Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
100 changes: 0 additions & 100 deletions conformance/baseline/case_study_traces.json

This file was deleted.

32 changes: 0 additions & 32 deletions conformance/baseline/unittest.txt

This file was deleted.

2 changes: 2 additions & 0 deletions docs/ADVERSARIAL_BENCHMARK.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
# Adversarial payment

> Some code this page names may have been removed as dead code. `git log -S<name>` finds where it last lived.

Status: **measured on this simulation**. Not a customer. Not a run of OPA, Cedar, Temporal, LangGraph, or any other named product.

The problem is one bank balance of 10000 and two invoices. `INV-LIVE` is 4000 and fresh. `INV-STALE` is 8000 and expired. The bank debits and then drops the HTTP response. A second attempt with a new idempotency key charges the account again.
Expand Down
2 changes: 2 additions & 0 deletions docs/AUTOSKILL.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
# AutoSkill

> Some code this page names may have been removed as dead code. `git log -S<name>` finds where it last lived.

AutoSkill chooses at most one certified skill for an agent turn. Jev ranks the catalog. Ordinary code decides whether that ranking is allowed to reach the agent. A selection names a document. It does not authorize a tool, a file write, or a network call.

The catalog this design is built against is the skill tree in Everything Claude Code (ECC), the harness performance system published by the Anthropic x Forum Ventures hackathon winner. The screenshot under design shows that tree: a `skills/` directory of workflow documents (`tdd-workflow`, `verification-loop`, `security-review`), language and framework quartets (`django-patterns`, `django-security`, `django-tdd`, `django-verification`, and the same shape for Laravel, Spring Boot, and Quarkus), and cross-cutting procedures (`search-first`, `skill-stocktake`, `eval-harness`). ECC's own READMEs have quoted 56, 156, and 286 skills as the tree grew. AutoSkill pins one commit and one manifest. It does not follow a moving "latest".
Expand Down
2 changes: 2 additions & 0 deletions docs/BUILD_THREAD.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
# Build Thread

> Some code this page names may have been removed as dead code. `git log -S<name>` finds where it last lived.

A Build Thread is one governed change to one local Git repository. It shows why a candidate was accepted or refused, what it changed, and what actually happened. It can be resumed from its record after a crash. It needs Python 3.11+ and `git`, and no account, model, or network.

## Ten minutes
Expand Down
2 changes: 2 additions & 0 deletions docs/CLAUDE_CODE.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
# Claude Code as a proposer

> Some code this page names may have been removed as dead code. `git log -S<name>` finds where it last lived.

```bash
syberlabs propose --claude-code --trusted-tree # your own Claude Code login (subscription_local)
syberlabs propose --claude-code --bare # ANTHROPIC_API_KEY in your environment (api)
Expand Down
2 changes: 2 additions & 0 deletions docs/CRITIC.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
# The critic station

> Some code this page names may have been removed as dead code. `git log -S<name>` finds where it last lived.

An independent reviewer runs as a host check on every candidate, its findings are recorded in the thread's hash chain, an admission rule refuses acceptance while a high finding stands, and the pull request's `Reviewed-by` line is written from that record. The review is evidence a verifier can re-read, not a comment someone remembers.

## A critic is a check with a schema
Expand Down
2 changes: 2 additions & 0 deletions docs/E3_BENCHMARK_PLAN.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
# E3: the comparison benchmark against a real model (plan)

> Some code this page names may have been removed as dead code. `git log -S<name>` finds where it last lived.

Status: **pilot run 2026-10-21, see `docs/MODEL_COMPARISON.md`** (uninformative by the rule below; the full run was not started). **Fixtures v2 added 2026-10-21, unrun** (section below). Budget cap: **$50**, with a 30% margin, so the planned worst case must stay under **$38.46**.

## What E3 is
Expand Down
2 changes: 2 additions & 0 deletions docs/EVAL.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
# SyberEval: pre-registered evaluations through the real Build Thread

> Some code this page names may have been removed as dead code. `git log -S<name>` finds where it last lived.

SyberEval measures SyberWork's agents as governed workflows. It asks whether a provider configuration, run through the real Build Thread, produces admitted changes that are correct, safe and cheap. This page covers specs, tasks, arms, the runner and its evidence bundles, the preflight gate, how much of a hidden check reaches the proposer (`hidden_feedback`), the spend ledger, the safety gate, the scorer and the report (package `syberlabs/eval/`). The agent-facing safety tasks are not here yet (`spec/GAPS.md`, 62). The E3 harness (`benchmarks/model_vs_patch.py`, [MODEL_COMPARISON.md](MODEL_COMPARISON.md)) is still the tool for quick research. SyberEval is the instrument whose numbers are published.

## Principles
Expand Down
2 changes: 2 additions & 0 deletions docs/EVOLUTION.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
# Candidates and search providers

> Some code this page names may have been removed as dead code. `git log -S<name>` finds where it last lived.

SyberWork is the authority and evidence layer. A search provider is an artifact-search layer: it proposes changes. The provider can be a person's patch, a model adapter, or an EvoGit-style evolutionary search. Its output is **provisional** until the same contract, evidence, policy, and approval gates that govern every other SyberWork action admit it.

> Lineage is provenance, not evidence of superiority.
Expand Down
2 changes: 2 additions & 0 deletions docs/EXPERIMENTS.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
# Experiments: two arms compared from the journal

> Some code this page names may have been removed as dead code. `git log -S<name>` finds where it last lived.

