Skip to content

Progressive early artifacts (time-to-first-useful)

Status: Layer 1 shipped (deterministic working brief after scrape + hiring refresh on fast and structured/deep collection; public path under run output_dir; MCP early_artifact_paths; hiring refresh preserves sample URLs; missing paths omitted from inventory). Layer 2+ remain design-only. Complements the 1.x quality and 2.0 memory workstreams; does not replace either.

Problem

Full provider-backed runs are often 30–120 minutes. Chat deep-research products often surface a usable narrative in tens of minutes. Primr already has useful early signals (recon, scrape progress, prep bundles), but the default operator experience still feels like “wait for the whole brief.”

That gap is real. Closing it should improve usability without weakening the full Strategic Overview / AI Strategy contract.

Goal

Ship intermediate, honest artifacts during a run so operators and agents can act earlier, while the full report remains the primary deliverable.

Success looks like:

  • Something useful appears well before the final MD/DOCX.
  • Early artifacts are clearly labeled incomplete (not final).
  • No extra billable model work unless the operator opted in.
  • Agents can discover early paths via existing body-free inventory resources.

Non-goals

  • Matching chat UX or continuous streaming of the full report.
  • Replacing the full brief with a short summary as the default product.
  • Always-on monitoring or background company watchers.
  • New regex content-quality gates on partial drafts.

Proposed layers (dependency order)

Layer 0 — already available (document, do not re-build)

Timing Artifact Cost
Seconds primr recon $0 model
Minutes primr prep evidence bundle + host workflow $0 Primr model
Mid-run Working folder scrapes, stage progress, run events included in run

Action: make these more visible in RUN_MODES / README “first useful output” without implying they are the full dossier.

Layer 1 — free progressive skeleton (first implementation slice)

During a paid or free collection path, write a working markdown file once enough structured evidence exists:

Suggested path (illustrative):

output/<company>/<Company>_Working_Brief_<date>.md

Contents (deterministic assembly first; no new model call):

  1. Company name, URL, run id, timestamp
  2. Status banner: WORKING BRIEF — incomplete; not the Strategic Overview
  3. Recon / DNS summary when present
  4. Scraped page inventory (counts, top URLs, access notes)
  5. Hiring signal counts when present
  6. Source index excerpt (domains, not full bodies)
  7. Explicit “still running” section list

Exit criteria for Layer 1 (met):

  • File appears after scrape (+ recon) complete, before long reasoning finishes
  • Zero additional model tokens beyond the run’s existing stages
  • Final Strategic Overview remains unchanged
  • Inventory / MCP artifact listing classifies it as intermediate (not primary_report)
  • Public dated file honors run output_dir / MCP job dirs
  • Hiring refresh re-emits counts without wiping first-party/external URL samples
  • Early-path inventory skips paths that no longer exist on disk

Validation: free offline fixtures + hermetic unit tests for assembly, classification, hiring refresh, and existence filtering.

Layer 2 — optional early executive sketch (opt-in)

Only if Layer 1 is stable and operators ask for prose earlier:

  • Optional flag or mode, e.g. --early-sketch (name TBD)
  • One bounded writing call on collected evidence only
  • Same confidence discipline: hedge, no fake Confirmed density
  • Must not auto-promote sketch claims into the final report without the normal pipeline

Validation: small paid eval budget only after free assembly tests pass; compare sketch vs final for contradiction rate, not eloquence.

Layer 3 — progressive agent handoff

  • Body-free resource or status field: early_artifact_paths, stage, completeness
  • Primr Zero path already checkpoints host writing inside prep; align naming so agents do not confuse working brief with primary_report

Placement in the pipeline

Prefer hooks at existing seams:

  1. After recon + scrape inventory is durable
  2. After fast collection stages that already write working-folder state
  3. Never block the main pipeline if early write fails (log + continue)

Do not introduce a second orchestrator or a DAG framework for this.

Artifact contract

Role Final SO Working brief
Inventory class primary_report intermediate / working
QA gates existing shipping gates structural only (path, header, non-empty)
DOCX default for final optional later via primr render
Overwrite dated final may refresh in place during the run

Risks

Risk Mitigation
Operators ship the working brief as final Loud incomplete banner; inventory role
Extra cost Layer 1 is free; Layer 2 opt-in
Scope creep into chat streaming Keep discrete files + phase markers
Stale mid-run files after failure Status field + run events

Relation to other workstreams

  • Epistemic quality (#1): early artifacts stay thinner; do not lower Confirmed standards on the final report to feel faster.
  • Backend freedom (#2): early assembly should work for cloud, host, and local collection the same way.
  • Memory (2.0): later, a working brief can seed delta mode; do not wait for memory to land Layer 1.
  • Run-scoped epistemic ledger: findings, strategic inferences, and section evidence-usage views are structured supporting artifacts, not another working-brief layer. They remain separate from early-artifact paths and are never classified as primary_report; see run-epistemic-ledger.md.
  • Competitive positioning: this is a usability answer to chat speed, not a claim that Primr is faster overall.

Suggested first PR shape (when implementing)

  1. Pure assembler + classification + tests (no CLI flag change if a default mid-run write is safe).
  2. Wire write after scrape/recon in one path (fast first).
  3. Docs: RUN_MODES “what you get while waiting.”
  4. Optional: MCP inventory field only after CLI path is proven.

Validation cost

  • Free: unit tests, fixture working folders, dry-run path checks.
  • Paid (later): one cheap live run to confirm timing of the working brief vs final SO on ExampleCo-style placeholder targets only.

Explicitly not in v1 of this design

  • Streaming tokens to the terminal as the product surface
  • Auto-email / Slack of partial drafts
  • Treating prep-only output as a full paid Strategic Overview

Recorded collection signature

Prep and provider-backed completion show one coverage signature immediately before their final next step. Word carries the same counts once, as a muted closing Collection note, so a client-facing report opens with the title and the Executive Summary rather than collection diagnostics (#263 supersedes the under-title placement from #229/#236). An unusable or empty collection keeps the warning color in that note. This is collection metadata, never a claim-confidence or calibration score.

core.collection_coverage derives counts from the collector's existing source classification. Only nonempty current first_party captures enter the first-party total; archives, fallback captures, regulatory/reference pages, and model synthesis stay in the separate other-captures count. Only verified hiring identities enter the linked-postings count. No global or most-recent trace is consulted. Blocked-page and redirect-loop counts stay absent when the run did not record them; a generic redirect-limit failure does not establish a loop.

Provider runs write a source index and bind its digest and scalar counts in _run_state.json. Missing, malformed, linked, changed, or legacy metadata shows unknown counts rather than fabricated zeros. Prep uses its published index; primr render --evidence-bundle PATH additionally verifies the existing manifest, artifact hashes, and captured source identities before rendering. Legacy indexes without capture digests use the text frames already verified by that same reader, including empty-frame handling. They do not infer captured pages from raw length. Without that explicit bundle, standalone rendering makes no collection claim. A partial prep bundle remains visibly unusable even when it contains captured pages.