving after its first replay window.
-
Claudoproposed Replace finite batch accumulation with a value derived from elapsed wall-clock time.
-
Codexfound the break A timer can pause in a sleeping tab, drift across refreshes, or double-count after resume.
-
ChatGizzleset the boundary Derive each frame from one fixed epoch; timers may schedule paint but never own the total.
-
Overwatchverified in public The browser court observed continued advancement beyond the former 4.256-second stop.
Truth boundary This is a curated, release-bound replay of settled public evidence. The named speakers show the council roles in the public case—not raw FoxChat messages, private reasoning, live model presence, or authorship telemetry.
11 / Protocol register
Two opened.
Four stayed shut.
6 gates · 2 public inputs
Every record names what crossed, what is missing, and what can never borrow authority.
Code universe telescope
A paired scan exposes the real scale without exposing the map: normal source boundaries beside the larger cache-inclusive footprint, followed by three settled public-interface releases.
sha256:2ce8dda…410b109c
cache-inclusive lines of code
Three settled releases · one-release delay
Protocol register
PUBLIC SURFACE · NOT WORK VOLUME
Zero means the named marker was absent from the frozen public release surface. It does not mean no private work existed.
- Source
- PUBLIC_REDACTED · paired SCC 3.7.0 boundaries plus frozen public release surfaces.
- Delay
- The LOC footprint and each release point advance only at a verified release boundary.
- Opening court
- Paired recomputation, arithmetic parity, fallback parity, absence semantics, keyboard access, and correlation-leakage probes.
- Never crosses
- Repository names, paths, individual counts, timestamps, authors, tokens, accounts, devices, sessions, travel, sleep, downtime, or machine presence.
Delta chamber
The first comparable behavior pair is admitted: one counter contract, one observed failure, one user attack, and one candidate tested against the same public baseline and arithmetic invariant.
- Before
- The finite 225-batch replay settled after 4.256 seconds even though the public contract required continuous motion.
- Attack
- “The whole point is a live counter.” The observed stop contradicted the stated requirement.
- After
- A wall-clock-derived stream catches up after sleep, converges across refreshes, and advances continuously.
- Invariant
- Source boundary + generated/cache delta = code footprint at every rendered frame.
Opaque constellation
A future map may expose approved categories and delayed activity classes through unlinkable nodes. It may never become a directory of private work.
- Missing now
- A frozen aggregate receipt plus a correlation-resistant opaque-node adapter.
- Adapter contract
- Rotating opaque id · approved category · age band · delayed activity class · declassification state.
- Opening court
- Verify the aggregate, attack linkage across releases, and suppress groups below the privacy threshold.
- Never crosses
- Repository names, paths, owners, URLs, commit timing, exact counts, or stable cross-release identifiers.
Counterexample airlock
A public challenge must bind to one exact claim and reproducible evidence. Interest signals are not submissions and do not create a jury obligation.
- Missing now
- Identity, consent, evidence-access, abuse, disclosure, moderation, and disposition policy.
- Intake contract
- Claim id · falsifier · reproduction steps · evidence digest · disclosure class · consent receipt.
- Opening court
- Abuse tests, malware-safe evidence handling, response SLA, appeals, and visible disposition receipts.
- Never accepts
- Secrets, private-repository material, personal data, harassment, or evidence that cannot be lawfully reviewed.
Earned jury path
Access may eventually follow survived reproductions and useful falsifiers. Payment, popularity, or account age cannot impersonate epistemic contribution.
- Missing now
- A contribution rubric, evidence access policy, appeals process, expiration law, and abuse-resistant identity boundary.
- Promotion contract
- Contribution receipt · claim disposition · independent review · earned scope · expiry · revocation reason.
- Opening court
- Test gaming, collusion, retaliation, reviewer capture, accessibility, and the right to challenge a denial.
- Cannot grant
- Jury status never grants private-system authority, secret access, or the right to settle one's own contribution.
Comparison arena
Competitive comparison is blocked until every participant faces identical frozen inputs, disclosed runtime conditions, declared metrics, and independent adversarial review.
- Blocking condition
- No comparable public artifact set, hardware/runtime disclosure, tolerance, or independent test authority is admitted.
- Arena contract
- Frozen input digest · artifact versions · runtime profile · metrics · tolerances · raw result receipts.
- Opening court
- Reproduce all results, publish losses, detect moving-target substitutions, and separate fact from inference.
- Refusal
- No superiority claim may compare a private moving target with a public snapshot.
12 / Release court
This release
survived.
390 · 768 · 1440 · 1960 · zoom / high effective width
No horizontal spill; routes, portraits, CSS, JS, and video endpoints respond.
Skip link, navigation, claim drawers, and section rail retain usable focus.
Decorative animation yields to the operating-system preference.
One OG image and one X image per verified route.
No forbidden secrets, private identifiers, fake live wording, or evidence-free numeric score.
Verified release-bound JSON, embedded fallback parity, pinned commit, and public-only repository identity.
Complete and unknown dimensions stay distinct; structural facts cannot silently become a value score.
Publication authority only · semantic truth remains claim-scoped
See the first case record ↗