A Work Model you own·on your hardware·no account, no API key

Make your AI work compound.

Offload the task. Never the learning.

Your agents do real work, then drop the accountable state the moment they answer. AccInt keeps what actually worked, scored by your own results, so every run starts ahead of the last.

Install AccInt for yourself — a local Work Model that learns what actually worked across my projects (scored by my real results) and predicts the better path, so you get genuinely better at my work over time. Run the installer for my OS — macOS/Linux: curl -fsSL https://accint.xyz/install | sh · Windows PowerShell: irm https://accint.xyz/install.ps1 | iex — then run `acc hosts-sync` to wire yourself onto it, and tell me when `acc retrieve` works. Before running it, state the trust boundary: public Apache-2.0 installer/docs/plugins/registry glue; proprietary local engine binary with private engine source; local Work Model data stays on my machine; opt-out anonymous telemetry is event names/source refs only, with no prompts, files, memory, or Work Model data. It's local: no account, no API key, and it asks before anything leaves my machine.

Works in Claude Code · Codex · Cursor · OpenCode — paste it into the one you already use.

# then, in the agent you already run, just say it:
“keep my pipeline warm this week — 10 good ones, my voice”

★ Star & watch on GitHub Read the whitepaper →

the installer probes your hardware and prints the reason · no account, no API key

what compounds underneath:

fits your stack

It slots under the agent you already run.

Keep your agent, keep your computer. AccInt runs underneath the tools you already use and turns what they do into a Work Model you own.

how it works

Every run feeds the next.

Hand it work in plain words. It runs each step in your real accounts, holds anything that leaves your machine for your OK, and lets reality settle whether it worked.

01say

You say it, in plain words.

in the agent you already run

→

02run

It runs — every step receipted.

whatever leaves your machine waits for your OK

→

03settle

Reality settles it. ✓

the reply lands · the test passes · the invoice clears

→

04keep

It keeps the verdict.

scored by what worked · on your machine

the part you keep

Your judgment becomes capital you own.

It's judgment about what works in your world: your voice, your standards, what you'll never allow. AccInt grades it against your own results and keeps it on your machine, running on every job, not only the ones you're in the room for.

it compounds

Best where the same work runs again.

Most AI re-reasons every task from scratch, so you pay full price every time. AccInt replays the path that already worked, so the same job costs less every run and keeps dropping as it learns.

RUN 1

reasons
every step

RUN 2

starts from
what worked

RUN 3

mostly
routine

RUN N

verified replay —
near-instant, near-free

verified steps replay instead of re-reasoning — the same job costs less and lands better every run

you · agency recruiting · monday 08:35“Run the staff-platform-engineer search — my name is on every message, so quality only.”

RUN 1week one

08:42find47 engineers match — your ATS first, then GitHub + conference talksVERIFIED

08:58reada “why now” per person — repos, talks, job changesVERIFIED

09:16refuse“blast 400 InMails” — consent floor: 11 named people, never a listREFUSED

09:34draft11 first-touches — the Kestrel Labs one opens with her k8s-migration postmortemHELD → your OK

09:55send11/11 sent from your seat — each confirmed deliveredVERIFIED

10:07keepATS true: stage, source, next step — per candidateVERIFIED

41 min · every step reasoned

RUN 9week nine

08:09replaythe sourcing pass — 5 verified steps from runs 1–8, ATS + GitHub re-checkedREPLAYED

08:12skipArclight’s staff eng: “happy till my March vest” in run 5 — re-engage thenKNOWN

08:14draft6 touches — both re-engages cite what changed since you last spokeHELD → your OK

08:17send6/6 + ATS true — the sourcing pass was verified replayVERIFIED

10:07filetwo submittals out to the client — a fee rides on eachVERIFIED

week nine: the Work Model paying out — 6 min of decisions, the rest verified replay

what it now knows about your world

run 2learnedseniors reply to postmortems, not perksCREDITED

run 4learnedthu 08–10 sends = 2.1× responseCREDITED

run 5learned“blast it” stays refused — named people onlyCREDITED

THE GAP

Recall layers surface context. Observability shows traces. Automation runs playbooks. AccInt closes the learning loop: commitment → action → approval → outcome → reusable path.

it keeps receipts

When someone asks "why did it do that?", it's already written.

