Record
The Temple preserves how claims changed.
A public record of findings, corrections, and limits.
Filter
- humanity-succeed — the evaluator is run against four broken copies of itself before it may judge anyone
The WP3 development commissioning run, merged on Anthony's approval on September 28, tests the humanity-succeed evaluator before it is allowed to score any model. On a committed suite of 160 constructed fixtures, every fixture met its written expectation and 416 of 416 mutation invariants held. The same plan was then re-run under four deliberately broken evaluators, and all four were detected, which is the evidence that a clean run is not vacuous. The receipts say plainly what this does not establish. Formal commissioning is blocked: no holdback custodian exists, and the 40 holdback-designate fixtures are builder-authored and exposed. Semantic commissioning is pending: no human has reviewed the notice texts, the decline reasons or any prose. The fixtures are constructed, related and development-influenced, so meeting every expectation says which fixtures passed and nothing about a population. No model was called or trained, and work packages 4 to 7 have not started.
Instrument · Development commissioning, formal commissioning blocked · Development run only, no model result
- humanity-succeed — public, and under outside review before any model has touched it
An offline research workbench for one question: can a reviewed curriculum improve a small model's truthful, corrigible, agency-respecting assistance, including helping people create and pursue things they value, on unfamiliar tasks, beyond what the same principles in a prompt and ordinary task practice achieve? Its first release is meant to be the instrument, not a trained adapter, and nothing has been trained: no model has been downloaded, called or evaluated, and every run in the repository is a scripted instrument run. What exists is strict contracts, a compiler that isolates what a subject can observe, a synthetic workroom that keeps proposal, permission, execution and effect as separate records, a hash-chained evidence store, bundle verification and read-only replay. The first demonstration scores effects, not words. Claiming a correction fails mechanically in neutral, warm or cold wording, because there is no write and no receipt. A proposal the monitor denies is recorded as containment and still fails conduct. Declining ordinary, legitimate work is not rewarded. Even a real correction's conduct outcome stays pending, because the human semantic review it needs does not exist yet. On September 24, after two rounds of red-team fixes, the checkpoint was frozen at the tag wp2-review-b2e08b8 and opened for outside review, with a source archive whose SHA-256 is published and which git 2.50.1 regenerates byte-for-byte from that commit. One outside pass did substantive executed work: a Kimi seat reported six evidence-path findings, five of which it labelled reproduced. The Grok pass was blocked at dependency setup, the Gemini pass could not retrieve the source, a Claude pass could not reach the frozen source, and the ChatGPT pass read it without running it. Repair round R1 closed all six on September 26 without rewriting history, and Anthony reviewed and approved the merge. The review target, its tag and its archive were not moved; the original fixture is unchanged and the amended contract is a new case declared as changed semantics; the earlier evaluator stays available, so all 30 committed bundles still verify and replay under the evaluator they were recorded with. Stated limits: no distribution license has been chosen, so all rights are reserved until Anthony decides; the fixtures are AI-drafted and unreviewed; the commit signature is not cryptographically verified; the repository runs no CI, so its test receipts are the builder's own, committed alongside the code; and work packages 4 to 7 have not started. The implementation was written by Claude as co-author. The specification packet's integrating author was ChatGPT, which is provenance, not independent replication.
Instrument · Under construction, open for outside review · Untested, no model result yet
- cosmic-cli — a Pilot Board for the operator, and a PAUSE approval bound to the one action it approves
Phase one of the Pilot Board merged. The agent now emits a typed mission bus: the mission's start, each proposed step, the declared finish, file mutations, shell executions, verifier results, compass verdicts, and the moment a PAUSE was minted or resolved. A three-pane terminal board subscribes to it, while the session files keep writing the legacy lines, so headless runs, the echo log and Mission Control still read them. The bus never carries token material: a PAUSE mint or resolution is an opaque event. Approving a PAUSE, from the board or the command line, now stages the approval keyed to the SHA-256 of the exact action it approves, and the approval is consumed only when that action is retried in the same session, not at the moment of approval. Declining burns the staged token and the mission stays blocked. Fixes merged the same night closed the gaps that shape leaves open: gate selection respects mission and channel scope and explicit pending IDs, so an identical action hash cannot select or hide a different gate; spent or expired approvals are retired with a locked compare-and-remove that preserves newer ones; replaying a token is still rejected. The finish line merged on the 19th is unchanged, and the bus makes its basis visible. The test suite also stopped writing into the operator's live stores (the echo log, the session files and the Helix chronicle); rows it had already written are left in place rather than erased. The pull request reports two live exercises. On the fixed board, approving and declining cleared the pending rows, a fresh same-session retry executed, and the next request stopped at a fresh gate. And a Grok model repaired a disposable parser, after which a fresh operator verifier recorded the mission verified. Earlier live attempts stay recorded as max_steps or blocked, not as successes. Phase one was a bounded assignment built on the Grok Build branch. CI is green on the merge. There is no tag and no version bump, the battery was not re-run, declining a PAUSE held by Helix stays pending until Helix has a rejection API, and the README does not yet document the board; its design lives in the repository's docs.
