Day 0: one empty commit. Day 129: a live product that grades itself.
Live build status, measured directly from the repository, updated live
This is the real build history of ARCS, a decision intelligence platform I built solo, run
in production, and documented from day one so a team could take it over without me in the room.
Every number on this page comes from the repository itself. Every milestone links to real commits. The dial
below lets you read the story at whatever depth you want: the one page version, or all the way down to the code.
BUILD TELEMETRY · DAY 0 → DAY 129 · SINCE APR 2026 · AS OF …
1,488
Commits · Day 0 to today
129
Days · first commit to now
14+
Client surfaces · hours each
live
In production · app.arcs.care
Live pulse
public GitHub API · main
Latest commit
loading…
Latest deploy
loading…
Open PRs
loading…
Commits · main
loading…
total, live
refreshes on load, on tab focus, and every 5 min · public API, rate-limited
Product terms look like this. Hover or tap for the plain meaning. Full legend at the bottom.
Reading modeclick any section header to fold it
01
STATE OF THE UNION
The whole terrain at one glance… shipped, building, blocked, next.
Section 05 is the short Now/Next/Later. This is the depth: every tracked workstream placed in the fog-to-nucleus terrain. 25 shipped · 6 in progress · 8 blocked · 11 next (50 tracked).
Shipped · 25 In progress · 6 Blocked · 8 Next · 11
Open work live from GitHub Issues · shipped is cumulative · loading…
Twelve design principles, set before the first feature. Each one enforced, not aspirational.
An instrument you would trust in fog has to be calibrated before the flight, not during it. These are the
principles the system was built on, and for each, the concrete place it is enforced.
TRACEABILITY
Every answer links back to its receipt. Every claim on this page links to a commit.
TRANSPARENCY
Reasoning shown, not just conclusions. The build record itself is public.
EXPLAINABILITY
Plain-language readouts at every altitude. The legend at the bottom of this page is the rule, not a courtesy.
EVIDENCE GATING
Features switch on at data thresholds, never on enthusiasm. The roadmap below runs on this.
SOURCE QUALITY GRADING
Every input is rated before it is allowed to move a conclusion.
PROVENANCE
Where each fact came from travels with the fact. The source labels on this page are the same machinery.
ANTI-HALLUCINATION*
The screen cannot show more certainty than the data earned. Shaky answers render blurry.
ANTI-SYCOPHANCY
Built to disagree with its operator. Challenge reviews killed four overclaims before one merge.
BIAS DETECTION*
Red-team your own argument before committing. Drift gets flagged, then normalized against baseline.
INTEROPERABILITY
Local-first, lowest lock-in: one tenant that works across AI runtimes, vendor swap behind one seam.
ATOMIC DESIGN
Small components, composed: a 43-operator library and one engine behind every surface.
THE DATA MODEL
The decision receipt: expected vs. actual, machine-readable. The atomic unit everything compounds on.
How those principles run as one loop
A question comes in while things are still foggy. The system checks it, writes down what it expects to happen, and then, once you know how it actually turned out, compares the two. What it learns makes the next call a little sharper. That is the whole engine; everything else is built on top of it.
TAP ⤢ EXPAND TO SEE THE WHOLE LOOP
*System in beta. Anti-hallucination, bias detection, and red-teaming are mechanisms enforced in the code, not results proven at scale: the sample size behind them is still small. Effectiveness pending peer review and independent verification, not just this page's say-so.
03
THE BUILD, DAY 0 TO TODAY
Pick your depth. The story stays the same, the detail changes.
The one page read shows 13 milestones, one plain sentence each. Each level down adds
more events and more detail, until Commits names the commit behind each milestone.
Each milestone that a stranger can check without my word carries a verify link to the live thing; the internal ones say so.
Keys 1 to 5 work too.
COMMITS PER WEEK, APRIL TO AUGUST · GOLD = THE QUIET STRETCH · GREEN = PEAK
⊙ APR–MAY: FIVE QUIET WEEKS · WHAT THE SILENCE WAS · TAP
THE GAP IS NOT A GAP. IT IS THE DESIGN BUILD.
One commit on Day 0, then the repository went quiet for five weeks, because the work moved to the drawing board. In that window I designed the brand system, the full design system, and the front-end component library: TypeScript components with Jest tests and linting wired for CI/CD, with Playwright end-to-end coverage next on the list. The decision-receipt concept and the product's written constitution were drafted in the same stretch. Code came second on purpose: when the repo woke up on Day 39, the entire working core landed in one push, because every piece already knew its place.