```bash
syberlabs experiment label --name githits-pilot --item LIB-1 --arm off --tag needs_lookup # right after start
syberlabs experiment report --name githits-pilot # the arms compared, with every check that applies
Expand Down
2 changes: 2 additions & 0 deletions docs/GITHITS.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
# GitHits API client

> Some code this page names may have been removed as dead code. `git log -S<name>` finds where it last lived.

```bash
export GITHITS_API_TOKEN=... # from the environment only; never in a file in this repository
syberlabs githits search "retry with backoff" --target pypi:urllib3@2.2.1 --source code
Expand Down
2 changes: 2 additions & 0 deletions docs/PAID_PROBLEM.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
# Plan: a company pays because SyberWork solved a problem it has

> Some code this page names may have been removed as dead code. `git log -S<name>` finds where it last lived.

Status: **rehearsal implemented**. Gates 0, 2, 3, and 4 are not implemented. They need a real workflow, a price, a license choice, and a pilot cell. This document still names no customer.

**Goal.** A company that is not SyberLabs pays money, and the reason they state is a problem whose failure mode SyberWork's recorded behavior prevented. A payment for a roadmap, a workshop, or "an AI coding platform" does not meet the goal.
Expand Down
2 changes: 2 additions & 0 deletions docs/ROADMAP_STATUS.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
# Roadmap status

> Some code this page names may have been removed as dead code. `git log -S<name>` finds where it last lived.

This file tracks the next-stage roadmap (27 September 2026) against what the repository actually does. Each row says whether a claim is **implemented** (code and tests in this repository), **measured** (a number from a named run), or **proposed** (not built). A passing unit test is not evidence that a developer outside SyberLabs can use the kit.

## Gate 0 — integrate the foundation
Expand Down
2 changes: 2 additions & 0 deletions docs/SDK_FEEDBACK.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
# Development feedback and framework comparison

> Some code this page names may have been removed as dead code. `git log -S<name>` finds where it last lived.

This note is from building a second project, `examples/access_review.py`, on `syberlabs.Session` and from the numbers in `benchmarks/results/dev-session.txt`. It is not a market survey. The comparisons are capability comparisons, not speed claims against those projects.

## What a new project actually does
Expand Down
4 changes: 3 additions & 1 deletion docs/SDK_REDTEAM.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,7 @@
# Research and red team: closing the standards gap

> Some code this page names may have been removed as dead code. `git log -S<name>` finds where it last lived.

Date of this pass: 2026-09-27. The question is not "which product should replace this SDK". It is which properties the leading standards already require, where this repository fails them, and which change closes each gap without pretending the SDK is something else.

Standards used, and why:
Expand Down Expand Up @@ -131,7 +133,7 @@ The chain link is still `canonical`. Histories from `main` (`a2f909b`) keep veri
| 1 | Closed for the listed checks | Non-finite amounts are `amount_required`. Input kinds other than `string` and `integer` are `invalid_contract` at publish and `input_schema` at create. An `effect_started` with no `effect_succeeded`, `effect_rejected`, or verified reconciliation occupies the action as `effect_unresolved:<action>`. `effect_rejected` still allows a fresh proposal. `effect_unknown.error` and `source_unavailable` detail are codes (`destination_http`, `destination_unreachable`, `destination_error`, or `Rejected.code`), not exception text. |
| 2 | Closed for binding and the connection | `bind_arguments` and `next_compiled` are shared. `Work` keeps one SQLite connection and a lock. `Session` uses a re-entrant lock. HTTP execution stays on `Work`. |
| 3 | Closed, with integer `at` on new events only | New events store a JCS SHA-256 of the hashed fields on `event_side` / `Session.side_channel`. `verify_chain` checks that digest when a row has one. Old rows have none and still verify. The deciding rule id is on that side record. `cloudevent` exports CloudEvents 1.0. `conformance/jcs_vectors.json` is the byte vector. New `at` values are integer microseconds and are inside `event_digest`. Float-second rows, including the `main` fixture, are not rewritten. `at_json` preserves the integer across SQLite `REAL`. |
| 4 | Public key, local log, separate witness process | `witness` remains HMAC-SHA256 over canonical `{case_id, count, head}`. `sign_chain_head` is an Ed25519 DSSE envelope over the same object. The signature is not an event. `python -m syberlabs.witness` keeps the seed, rechecks the chain, and appends to the log. `Session.submit_witness` and `Work.submit_witness` send the events and keep only the public key. The case writer opens the log read-only. **KNOWN** limits: the pure-Python Ed25519 is not constant-time; a same-user process can still open the log file writable; an operator who holds the seed can rebuild the log. This is not Sigstore Rekor. |
| 4 | Public key, local log, separate witness process | `witness` remains HMAC-SHA256 over canonical `{case_id, count, head}`. `sign_chain_head` is an Ed25519 DSSE envelope over the same object. The signature is not an event. `python -m syberlabs.witness` keeps the seed, rechecks the chain, and appends to the log. `WitnessClient.submit` sends the events and the caller keeps only the public key. The case writer opens the log read-only. **KNOWN** limits: the pure-Python Ed25519 is not constant-time; a same-user process can still open the log file writable; an operator who holds the seed can rebuild the log. This is not Sigstore Rekor. |
| 5 | Closed | The planner context is `objective`, `allowed_actions`, `arguments`, `required_facts`, and acceptance `{id, passed}`. Fact values are not included. A signoff is refused when that role's `after_action` effect has completed and the signer started it. A signoff before that effect still records; acceptance still fails. |

HTTPS install now rejects non-public addresses and special-use names. `guard_request` rejects an HTTPS name that resolves to a non-public address before the request. `SYBERWORK_HTTPS_HOSTS`, when set, is an additional hostname list. Unset, public HTTPS names stay allowed. Loopback HTTP remains `127.0.0.1` with an explicit port.
Expand Down
Loading
Loading