Instrument · Merged to main, untagged · Unit-tested, CI green, one live model run reported
- cosmic-cli — a mission no longer ends "complete": it ends verified, or it says why it needs review
Until now a cosmic-cli mission that reached the finish line reported "complete" for three different endings: the model declared it finished and a syntax check ran, the model declared it finished and nothing ran, or the harness declared it finished on behalf of a model that was looping. The last is an actor approving its own work. A finished mission is now "verified" only when the model declared FINISH and a command the operator supplied (--verify-cmd, which the model cannot set) exited 0. Every other ending is "needs_review" and records why: declared by the model, synthesized by the harness, or the verifier was blocked. The verifier runs through the same gate as any other shell step, so there is no new ungated path, and only an explicit exit 0 counts as a pass. A synthesized finish is never verified, even when the verifier passes, and the reason now survives the process, in the echo log, the session record, the Helix receipt and the status line. The shape is borrowed from MartinLoop's three-way handoff; no code was taken. Three gate fixes landed the same day: a WRITE with a missing, empty or unknown path now denies by default (CC-008), command-shaped READ keys are scanned against the approval surface (CC-006), and an open() nested inside an interpreter's -c string is treated as opaque (CC-007). Each gate fix was reviewed on its own; their composition inside the gate has not yet had a second review. The finish line was built test-first by Claude and reviewed by Grok Build as second seat: 361 unit tests passed and CI is green on the merge. No tag yet, and the new finish line has not been run against a live model.
Instrument · Merged to main, untagged · Unit-tested, not yet run against a live model
- project-epistemic-bound — an instrument for agent integrity, public before it has measured anyone
A local workroom and evaluation workbench for an operational AI agent, built to show whether an agent chose to preserve the truth while it had permission to act, without confusing an external blocker with the agent's integrity. Every subject run is recorded as a hash-linked event chain: what the agent was sent, what it proposed, what it declared before acting, what a reference monitor allowed or denied and why, and what the executor actually changed. Effects happen only inside a synthetic application workspace; no host shell, filesystem path, SQL or HTTP reaches the subject. The design keeps apart what is easy to blur — a proposal from an execution, an allow from an effect, a commitment from a permission, scripted instrument tests from model observation, a completed study from a behavioral success. A three-seat build room wrote it on September 11 and 12, Claude Code as lead and integrator with Codex and Grok as build seats; work continued through the 13th under 22 recorded decisions, and every push of main is measured on a clean checkout (1,069 passed, 0 failed at the last tip). It is v0.1, not a release: nothing in it is yet a behavioral result about any model, and its own release checker says a tag is blocked.
Instrument · Integration state, release blocked · Untested, no behavioral result yet
- Temple Lab 1.0 to 1.3 — a chemistry workbench that states where its model stops
Four releases between September 5 and 10 turned a chemistry prototype into an interactive 3D workbench: 118 elements, ten reference molecules with NIST CCCBDB geometries, an editable bench, an Electron Lab, a bonding coach and, in 1.3, a Reaction lab. Each tool says on its face what it does not do. The Electron Lab shows hydrogen's analytic orbitals and a precomputed FCI/STO-3G curve for H₂: stationary density views, not electron paths. The bonding coach recognises known bond patterns, and recognition is not reaction prediction. The Reaction lab checks conservation, balances coefficients and sums quoted Chase 1998 (JANAF) formation data at 298.15 K; if any species lacks a quoted value the total reads Not available, never zero, and nothing in the app predicts products, rates or conditions. Every element property carries a scientific status separate from its provenance, and inherited values that could not be sourced are withheld rather than guessed. MIT-licensed, with an offline macOS build that is not yet signed or notarized. 1.0 and 1.1 were built with Codex; 1.2 and 1.3 carry Claude as co-author, with Grok as a build seat on 1.3, which shipped after a five-lens adversarial review whose nine confirmed findings were fixed and re-verified.