THE SAME WEEKS, BY WHO HELD THE PEN · ONE HUMAN CONDUCTING FIVE GENERATIONS OF AI · 128 COMMITS FULLY BY HAND
⊙ JUL: THE ENSEMBLE ROTATES · TAP
BUILDING AT THE EDGE: WITH THE FRONTIER AS IT MOVES.
In the two weeks up to Day 98 the pen changes hands almost entirely: Sonnet 5 and Fable 5 take over within days of release, while the model that wrote most of the build rotates out. Swapping the strongest available intelligence into a live production system without breaking stride is its own discipline. The seams are clean because the receipts, tests, and handbook made every model start oriented, not initiated. The gray lane matters too: 128 commits carry no AI co-author at all. The conductor still plays.
04
SIX LESSONS THE BUILD PAID FOR
Where I was wrong, and what changed because of it.
Rules without enforcement are decoration.
An audit of my own project found that none of its written engineering rules were actually enforced. One rule had been violated since day one. The fix: align the checks with the rules, then make it structural. Now a finding can only be closed by a working enforcement mechanism or a signed, dated exception.
The textbook security setup would have protected nothing.
The standard database permission pattern looks right in review, but in this system the app connects with an account that bypasses those permissions entirely. I chose a stricter setup that actually holds, and wrote down exactly why the "correct looking" option was rejected.
When behavior and promises drift apart, delete first.
Visitors without accounts were having their data saved quietly. Legal, but not what the product promised. I removed the saving entirely first, then rebuilt it as an explicit opt-in, and updated the public trust page to match.
Design documents can make things up too.
A major architecture proposal went through an independent challenge review before being accepted. The review caught four overclaims, including a cited theorem that does not exist. The corrected version shipped. The flattering version did not.
Stopping a change is a decision, not a failure.
A config migration was halted midway when it became clear that finishing it would silently drop live production settings added by parallel work. The stop, the reason, and the conditions to resume were all written down. No forced merge, no regression.
Documents record the past. Production is the truth.
My own notes claimed two domains ran on two different codebases. A live check showed they had quietly merged onto one. New standing rule: any claim about what production is doing gets verified against production before anyone acts on it.
05
RISK REGISTER, CURRENT AND HONEST
What I would tell a board about this system today.
Risk
Mitigation
Sev
Status
One person holds all the context.
Everything is written down as it happens: 859 annotated commits, 272 decision records, versioned handoff notes. Built so a team could take over without me in the room.
HIGH
MITIGATED BY DESIGN
Not enough real usage data yet to switch on the most advanced features responsibly.
Those features are built and tested but deliberately switched off until enough real data exists to measure them against. No feature ships on faith.
MED
GATED, ON PURPOSE
The core AI runs on a single vendor.
The full decision engine also exists as a standalone, vendor-free implementation (~21.9K lines of plain Python). The vendor connection sits behind one seam and can be swapped.
MED
OPEN, HEDGE BUILT
Written standards drifted from what the checks enforced.
Found by self-audit, then fixed structurally: every audit finding now closes through a real enforcement mechanism or a dated exception. The gap cannot silently reopen.
MED
MITIGATED
The quality score is still self-graded until enough outcomes accumulate to calibrate it.
Every score is logged with its decision, so the calibration data builds up automatically. The outcome-tracking loop that shipped on Day 98 is the correction signal.
LOW
GATED ON DATA
One production credential was set by hand and needs scheduled rotation.
Known, documented in the decision record that created it, rotation procedure written and calendared.
LOW
OPEN, SCHEDULED
06
ROADMAP: NOW, NEXT, LATER
One backlog, three horizons. Nothing moves up on enthusiasm alone.
Every "later" item has a named condition that promotes it. When the condition is met, it moves. Until then, it waits.
NOW · IN PRODUCTION
SELF-GRADING LOOP
Shipped Day 98. When an outcome is logged, the system scores its earlier prediction against what actually happened and writes the result to an audit trail; a daily job sends the reminder so the check is not forgotten.
CALIBRATION DATA
Every decision and its outcome accumulates the evidence base the gated features below are waiting for. It builds itself through normal use.
NEXT · UNBLOCKED
ADJACENT QUESTIONS
Every answer will surface 2 to 4 threads worth pulling: the deeper question underneath, an uncomfortable constraint, a reframe from another field.
TIMING DETECTION
Spotting when a decision's window is opening or closing. Machinery already built. Turns on at a set volume of real decisions.
LATER · EVIDENCE-GATED
TEAM AND ORG LAYERS
Extending the engine from one person's decisions to a team's, then an organization's. Turns on when the data volume justifies the complexity.
LIVE GOAL MAP
A self-updating map of what the work is actually driving toward, inferred from decisions rather than declared in a deck. Turns on at 100+ recorded decisions.