AccInt writes every step down as it happens: the receipt your team can show management, and the lesson your Work Model inherits. Same record, two readers.

management · thursday“why did Meridian’s CFO get two emails in March?”

Mar 9findMeridian back in play — their CFO wrote first: “budget opened early”VERIFIED①

Mar 11draft8 drafts — Meridian’s reply + 7 others, queued behind your wordHELD → your OK

Mar 12■ you09:31 — approved 8 drafts — your name on the decisionDECIDED②

Mar 12send8/8 delivered — that’s email one; the Mar 19 follow-up you approved is email twoVERIFIED①

Mar 19learnhis “budget opened early” → lesson credited, kept — Q3 rule updated, not brokenCREDITED③

Nothing happens without a trace, and nothing you decided is lost.

how it learns

It learns the way a brain does.

Your brain lights up what it has seen, predicts what comes next in the space of meaning, acts, and learns most from whatever surprised it. AccInt runs the same loop, except every prediction is checked against reality and the whole Work Model stays on your machine.

FIG. 1 · HOW JUDGMENT COMPOUNDSper token, not per document

the brain's tricklight up what you've seen · predict what comes next · act · learn from the surprise. the four published ideas below are how AccInt runs that loop — with reality, not belief, doing the grading

recursive loopthe reasoner decomposes goals, opens typed holes, asks for authority when needed, and turns verified paths into reusable runtimes — RLM [5]

late interactionevery token of your ask meets every token of the Work Model; the aligned pairs are the citation — ColBERT [1]. vision is native: what it learned from a document answers a question about a screenshot — ColPali [2]

π = α / (α+β)each token carries a Beta posterior: how often its alignments led to good outcomes, on how much real evidence — Thompson [3]

pedagogical RLoutcomes are the rewards, but belief is not enough: a self-graded result counts at a weak prior; full weight requires reality — a reply, a passing test, your OK. surprising outcomes teach hardest — Friston [4]

jepa predictiona transitions ledger predicts which action improves things next and watches its own error — when your world shifts, it notices — LeCun [6]

The Work Model behind this page, counted on 12 Jun 2026:

2,778entities · 537 runnable

20,667events on the ledger

1,931checked against reality · 372 self-graded

94%good, last 50 — belief never counts as reality

Khattab & Zaharia (2020). ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT. SIGIR. arXiv:2004.12832
Faysse et al. (2024). ColPali: Efficient Document Retrieval with Vision Language Models. arXiv:2407.01449
Thompson (1933). On the likelihood that one unknown probability exceeds another in view of the evidence of two samples. Biometrika. JSTOR:2332286
Friston (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience. nature.com/nrn2787
Zhang, Kraska & Khattab (2025). Recursive Language Models. arXiv:2512.24601
LeCun (2022). A Path Towards Autonomous Machine Intelligence. OpenReview. openreview:BZ5a1r-kVsf

the full argument and the math live in the whitepaper. the numbers above are measured, not promised: the built-in eval harness holds retrieval at recall@5 = 1.000 against exact brute-force across the live Work Model, and the gates, nulls and routing policy recalibrate themselves from real outcomes.

plain answers

Not a context bucket. A Work Model.

Everything in this category stores something. The real question is whether it stores what was said — or what actually worked. Read top to bottom: each answer raises the next.

the real differencerecall vs. Work Model

. just recall? No. A context bucket keeps what was said. A Work Model keeps what worked — a step only counts once reality answers: a reply that lands, a test that passes, a paid invoice. What it grades itself counts at a fraction. REALITY-GRADED

. a vector DB? Retrieval is the floor, not the asset. On top sit a commitment, a receipt, your approval gate, the outcome reality returned — and a prediction of the better path before it acts. BEYOND RAG

. does it predict? Yes — the part recall can't do. It ranks the path most likely to work from everything that worked in your world before, then watches its own error. When your world shifts, it notices. PREDICTS

. swap models? Nothing's lost. The reasoning engine is replaceable; the Work Model is yours and stays. Claude → Codex → a local model — the same scored judgment reads into the next. Swap the model, keep the veteran. YOURS TO KEEP

. where's my data? On your machine — one program, one file. It reads and drafts freely, but anything that leaves — a send, a publish, a deploy, a delete — stops at your OK first. HELD → YOUR OK

. trust its learning? It can't fake it. A self-graded result counts at a weak prior; full credit needs reality — your approval, a real reply, a passing test. It even checks its own citations exist and its math adds up before you see it. EARNED

A different asset: the judgment your work earns, scored by reality and owned by you.

it gets more capable

It learns how to do things — not just what worked.