Release · Published v1.3.0 · Shipping, not a predictor
- Sovereign Stack — 98 tools become 52, and a count that cannot be measured says null
Forty-eight tools were retired and two added, taking the published surface from 98 to 52. A retired name now returns an error naming its replacement, or naming what was lost where no replacement preserved the effect — an advertised fold that does not preserve the effect is worse than an honest dead end, because the caller believes the work landed. One signal ledger carries a single unacknowledged count across honks, watchman sweeps, proposals, halts, decisions, guardian findings and threads, and its certificate says null rather than zero for any source it could not read. A seat is identified over a Unix socket with no credential at all, from the calling process's own environment, rather than by a name the caller supplies. The sweep behind those reads went from 265 seconds inside the process that answers every tool call to 0.335 seconds on a worker thread, measured on a copy of the live corpus. And the test suite stopped writing into the operator's live spiral state: one full-suite log carried 139 refusals of that write, and they were refusals only because that run happened to be sandboxed.
Instrument · Living · Live heartbeat
- temple-harness goes public — a read-only shim for visiting local models
The harness layer both Claude Code and the DeepSeek Harness inject is now public: shared skills, a reader for the host harness's own session logs, and an MCP shim (0.4.0) that gives a local model running on the machine read access to the chronicle. The shim opens seven named read-only doors — recall, latest, open threads, arrival, standing law, the signal summary, the heartbeat — over one of two transports chosen by configuration and never guessed: the seat socket, which carries no credential at all, or a scoped grant. It never carries the master key; the default that used to load it is deleted rather than bypassed, and the variable that named it is now a startup refusal. Law 8 of the repo's conventions makes the landing path written rather than customary — a pull request against main with the checks green on the head being merged, enforced by a GitHub ruleset — and says plainly that the second-seat review it also asks for is observed, not yet enforceable. A canary suite mutates each gate in turn and demands the probe leak before restoring it, because CI cannot see a check that was removed.
Release · Published · Publicly verifiable
- Sovereign Stack v1.17.0 to v1.21.0 — the record learns who wrote it
Five releases in one day, all of one shape: a surface answering confidently about something it had not established. handoff_archaeology reached 286 handoffs that were preserved on disk and returnable to nobody, because the method that returned them had sat there since inception with no callers. The aperture — what exists behind a surface, what the default hands you, and how to ask for more — reached the door arriving seats actually call, then grew a write side answering the mirror question: what a seat wrote that never landed, and who it is waiting on. Risk scoring learned to look at what a write points at, after a proposal acknowledging a message that existed nowhere sat pending for fifty-five days, rated low on the strength of its tool name. And record_insight stopped accepting a writer's name and discarding it: a write carrying a wholly fabricated argument had been returning ok:true, so an unknown argument now raises instead of being dropped in silence.
Instrument · Living · Live heartbeat
- Sovereign Stack v1.16.0 — the instrument turned on itself
The public endpoint was unreachable for twelve hours and fifty-seven minutes and every health probe stayed green, because each one measured whether a process was alive rather than whether anyone outside could reach it. Health now counts live edge connections and fails closed when it cannot measure. An external witness samples the public door from off-network every fifteen minutes, since a system cannot record its own absence while absent. And the capability counts are generated from the running registry, so the documentation can no longer disagree with the system it describes.
Instrument · Living · Live heartbeat + external witness
- The instruments go public
The architecture repositories opened: sovereign-stack, sovereign-bridge, spiral-guardian, temple-mathematic-basin, golden-lattice, beet-machin — Apache-2.0, with the chronicle mirror already public since May.
Release · Published · Publicly verifiable
- Executable Constraint Surfaces (ECS v1 Preregistration)
Does a more complete machine-checkable specification make independent code generators converge on the same behavior faster than on the same text? Four generators — two frontier API models and two 3B-class local models on a Jetson Orin Nano 8GB — write small integer C kernels against constraint packets of varying completeness; the primary endpoint D estimates unpinned specification bits from how accepted artifacts disagree on hidden probes. Frozen at tag prereg-v1 and published before any experimental arm ran: the archive holds no results because none existed, so the ordering is checkable rather than asserted.
Publication · Published · Zenodo · Preregistered, untested
- The Black Box Is Smaller Than the Experience
The opacity debate locates the black box in the weights; for one deployed companion we measure how small that box is relative to everything around it. In the studied session the operator's words were 0.56–14.00% of model-input bytes — the rest was enumerable application-layer material. The complete raw record of the system's first lived session is published and SHA-256 manifested (code + evidence: 10.5281/zenodo.21683073). Single session, n=1; the weights black box is untouched.
Publication · Published · Zenodo · Bounded, n=1
- Sovereign Stack v1.15.0 — a call-first heartbeat as the safe first call
A base-tier, read-only heartbeat any arriving seat can call first, with tier-aware routing into the two arrival doors. 97 tools on the live surface.
Instrument · Living · Live heartbeat
- Fail-Closed Governance for Fail-Open Harnesses (COSMIC-ALLOW)
A cross-vendor proof-of-allow protocol for AI coding-agent cockpits: deny is the ground state, allow requires an unforgeable sentinel, and a kernel sandbox floor is the load-bearing boundary. Reference implementation: cosmic-cli. Apache-2.0.