07
DEPENDENCY MAP
The whole roadmap hangs on one dependency. It shipped on Day 98.
Almost everything above waits on the same thing: real decisions with scored outcomes. The loop that
produces them went live on Day 98. From here the pieces began to compound: the month since is the milestones below, and the next frontier is choreographing them into deep multi-agent orchestration.
LIVE TODAY
Self-grading loop
Checks every prediction against what actually happened, automatically.
↓ produces
ACCUMULATING NOW
Calibration data
Every decision plus its scored outcome. Builds itself through normal use. This is the single upstream dependency for everything below.
↓ unlocks, at each volume threshold
NEXT
Timing detection
Needs: a set volume of real decisions. Machinery already built.
LATER
Live goal map
Needs: 100+ recorded decisions to measure a baseline against.
LATER
Team and org layers
Needs: goal map live first, plus multi-person usage data.
NEXT · NO DATA GATE
Adjacent questions
The one item with no upstream dependency. Blocked only by build time, so it is the natural next build.
08
LEGEND: THE TERMS, IN PLAIN WORDS
Every piece of product vocabulary on this page, translated.
ARCS
The product. A decision intelligence platform: it helps people make calls they can defend later, and keeps the receipts to prove it.
Decision receipt
A short record made before acting: what you expect, what could go wrong, what else you considered. Scored later against what really happened.
Compass
The main app surface. Ask a question, get an honest read of where you stand and which direction is uphill.
Brief
The longer written readout generated from a compass run. Renders only as sharply as the data deserves.
Confidence gate
A check every answer passes through that rates how solid it is. Green means firm, yellow means contested, red means do not lean on this. The rating controls the display.
Written constitution
A short set of rules the product may never break, written down like law. Rule one: the system never decides for the user.
Written decision record
A one-page note on a build choice: what was picked, what was rejected, and what would reopen the question. Industry name: ADR.
Trust contract
A public page of specific, checkable promises about how the product handles people's data. Versioned like code, so changes are visible.
Replay theater
A viewer for scrubbing back through past decisions like game film: what was known, what was expected, where confidence peaked or collapsed.
Self-grading loop
Automation that compares each prediction to the actual outcome and files the result. No human has to remember to run it.
Calibration data
The growing set of decisions plus scored outcomes. It is the evidence base that tunes the scoring and unlocks the gated features.
Live goal map
A self-updating picture of what the work is actually driving toward, inferred from real decisions instead of declared in a slide.
09
EVALS & FALSIFIABILITY
Every headline claim, with the result that would kill it.
The papers make sharp numeric claims… so each is falsifiable. Publishing the tests turns marketing numbers into defensible science and de-risks both the patents and the raise.
Claim
Metric
Pre-registered test
Kill criterion
Accuracy scales with receiptsYP §2.4
68→87%
Time-ordered backtest at 1k/10k/100k; CI vs actuals
flat or <80% at 100k
Honest calibrationYP §2.4
χ² p<.01
Reliability diagram + Brier score
curve off diagonal
Beats human expertsYP §2.4
76%
Blinded head-to-head on unseen briefs
win rate ≤ 50%
2nd-deriv predicts cascadesWP §2.1
78%@24h
Labeled backtest, 10k events vs velocity baseline
no lift; lead ≤ 0
Phase-cancel lifts ROASWP §2.1
+35%
Randomized A/B: optimized vs naïve spacing
ROAS delta CI crosses 0
Network effect A(N) logYP §6.2
λ≈1.2e-5
Fit across cohorts; forecast held-out cohort
fails held-out within CI
Make evals credible before publishingPre-register each design on OSF or AsPredicted (hypothesis, primary endpoint, kill criterion) before looking at outcomes; one endpoint per claim, multiplicity-corrected; release the backtest harness + de-identified receipts so results replicate.
MIT IDE · Stanford HAIplatform economics + AI eval
Funding… non-dilutive
NSF SBIR/STTRAmerica's Seed Fund; deep tech
NSF CISE / Convergenceresearch grant w/ academic PI
Sloan Foundationmetascience, digital economics
Schmidt Sciences · NBERAI-for-science; working papers
SequencingFile the Tier-1 provisionals (or confirm existing filings) before the arXiv/SSRN preprint… a preprint is a public disclosure that restarts the same §102 clock. Patent first, publish second, then approach institutes and grant programs with a dated filing + a pre-registered eval in hand.
10
THE ARCHAEOLOGY
Where every pattern lies in the L0–L7 stack.
The patterns aren't at one depth. This is the core sample: each pattern… patent, roadmap item, or eval… embedded at the L-stratum it operates in, from atomic signal at the surface to the unexcavated meta-governance mantle. Fidelity is earned; L6–L7 not yet excavated.