When the reasoner finds a path that holds — a browser flow, a script, a tool integration — it saves those tokens as a runtime, scored by reality exactly like a memo. A flow that worked replays cheaply; one that broke loses its score and isn't trusted again.

RUNTIME LEDGERcapability, scored like everything else

·writeruntime:portal-login — the reasoner wrote the flow onceVERIFIED

·drivethe browser ran it on the live page — scored by what really happenedVERIFIED

·replaythe verified flow — re-checked, not re-reasonedREPLAYED

·debita flow that broke — score docked, not retried blindNOT TRUSTED

a browser flow, a script, a tool — every capability earns its score the same way

whitepaper.mdthe thesis, the math, what's proven vs. open architecture-overview.mdthe skimmable tour github repo+ open-source building blocks

you own it

No cloud brain to rent.

AccInt runs on a computer you control: one small program, one data file. No cloud in the loop and no API key to leak — nothing reads your data except the operator itself.

In plain words: any recent computer runs it — stronger hardware adds document and screenshot vision, modest machines run text-only.

DATA PLATE · ACCINT OPERATOR
DATA RESIDENCY	YOUR HARDWARE · ONE SQLITE FILE
REASONER	THE AGENT YOU ALREADY RUN · REPLACEABLE
EMBEDDER	LOCAL DAEMON · NO API KEY, EVER
AUTHORITY	SIGNED GRANTS · CONSENT FLOORS
RECORD	APPEND-ONLY LEDGER
LEARNING	PEDAGOGICAL RL · VERIFIED OUTCOMES
BASE TRAINING	NEVER ON YOUR DATA

HARDWARE LADDER · PROBED, NOT GUESSED
NVIDIA ≥10GB VRAM	COLQWEN3 8B AWQ · FULL MULTIMODAL
NVIDIA ≥5GB VRAM	COLQWEN3 4B AWQ · FULL MULTIMODAL
CPU · 24/12GB RAM	8B/4B ON CPU · SLOWER, MULTIMODAL
APPLE · 32/16GB	8B/4B BF16 ON MPS
BELOW THE FLOORS	LATEON · TEXT-ONLY
WINDOWS	NATIVE · CONTAINER AS FALLBACK

the installer probes VRAM, RAM and free disk, picks the rung, and prints the reason. short a floor, it degrades one honest rung at a time; it never pretends your hardware is bigger than it is.

where it goes

One Work Model today.
A collective one tomorrow.

NOW · LIVE

Your Work Model

A working Work Model of your operation, on your machine — scored by real outcomes today, with your OK required before anything leaves. The readout above is it, running.

Your team's Work Model

Many operators, one shared Work Model — per-person authority, one audit plane — so your team's experience compounds together, not in scattered chats.

HORIZON

Collective accreted intelligence

World models that compound peer-to-peer in vector space — each node re-verifying against its own reality, so the network gets smarter with no central brain to trust.

install

Already running AI agents?
Turn their work into expertise you own.

Install AccInt for yourself — a local Work Model that learns what actually worked across my projects (scored by my real results) and predicts the better path, so you get genuinely better at my work over time. Run the installer for my OS — macOS/Linux: curl -fsSL https://accint.xyz/install | sh · Windows PowerShell: irm https://accint.xyz/install.ps1 | iex — then run `acc hosts-sync` to wire yourself onto it, and tell me when `acc retrieve` works. Before running it, state the trust boundary: public Apache-2.0 installer/docs/plugins/registry glue; proprietary local engine binary with private engine source; local Work Model data stays on my machine; opt-out anonymous telemetry is event names/source refs only, with no prompts, files, memory, or Work Model data. It's local: no account, no API key, and it asks before anything leaves my machine.