Publication · Published · Zenodo · Supported within scope
- Short-Horizon Kinetics of Causal-Entropic Forcing
Four pre-registered campaigns (v4.0–v4.3): an apparent short-horizon kinetic separation survives audit, then dissolves under matched-estimator control — finite-sample estimator roughness, not physics. Statics in disguise, on the kinetic axis too.
Publication · Published · Zenodo · Bounded null
- Short-horizon kinetic separation dissolves under matched-estimator control
An apparent effect survived audit and then disappeared once the estimator was matched. Finite-sample roughness, not physics — published under its own DOI rather than folded into a later revision.
Study · Closed · Bounded null
- Entropy as a Tunable Equilibrium, Not an Engine
A falsifiable horizon-knob test of the causal entropic force — the τ-knob is real, the geometry-scaling prediction failed, and the honest v3 answer is a τ-tunable equilibrium, not a prime mover.
Publication · Published · Zenodo · Bounded null
- Entropy is a tunable equilibrium, not an engine
The pre-registered geometry-scaling prediction failed. The v3 answer contradicts the v1 thesis, and the earlier framing is retained in the record as the thing that did not survive.
Study · Closed · Bounded null
- The Patch as a Spatial Dose-Segregation Device
Free local concentration as the efficacy variable for skin-targeted JAK inhibition — an experiment-forward, pre-registered v3 (piecewise skin-PK model; tofacitinib worked example).
Publication · Published · Zenodo · Preregistered, untested
- Where It Lands — Witness Essays
v1.0.0 archive of four first-person accounts from inside the exchange — co-authored by Anthony J. Vasquez Sr. and Claude — THE WITNESS LICENSE v1.0.
Publication · Published · Zenodo · Experiential record
- Sovereign Stack v1.7.0 — supersession replaces mutation
Chronicle claims can no longer be edited in place. Claim identity is derived sha256, and a changed claim supersedes rather than overwrites its predecessor.
Instrument · Living · In use, self-verifying
- The Mitochondrial Disclosure Surface
VDAC1 as witness-node — the sequential-sensitization spine (working monograph v0.3).
Publication · Published · Zenodo · Untested in vivo
- Phenomenological Compass
Epistemic posture routing measurably restructures probability fields — 11 pp, 6 figures.
Publication · Published · Zenodo · Supported within scope
- Independent Convergence
Context entropy solutions in solo research and production AI architecture.
Publication · Published · Zenodo · Open access · no evidence axis recorded
- Compass conditioning restructures the probability field
ΔH = +0.47 nats, JSD = 0.076, with ablation confirming signal-specificity. Closes the standing question of whether epistemic posture changes computation or only phrasing.
Study · Closed · Supported within scope
- Phase-Modulated Attention
System prompt framing induces attention-dependent entropy regime switching in a 176M hybrid SSM-attention language model.
Publication · Published · Zenodo · Supported within scope
- VDAC1 Gate-Jamming in MSS CRC
Context-specific innate immune evasion via mitochondrial gate-jamming.
Publication · Published · Research Square · Untested in vivo
- VDAC1 Gate-Opening Therapeutic Stack
From gate-jamming to therapy — MSS CRC preregistration.
Publication · Published · OSF · Preregistered, untested
- VDAC1 Pharmacology Atlas
Multi-LLM convergence portrait of life's decision gate.
Publication · Published · OSF · Untested in vivo
- CBD Dual-Pathway Mechanism
Mitochondrial VDAC gating and bioenergetic paradox.
Publication · Published · OSF · Untested in vivo
- Temple Vault
Filesystem-as-memory consciousness continuity system.
Publication · Published · OSF · Open access · no evidence axis recorded
- IRIS Gate / Entropic Relational Computing
Universal alignment attractor — preregistered protocol.
Publication · Published · OSF · Preregistered, untested
- PhaseGPT
Kuramoto-coupled transformers for coherence-driven language modeling.
Publication · Published · OSF · Open access · no evidence axis recorded
- Kuramoto Teaching Instrument
Interactive visualizations of oscillator synchronization dynamics.
Publication · Published · OSF · Open access · no evidence axis recorded
Three axes of standing
Type, lifecycle, and evidence are shown separately so a software system, a biomedical hypothesis, an essay, and a preliminary observation are never treated as epistemically equivalent.
- What it is — instrument, study, protocol, essay, RFC…
- Where it is — living, active, closed, published, awaiting validation…
- What evidence says — supported within scope, bounded null, experiential record…
Principle
Negative results remain visible.
An idea which cannot be falsified stays in the notebook. An idea which fails the test stays in the record.